Learning in Standard-Form Contracts: Theory and Evidence

Abstract

Why are some contractual terms revised continuously while others are stubbornly fixed? We offer an account of both change and stickiness in standard-form contracts. We hypothesize that drafters (sellers) are more likely to revise their standard terms when they have an opportunity to learn about the terms’ costs from experience. Consider a warranty. Offering a warranty in an initial period will expose sellers to claims about malfunction by purchasers, allowing sellers to learn whether it is desirable to offer it going forward. When drafters are unable to learn in this manner, either because they fail to experiment or because the term in question is one where there is no increased opportunity to learn from experience, such terms will be revised relatively less frequently. While learning and change occur through various channels, we posit that, all else equal, terms that carry an opportunity to learn from experience will be revised more frequently, whereas terms or term modalities that do not will contribute to stickiness and stagnation. Our results support this hypothesis. Using a large sample of changes in business and consumer standard-form contracts over a period of seven years, we find that sellers are more likely to revise terms that offer an opportunity to learn from experience than those that do not. These findings are further illustrated and supported by interviews with in-house counsel. The results suggest that standard-form contract terms evolve over time as sellers learn experientially about their costs and risks. Our analysis offers new accounts for the use of boilerplate, stickiness, and change and has normative implications for the optimal design of default rules and product features (JEL codes: K12).

1. INTRODUCTION

On February 9, 2018, L.L. Bean sent a letter to its customers, announcing it would terminate its lifetime warranty for its products, which had famously been in effect for over one hundred years. Why would the firm abandon the warranty that had helped make it trusted and famous?

Like most contract terms offered in mass-market transactions, the warranty was part of a standard-form contract. A common feature in standardized agreements is that their terms tend to be “sticky”, even though contracting parties should revise their agreements when doing so enhances the value of their transaction. Yet stickiness is not a general phenomenon. While some terms seem to be resistant to change,¹ others, like the L.L. Bean warranty (though it took over a hundred years to change), are revised at a later stage. In its letter to customers, the company explained:

Increasingly, a small, but growing number of customers has been interpreting our guarantee well beyond its original intent. Some view it as a lifetime product replacement program, expecting refunds for heavily worn products used over many years. Others seek refunds for products that have been purchased through third parties, such as at yard sales. Based on these experiences, we have updated our policy.²

What led L.L. Bean to change its long-time warranty? Learning from the experience with their customers’ claims under the warranty. This is the topic of our paper: we examine contract change as a result of learning from experience. We posit that contract terms that offer an opportunity to learn in this way are more likely to be revised, all else equal, than terms that do not confer such an opportunity.

Recognizing the workings of learning from experience leads to a novel account of both stickiness and change in standard-form contracts. Not all terms are created equal. Some terms—what we call “asymmetric-learning terms”—have a learning modality and a nonlearning modality. For instance, L.L. Bean could learn the cost of offering the lifetime warranty only by offering it; offering no warranty does not generate experiential learning about the costs associated with it. This asymmetry creates what is known as a “real-option value” for the learning modality, that is, by offering the warranty now, the drafter purchases the option to make an informed revision decision later. Since the value of this “option to revise” makes offering an asymmetric-learning term in its learning modality more desirable than short-term considerations would warrant, such terms are offered more often initially. But precisely because they generate additional information that is valuable in the revision process, they are also revised more often later as compared to terms where learning is symmetric, that is, occurs irrespective of initial contracting choices.³

In addition, when learning is asymmetric, change has a predictable direction. Terms adopted initially in their learning modality are likely to change to their no-learning modality upon learning; indeed, L.L. Bean eventually withdrew its lifetime warranty. Conversely, the initial choice of the no-learning modality generates no new information and hence results in stickiness. In contrast, symmetric-learning terms lead to stickiness or change—all else equal—at the same rate. Our approach differs from extant accounts and helps explain not only why terms change but also which terms are more likely to change and in which direction.

Do firms serendipitously and unintentionally experience the costs associated with offering a particular term, or do they intentionally experiment with contract terms in order to learn? While we know little about Leon Leonwood Bean’s mindset in 1912, Zappos’ CEO Tony Hsieh was deliberately experimenting when he started making a $100 pay-to-quit offer to its new employees—in fact, a contract-modification proposal, a.k.a. “The Offer”—in the hope that resisting the temptation to leave would improve employee engagement. The risky side of this offer is that employees might take it, resulting in high turnover costs for the firm. As he learned from experience that such costs were low, Hsieh retained and even increased the offer.⁴ The difference between mere experience and outright experimentation is that experience only affects the firm’s revision strategy, while experimentation also shapes their drafting choices. That is, a firm may adopt a term that leads to short-term losses because of the long-term benefits of learning, which in turn makes the probability of ex-post revision even greater.⁵

Our paper explores these considerations by first offering a simple model of experimentation and learning in standard-form contracts. To keep our focus on learning from experience, the model zeroes in on cases in which screening customers through a menu of differently priced options is not possible,⁶ so the firm offers the same, standardized contract to all customers,⁷ just like in the L.L. Bean example. The model generates predictions on drafting choices, stickiness, and change for asymmetric-learning terms versus symmetric ones while accounting for the costs of opting out of a default rule, switching costs at the revision stage, and the firm’s expected customer-base composition and growth prospects. Stickiness, in our framework of analysis, results from the inability to learn. In particular, we show that default terms, which are typically very attractive for firms and customers because they are known and easily adopted, might for this very reason also be more prone to be revised over time.

Our focus is exclusively on experiential learning. Firms, of course, also learn in other ways, such as by studying the terms adopted by their competitors or from case law, technological innovations, news reports, and other sources. A common feature of these learning channels is that they, in contrast with experiential learning, tend to function largely independently from the specific contractual choices firms make.⁸ There are certain terms whose relative costs are best ascertained through the direct feedback only their use generates. In the L.L. Bean case, the company learned from experience that the extensive warranty had cost them $250 million in the five years prior to dropping it.⁹ We posit that learning from experience is more likely to be important when learning is about a term’s cost for the firm rather than its value for customers or customer characteristics, when changing product features is more cumbersome than revising the contract, and when experimenting with prices and product features is unfeasible.¹⁰

We explore the role of experiential learning empirically by studying the changes in contracts of a sample of software firms across various markets over a seven-year period. Recognizing that there are other drivers of term stickiness and innovation identified in the literature, we also conducted interviews with in-house counsel at Amazon, Fiat Chrysler Automobiles (FCA),¹¹ Google, Swapfiets, and a large media company, to explore the extent to which our hypotheses about experiential learning and firms’ propensity to revise terms track real-life contracting practices.

For our systematic analysis, we examine 32 common terms of a sample of End User License Agreements (EULAs) used by 246 firms that track contractual changes from 2003 to 2010. We grade each term against the default rules of Article 2 of the Uniform Commercial Code (UCC) and further classify each into categories that reflect their opportunity to learn. We first determine whether each term is associated with symmetric versus asymmetric experiential learning. For the latter, we identify whether the learning modality of the term occurs at default or when the firm opts out of the default. For example, we classify implied warranties as asymmetric terms where learning occurs only at the default. Opting out of the default does not expose firms to breach of warranty claims but also deprives them of experiential learning.

We find that firms are initially more likely to adopt terms at their default modality, consistent with theory and evidence that defaults are sticky. Yet terms set at the default are more likely to be revised at a later stage regardless of their learning modality, suggesting that such defaults might have been chosen sub-optimally precisely because of their attractiveness. Dividing terms into their respective learning modalities yields a strikingly different result. Consistent with our hypothesis, we find that symmetric-learning terms are equally likely to be revised at a later period, irrespective of the modality selected at an initial period. In contrast, we find that asymmetric terms are several times more likely to be revised at a later stage if the firms included the term in its learning modality at an initial period (and the reverse for asymmetric-learning terms adopted in their nonlearning modality). For example, we find that sellers are more likely to revise warranties and intellectual property restrictions when these are set in their learning modality at the initial period, among others. This is consistent with the hypothesis that firms may, to a large extent, willfully experiment with contractual terms. The results are robust to alternative characterizations of learning modality classifications. Specifically, this difference is also present when comparing the opt-out rates of non-default symmetric- and asymmetric-learning terms, offering further support that learning is a distinct effect from sellers’ tendency to switch away from sticky defaults.

We also consider alternative explanations for our findings. In addition to being sticky, default rules (such as implied warranties) tend to benefit consumers, a factor that may drive sellers to opt out and revise terms in a more self-serving manner at a later time (Marotta-Wurgler & Taylor 2013). In our data, consumer-friendly defaults are evenly distributed among symmetric and asymmetric terms. If such default terms confer similar benefits to consumers relative to their opt-out modality, then we would expect a shift away from all consumer-friendly defaults with the same frequency; yet, we do not see this. Rather, it is those defaults that carry an opportunity to learn that are revised more frequently. For symmetric learning terms, we find a similar rate of opt-out from consumer-friendly defaults relative to their alternative modality. A critical assumption is that asymmetric defaults are not more pro-consumer than symmetric defaults, as this would give sellers a more urgent reason to revise such terms later on. While we have no reason to believe that consumer-friendly asymmetric-learning default terms are more pro-consumer than other types of consumer-friendly defaults, we cannot rule this possibility out exclusively with the dataset of EULAs.

The interviews with in-house counsel offer additional support for the role of experiential learning as a mechanism that drives firms to revise their terms. All interviewees underscored the need to rely on standardized terms to manage scale but also acknowledged that terms need occasional revisions. Firms’ anticipation of contractual revisions, as well as learning, both direct and indirect, influence firms’ contracting decisions both at an initial stage and later on.¹² Importantly, during the interviews, the subjects identified experiential learning as a mechanism motivating contractual revisions for a particular set of terms. Younger firms, such as Swapfiets—a successful startup operating a novel long-term bike-rental service—were particularly keen on learning experientially, which they did purposefully, while older firms, such as Fiat—which makes automobiles—may have had less to learn from their mature markets. In addition to further supporting our hypothesis, these accounts offer rich details regarding the ways in which firms learn from experience and relay such knowledge to those individuals within the firm responsible for revising agreements.

Our paper makes three main contributions. First, it offers a new theory of contractual change that is driven by experiential learning. This additional mechanism has not been previously identified in the literature, to the best of our knowledge. Second, the theory accounts for both stickiness and change. Third, terms that allow firms to learn from experience generate a real-option value if offered in their learning modality, which may affect ex-ante contractual decisions. This real-option value is not unique to terms; it also applies to product and service features when firms can learn experientially and revise such features later on based on newly acquired experiential information.

The paper proceeds as follows. In Section 2, we situate our paper within the literature on standard-form contracts and contractual innovation. In Section 3, we propose a simple theoretical model of experiential learning and derive testable predictions. In Section 4, we explore these predictions using a unique dataset of standard-form contracts. In Section 5, we report our findings from interviews with in-house counsel. In Section 6, we conclude with additional normative implications of our theory. The Appendix contains theoretical proofs and details of the empirical analysis.

2. LEARNING, STICKINESS, AND INNOVATION IN STANDARD-FORM CONTRACTS

2.1 Stickiness and Change

The benefits of standardization are well understood and have been explored extensively. As terms become increasingly common and well-known, they are easier for contracting parties and courts to interpret. They also confer various spillover effects, such as lower reading costs, increased certainty of legal interpretation, and reduced litigation risk (Kahan & Klausner 1997; Gillette 1998; Choi & Gulati 2004). Yet these and other benefits resulting from the use of boilerplate terms may stand in the way of change, even when doing so might be efficient (Kahan & Klausner 1997). Reluctance to change in light of a superior alternative could give rise to agreements with terms that no longer serve the contracting goals of the parties, either because they no longer reflect the optimal allocation of rights and risks between them, or because they might be interpreted unfavorably by a court (Choi, Gulati & Scott 2017), among other reasons.

A number of factors contribute to stickiness. Stickiness tends to be associated with markets that experience network benefits that arise from firms’ simultaneous adoption of a term (Kahan & Klausner 1997), as well as with agreements drafted by law firms, whose hierarchical structure favors the re-use of old forms (Hill 2001; Gulati & Scott 2013). Firms’ incentives to innovate are further diluted by weak property rights in contractual innovations (Davis 2006) and the existence of default rules, which may have ceased to reflect an efficient allocation of rights and risks for a significant majority of contracting parties due to obsolescence (Schwartz & Scott, 2021). When states enact particular defaults, contracting parties might find it cost effective to just adopt them (Goetz & Scott 1985; Schwartz & Scott 1995, 2003), a tendency further reinforced by the status quo bias (Korobkin 1997). Such parties might also be reluctant to deviate from them when they perceive that opting out might signal negative information, even if value generating (Johnston 1990; Kahneman, Knetsch & Thaler 1991; Bernstein 1993; Ben-Shahar & Pottow 2006).

Despite these obstacles, change and innovation can and do still happen. Large repeat players, such as law firms and investment banks, can find it profitable to invest in innovation—even in the absence of strong property rights—through their ability to spread costs among clients (Kahan & Klausner 1997; Gulati & Scott 2013). In-house counsel in legal departments of firms that engage in mass-market commerce work closely with management and understand changes in technology that might give rise to new terms. In-house counsel is also more likely to receive feedback from offering or refraining from offering particular types of terms, allowing them to revise the agreements to adapt to new legal and market environments (Macaulay 1966; Triantis 2013). Innovation also results from the role of lawyers negotiating agreements in multiple transactions (Jennejohn, Nyarko & Talley 2022) or in related deals (Bishop, Jennejohn & Jones 2022), and can also be spurred by exogenous “shocks”, such as new laws, changes in legal interpretations of terms, or technological advances, driving firms to revise their agreements (Choi & Gulati 2006).

Finally, there is a rich literature exploring the benefits of strategic experimentation (Bolton & Harris 1999). In this literature, as in our model, the “risky” choice creates a real option (the option to revise the decision later on, after learning), while the “safe” choice induces no learning and hence does not carry an option to revise the decision later. Our model incorporates a simple version—a two-period, binary, two-armed bandit problem—of the more general multi-armed bandit problem (Robbins 1952) studied in this literature and various other disciplines, from computer science to clinical trials and portfolio theory. To the best of our knowledge, we are the first to apply this approach to the design of standard-form contracts.

Most of the empirical evidence on contract change, innovation, and stickiness comes from studies of bond covenants and financial products. Kahan & Klausner (1997) and Choi and Gulati (2006) found evidence of switching and learning costs in the corporate bond covenant context. Most strikingly, Choi, Gulati & Scott (2017) found that even in the trillion-dollar sovereign debt market, inefficient terms encrusted in boilerplate contracts were slow to change. Choi, Gulati & Posner (2013) found an S-shaped innovation pattern in sovereign debt contracts, where parties slowly move from the old standard to a new one in response to various exogenous shocks. In the law firm context, Gulati & Scott (2013) found that lawyers in law firms failed to revise terms even after those terms had acquired ambiguous meanings that increased litigation risk. In the insurance context, Schwarcz (2011) found evidence of innovation away from the ISO form, which is the standard insurance document. Coates (2016) found significant changes in merger agreements over time, unveiling that such contracts had doubled in size and that about 20 percent of such change could be attributed to new terms. Nyarko (2021) found stickiness in dispute resolution clauses in a large sample of template contracts negotiated by sophisticated parties. Finally, Marotta-Wurgler & Taylor (2013) found evidence of terms changing in reaction to litigated cases and changes in the enforceability of terms.

To summarize, there have been numerous accounts that explain and document either stickiness or change in standard-form contracts. In this paper, we propose a new mechanism to account for contract change: learning from experience. To the best of our knowledge, this is the first paper to explore this mechanism in the standard-form contract setting and to offer an account that explains both stickiness and innovation in the absence of external shocks.

2.2 Experiential Learning Versus Other Forms of Learning

Learning is a fundamental driver of change. Yet learning occurs in different ways. Firms can learn directly, by interacting with customers, through the experience earned by experimenting with particular terms or term modalities. Firms can also learn indirectly through other channels in ways that are unrelated to the contractual choices made at an initial period. Consider a term in a software End User License Agreement (EULA) limiting the number of devices where the software can be installed, which can clearly affect demand for the product. Learning in this context occurs largely independently from the form of the contract term offered related to the number of devices. The firm learns about its demand regarding customer use, whether the use is limited to one, five, or one hundred users. That is, it is unlikely that a firm will learn more about the value of a term imposing a limit on the number of authorized users of a particular software product by experiential learning from customers with one particular version of the term versus another. All modalities of the term lead to the same amount of learning. More likely for this type of term, learning about customer preferences and uses of software can be achieved by looking at purchasing patterns or examining the offerings of competitors in their own market and adapting terms accordingly.

This type of indirect learning takes multiple forms. Firms can learn from litigated cases about possible contractual choices as well as about the enforceability of particular clauses (e.g., a “change of terms” clause that allows firms to modify standard agreements unilaterally) (Gulati & Scott 2004; Marotta-Wurgler & Taylor 2013; Schwarcz 2021). The literature on contractual innovation has also pointed out that firms can learn from each other’s contracts, creating a well-known free-rider problem (Goetz & Scott 1985; Davis 2006).¹³ Law firms are also a conduit of indirect learning by transmitting knowledge to their seller-clients, who can then revise their terms accordingly (Jennejohn, Nyarko & Talley 2022).

More recently, legal service firms like Bloomberg Law and Legal Zoom have begun offering standard terms for different types of contracts, allowing firms to innovate at a relatively low cost (Triantis 2013). Blogs, trade publications, word of mouth, and internet forums offer additional sources of free advice regarding terms. All of the aforementioned channels enable learning and change, but the mechanism by which this happens is unrelated to the firms’ experience with its adoption of a term or term modality. We refer to these forms of learning as “indirect” learning mechanisms because they can occur independently from contractual choices made or direct interactions with customers implicating a term.

In other circumstances, learning is not indirect but rather the result of firms’ interactions with customers and, to some degree, the result of the products, services, and terms offered. This form of learning is experiential in nature, but the experience is not conditioned on the occurrence of an event that might implicate the contract in specific ways. For example, firms can acquire valuable knowledge from interacting with customers through employees and customer service channels, where learning is not necessarily mediated by the contract terms themselves (Hoffman 2018).¹⁴ Customers can call firms and inquire about the meaning or implications of a particular term without demanding rights under the contract. This might allow firms to learn about how customers understand the contract or particular features of a product or service. Firms also gain valuable information from feedback offered through online customer reviews (Ghose & Ipeirotis 2011), which can lead to change. We refer to these forms of learning as “direct” learning because they result from customer interactions or aspects related to the product or term offered.

One particular form of direct learning is experiential with respect to the contract. In particular, firms can learn from experience resulting from the use and implication of a particular term. The most natural example is a warranty. A seller can offer it or not; if the seller offers the warranty and the product breaks, the customer can bring a claim for breach of warranty. In honoring the warranty, the seller learns the costs of offering this particular term. Learning occurs if three conditions are met: the seller offers the warranty, the product breaks down, and the buyer brings a claim for breach of warranty. Note that the difference with this form of experiential learning as compared to other forms of direct learning, such as when customers provide feedback, is that additional learning only occurs when a particular term or term modality (e.g., adopting the default or opting out of it) is offered (and implicated) and not otherwise. There are some terms that tend to lend themselves to experiential learning more than others. Warranties, terms offering maintenance and support, and terms where firms are able to experience the cost of offering them from customer claims or actions tend to fit well in this category.

Terms differ in purpose and effect. A warranty is different from a term in a EULA restricting the number of users of a product. Given these differences, the ability of a firm to learn from experience, as well as what is learned, will also likely depend on the type and function of the term involved. L.L. Bean’s experiential learning from offering the warranty took the form of receiving breach of warranty claims from customers and allowed it to learn about the costs of offering such an extensive warranty. Firms can and do likely adjust a number of dimensions related to their product or service as a result of this learning.¹⁵ Yet, in many instances, experiential learning can lead to the revision of terms. Warranties are the most salient example. Upon learning about their cost from experience with customers bringing a breach of warranty claim, firms can react more nimbly by revising the warranty instead of changing their products or product quality entirely. L.L. Bean’s new one-year warranty incorporates what it learned from its experience with customers giving the firm discretion to accept later returns that are not caused by customer advantage-taking: “After one year, we will consider any items for return that are defective due to materials or craftsmanship”.¹⁶ Other terms, like those related to defining the scope of use or assigning risk of loss, are also likely to be revised upon learning. Section 5 offers more details regarding firms’ possible courses of action upon learning about the relative cost of the terms they offer.

While there are several types of learning that may lead firms to revise their terms, we focus exclusively on the direct experiential learning that occurs when a particular modality of a term is invoked by an event that allows the firm to interact with its customer or other contracting party—such as a supplier¹⁷—and thus experience and learn about the term’s cost. We expand on the different forms of direct experiential learning in the next section.

3. A FORMAL THEORY OF EXPERIENTIAL LEARNING

3.1 Symmetric Versus Asymmetric Experiential Learning

A core feature of our theory of experiential learning in contracts is that learning may depend on the contract terms offered in the past. The firm’s contractual choices in an initial period (time 0) determine learning in the interim period and, consequently, affect the firm’s contractual choices in the subsequent period (time 1). (This simple timing matches the structure of our data and is replicated in the formal model that we will introduce momentarily.) Experiential learning may occur symmetrically when the default and the opt-out options offer analogous opportunities for learning, or asymmetrically, when the feedback generated by the two (or more) modalities is different. To illustrate, consider two stereotypical examples of symmetric (S) and asymmetric (A) learning terms.

Consider firms selling tax preparation software packages to customers. Unless modified or opted out, the standard default rules offered by Article 2 of the UCC will apply to the transaction, including implied warranties. As they write their contract, firms can choose to keep the defaults or opt out of them (Marotta-Wurgler 2008).¹⁸

Firms may consider offering a term restricting the use of the product by the customer, a term that opts out of the default term, which imposes no restriction. A common example is a restriction on commercial use, whereby the customer is only able to use the product for personal or household purposes. Customers value unrestricted use, but the firm might benefit from carving out specific uses that, depending on market conditions, could be licensed separately. At the time of writing the contract, the firm might be uncertain about the relative costs of offering the default. Alternatively, the firm may not anticipate at that time the need to revise the term later on. These costs, however, may become known later on through the firm’s interactions with its customers. With this information at hand, at time 1, the firm can confirm or revise the contract. Crucially, the firm learns new information from interacting with customers or through other means both if it adopted the default term of no restrictions at time 0 and if it opted out of it.

Consider a firm that adopted the default of no restriction. The firm might monitor use by customers (a form of indirect learning) or receive feedback from them in the form of inquiries about the scope of use (a form of direct learning) and notice that a specific category of its customers uses the product particularly intensely along a specific dimension. For instance, it may discover that small retailers use the software produced by the firm for marketing purposes. As a result, it could in the future restrict the use of the software and sell a license for marketing uses separately at a higher price. Yet, the firm would learn the same information even if it adopted the restriction-on-use opt-out term. The firm might start out by restricting use to non-marketing uses and then learn about the demand for the license for marketing uses directly by offering it for sale separately. In both cases, the firm may acquire valuable information on the net value of allowing or restricting specific uses. Since the feedback generated by the two modalities of the term is unlikely to differ very much, we classify such terms as symmetric-learning terms. Of course, the firm might also not anticipate learning, yet learn nonetheless, naïvely, and revise its term based on the newfound information. Learning here might be direct or indirect. Regardless of the mechanism, learning is symmetric for both modalities of the term.

Now consider a different term that firms selling goods may offer, like the warranty of merchantability. This is a term implied by UCC Article 2, which a firm can adopt by staying silent (the default term) or may disclaim by including a standard implied warranty disclaimer (the opt-out term).¹⁹ The decision whether to offer the default warranty or to opt out of it will depend on the tradeoff between the price increase that the firm may be able to capture by offering the default—assuming customers will be willing to pay more for the product with a warranty—and the costs of such offering, in the form of claims against the firm. At the outset, firms may not have enough information to assess the costs associated with offering the warranty with absolute accuracy, yet the available information may be enough to assess the probability of facing high or low costs ex post. As before, the firm may learn new information by interacting with customers. Now, however, learning depends on the term offered at time 0.

As in L.L. Bean’s case, firms adopting the default warranty will face future claims with some probability and learn the amount of damages resulting from breach. The claims will likely depend on the product, the way customers use it, the activities customers are involved in, local conditions, and other factors that will typically be unknown at the outset. In L.L. Bean’s case, the firm also learned how customers interpreted and invoked the warranty. Facing claims from different customers allows firms to better learn the cost of being exposed to liability as a result of offering the warranty.

Firms that opt out of the default by disclaiming implied warranties will not face such claims and thus reduce the chances to learn the costs of offering such terms. Of course, not offering a warranty is not set in stone; a firm not offering one in an initial period might still decide to offer one in a later period as a result of changes in demand, or for competitive reasons, among other factors. All else equal, a firm that offers a warranty in an initial period benefits from additional learning as a result of its experience with customers. Such a firm will be relatively more likely to revise it at a later period due to such learning. The implied warranty of merchantability is an asymmetric-learning term. More precisely, this is a term where the firm learns more if it offered the default option.

Table 1 provides an overview. Terms of type S are symmetric-learning terms and are such that the firm receives new information at time 1, irrespective of whether it adopted the default or opted out of it at time 0. An example of such terms are restrictions-on-use terms, described earlier. The distinctive characteristic of a symmetric-learning term is that a decision of whether to revise the term at a later period will arise irrespective of the term’s modality at an initial period. In contrast, terms of type A are asymmetric-learning terms: between time 0 and time 1, the firm learns the costs associated with the term only if it has adopted what we call the learning modality of the term, which could be either the default (as in the implied warranty example above) or the opt-out. If the firm does not offer the learning modality at time 0, it will not learn anything—or, more generally, it will learn less—from interacting with customers.

Table 1.

Information types and modalities of contract terms

Information type		Default	Opt-out
Symmetric-learning terms (S)		Learning	Learning
Asymmetric-learning terms (A)	From default	Learning	No learning
	From opt-out	No learning	Learning

3.2 Model Setup

In the model, we consider a monopolistic firm (a seller) of type p—which is randomly drawn uniformly from the unit interval—where p is the probability that the firm’s customers (buyers)²⁰ are high-intensity users (type H, or intense users for short). With the complementary probability, 1 – p, the firm’s customers are low-intensity users (type L, or regular users for short). While the firm knows its type, it ignores the type of its customers. Crucially, customers also ignore their type, and hence the firm cannot screen among them by using a menu of differently priced contractual options.²¹

At time 0, the firm offers all of its customers the same standard-form contract, which can come in two guises. The firm can either adopt a default term prescribed by the law or opt out of it. In the interim period, the firm has an opportunity to learn the costs associated with the terms it offered and infer from them the type of its customers. Then, at time 1, the firm can offer the same contract as at time 0 or revise it.

If the firm offers the default term, all customers—irrespective of their type—are willing to pay v^D for the product, while if the firm offers the opt-out term, customers are willing to pay v^O. We can then write the relative value of the default term as v = v^D – v^O.

The two terms also entail different costs for the firm, which depend on the type of customers. If the firm offers the default term, costs are $c_H^D$ if customers are intense users or $c_L^D$ if customers are regular users. Similarly, the opt-out term yields costs $c_H^O$ or $c_L^O$⁠. We can write the relative costs of the default term as $c_H=c_H^D-c_H^O$ or $c_L=c_L^D-c_L^O$ depending on the type of customers. Let $c\in \{c_L,c_H\}$ be the actual relative cost of the default. While the firm knows the values of $c_L$and $c_H$⁠, the firm ignores c because it ignores the type of its customers.

To capture the interesting scenarios, we focus on cases where:²²

 

Assumption 1: c_L < v < c_H

The firm maximizes profits. Given Assumption 1, if the firm knew the type of customers it faces, it would offer the default term to regular users (L)²³ and the opt-out term to intense users (H).²⁴ The firm, however, ignores the type of customers when drafting the contract, and hence its initial contractual choice will depend on the expected short-term costs at time 0 and on the possibility to learn and hence revise the contract at time 1. We consider three learning scenarios:²⁵

• Symmetric learning: If $c_H^D\ne c_L^D$ and $c_H^O\ne c_L^O$⁠, then the firm experiences different costs irrespective of the initial contractual choice and hence learns both from the default and from the opt-out.

• Asymmetric learning from the default: If $c_H^D\ne c_L^D$ but $c_H^O=c_L^O$⁠, then the firm experiences different costs when offering the default and hence can learn the type of customers from it, while if it offers the opt-out term, costs are the same for both types of customers, and hence there is no learning.

• Asymmetric learning from the opt-out: If $c_H^D=c_L^D$ but $c_H^O\ne c_L^O$⁠, then the firm experiences different costs when offering the opt-out but not when offering the default and hence only learns from offering the opt-out.

In the next section, we present our results. Most of the technical details of the derivations are in the Appendix, where we also relax two assumptions made here for simplicity and show that they have no bearing on our findings: we allow for a general distribution function for the population of firms (here assumed to be uniform) and we allow for a continuum of degrees or probabilities of learning (here restricted to be either 0 or 1).

3.3 Main Results

With symmetric learning, the firm’s decision at time 0 is purely driven by a balance of expected short-term costs because learning occurs irrespective of this initial choice. The firm will choose the default if its value is greater than its expected costs, that is, if v > p_cH + (1 – p)c_L, from which we can define

$${\displaystyle \begin{array}{l}{\displaystyle p^S\equiv \frac{v-c_L}{c_H-c_L}}\end{array}}$$

(1)

so that the firm will offer the default term if $p<p^S$ (low expected costs) and opt out otherwise (high expected costs). At time 1, adopters may discover that costs are high—this happens with probability p—and decide to switch to the opt-out. The area of the gray triangle in the upper-left corner of Figure 1(a) depicts the ex-ante probability mass of switches from the adoption of the default term to opt-out, which is equal to ${(p^S)}^2/2$⁠. (To draw Figure 1, we have set $v=c_L+c_H/2$ and hence obtained $p^S$ = ½ and $({p^S)}^2/2=1/8$⁠.) Conversely, firms with $p>p^S$ opt out at time 0 and decide to switch with probability 1 – p to the default at time 1. The gray triangle in this part of the graph depicts the ex-ante probability of switches from opt-out to default, which is equal to ${(1-p^S)}^2/2$ (which is again equal to $1/8$ in the figure).

Figure 1

Adoption decisions at time 0 and time 1.

Consider now asymmetric learning from the default option. If the firm chooses the default at time 0, it will face expected costs equal to pc_H + (1 – p)c_L but will also have an opportunity to learn and be able to revise the contract. After learning, the firm’s time-1 payoff is v – c_L with probability 1 – p (the firm learns that costs are low and confirms the time-0 choice) and zero²⁶ with probability p (the firm learns that costs are high and switches to the opt-out). This—that is, (1 – p) (v – c_L)—is the real-option value of the default term, which pushes the firm to choose the default for its dynamic gains, even in cases when it yields static losses. Now the firm chooses the default if v + (1 – p) (v – c_L) > pc_H + (1 – p)c_L, which yields a higher threshold for opting out:

$${\displaystyle \begin{array}{l}{\displaystyle p_D^A\equiv \frac{2(v-c_L)}{(c_H-c_L)+(v-c_L)}.}\end{array}}$$

(2)

We can visualize the firm’s choices in Figure 1(b) (where $p_D^A=2/3$⁠). Compared to the symmetric-learning term, more firms choose the default at time 0. Adopters, however, switch to the opt-out with relatively high probability, especially in the range $[p^S,p_D^A]$⁠, that is, in cases that would have resulted in opt-out at time 0 had the term been of a symmetric type. These are instances in which the default has a negative expected short-term value, and it is chosen purely for its learning value, that is, for the option to make a perfectly informed decision at time 1. The probability of switches away from the default is ${(p_D^A)}^2/2$ (which is equal to $2/9$ in the figure). Instead, firms that choose to opt out irreversibly confirm that decision at time 1 because, differently from what happens with symmetric-learning terms, no new information is acquired in the meantime.

Similarly, it is easy to show that with asymmetric learning from the opt-out, the threshold is:

$${\displaystyle \begin{array}{l}{\displaystyle p_O^A\equiv \frac{v-c_L}{(c_H-c_L)+(c_H-v)},}\end{array}}$$

(3)

which is less than p^S. In this case, there are no switches away from the default, which is now the no-learning modality, and the probability of switches away from the opt-out is equal to ${(1-p_O^A)}^2/2$⁠, as depicted in Figure 1(c) (where $p_O^A=1/3$ and ${(1-p_O^A)}^2/2=2/9$⁠). These results are summarized in the following propositions:

 

Proposition 1. The learning modality of a term is chosen more often at time 0 when learning is asymmetric as compared with symmetric learning.

 

Proposition 2. Switches away from the learning modality of a term occur more often at time 1 when learning is asymmetric as compared with symmetric learning and, vice versa, switches away from the no-learning modality of a term occur less often at time 1 when learning is asymmetric as compared with symmetric learning.

Note that these results crucially depend on the assumption that firms anticipate the advantages of learning, that is, that they willfully experiment, at least to some extent. If that were not the case, the firm would not factor in the option value of the asymmetric-learning modality, and hence the initial drafting choice between the default and the opt-out would be driven by identical short-term considerations both under symmetric and under asymmetric learning, and we would have $p^S=p_D^A=p_O^A$⁠. Therefore, with mere, passive experience, the difference between the probability of switches away from the learning modality of an asymmetric-learning term and the probability of switches away from the same modality of a symmetric-learning term would—all else equal—vanish.²⁷ Although overall revision rates are lower with experience as compared with experimentation, even with passive experience, there is still a difference in revision rates between the learning modality and the no-learning modality of asymmetric-learning terms and between the no-learning modality of an asymmetric-learning term and a symmetric learning term.

 

Proposition 3. If firms do not anticipate the possibility of learning, switches away from the learning modality of a term occur equally often at time 1 when learning is asymmetric as compared to symmetric learning.

3.4 Other Determinants of Contractual Choice

Let us now enrich the model with three additional ingredients. First, firms that opt out of the default term may incur opt-out costs, such as the costs of additional marketing and legal research, riskier litigation, and customer aversion to unfamiliar arrangements.²⁸ It is easy to see that these costs simply raise the value of the default, v.

Second, time-0 and time-1 profits might weigh differently on the firm’s time-0 choices because the expected volume of sales at time 1 might be greater or less than that at time 0. Let w > 0 be the weight of time-1 sales. Firms that expect to grow are characterized by w > 1, while firms that expect to shrink are characterized by w < 1.²⁹

Third, switching at time 1 entails a cost for the firm, denoted s ≥ 0, which captures the costs of rewriting the contract, doing additional legal research, informing customers, and so on. If the costs of switching are too high, then switching does not take place and learning becomes an uninteresting option. Therefore, we focus on cases in which switching costs are moderate:

 

Assumption 2: s < min{w(c_H – v), w(v – c_L)}

That is, the firm’s switching cost is less than the efficiency gains from switching in either direction. It is easy to verify (and it is proven in the Appendix) that the three adoption thresholds for the default term in the three learning scenarios defined in the previous section are now:

$${\displaystyle \begin{array}{l}{\displaystyle p^S\equiv \frac{v-c_L+s}{c_H-c_L+2s},}\end{array}}$$

(4)

for symmetric-learning terms,

$${\displaystyle \begin{array}{l}{\displaystyle p_D^A\equiv \frac{(1+w)(v-c_L)}{c_H-c_L+w(v-c_L)+s},}\end{array}}$$

(5)

for asymmetric-learning-from-default terms, and

$${\displaystyle \begin{array}{l}{\displaystyle p_O^A\equiv \frac{v-c_L+s}{c_H-c_L+w(c_H-v)+s},}\end{array}}$$

(6)

for asymmetric-learning-from-opt-out terms. The probabilities of switches are defined in the same way as above. From these values, we have the following results, which are formally proven in the Appendix.

3.4.1 Opt-out Costs

Opt-out costs add to the value, v, of choosing the default and make it less likely that a firm will opt out both at time 0 and at time 1. However, if the firm learns in the interim period, opt-out costs are irrelevant to the firm’s choice as long as Assumption 1 holds. If the firm does not learn, the firm simply confirms its time-0 choice irrespective of the opt-out costs. Therefore, opt-out costs are felt particularly at time 0, when the firm chooses under imperfect information, and make the choice of the default option more likely. In turn, this expands the set of firms that, having chosen the default, may learn and switch to the opt-out option at time 1. Proposition 4 formalizes these intuitions.

 

Proposition 4. If opt-out costs increase, then the default term is chosen more often at time 0 and there are more switches away from the default at time 1.

3.4.2 Growth Prospects

Growth prospects have two effects: not only do they magnify the (dynamic) gains from learning but they also increase the (static) net value of a term. Hence, an increase in growth prospects may or may not result in more learning, depending on which effect dominates. To elaborate, the two modalities of a term may or may not be symmetric with respect to their net values. If $v=(c_L+c_H)/2$⁠, we have v – C_L= c_H– v (the net values of the default and the opt-out are the same), and the firm’s contractual choices are driven purely by learning. In this case, growth prospects stimulate learning: firms with greater growth prospects choose the learning modality of a term more often at time 0 and, hence, also switch away more frequently from the learning modality at time 1. However, if the former condition does not hold true, we may have V – C_L > C_H-V (or vice versa). In this case, growth stimulates learning only if the net value of the learning modality is large enough. Otherwise, the result is reversed. Proposition 5 provides a general formulation of these results.

 

Proposition 5. If v is close to$c_L+c_H/2$, if the firm’s growth prospects increase, then the learning modality is chosen more often at time 0, and there are more switches away from the learning modality at time 1. This result is reversed if the net value of the learning modality is sufficiently small, that is, if v is either much greater or much smaller than$c_L+c_H/2$⁠.

An obvious corollary to Proposition 5 is that in the case of symmetric-learning terms, both modalities of the term imply learning, and hence the choice is purely driven by the term’s net values. In particular, greater growth opportunities result in more frequent choices of the default option if v – c_L > c_H – v, and of the opt-out if v – c_L< c_H– v.

3.4.3 Switching Costs

Switching costs add an implicit tax on learning, thereby making learning a less attractive option, restricting the choice of the learning modality at time 0 and, consequently, the frequency of switches away from the learning modality at time 1. This result, however, may be reversed if the learning modality also yields a smaller static net value, which, as above, weighs against learning. Proposition 6 formalizes these observations.

 

Proposition 6. If v is close to$c_L+c_H/2$, if switching costs increase, then the learning modality is chosen less often at time 0, and there are fewer switches away from the learning modality at time 1. This result is reversed if the net value of the learning modality is sufficiently small, that is, if v is either much greater or much smaller than$c_L+c_H/2$⁠.

As above, Proposition 6 implies that in the case of symmetric-learning terms, contractual choice is purely driven by the term’s net values. In particular, greater switching costs result in a higher frequency of choice of the default option if v – c_L< c_H– v, and of the opt-out if v – c_L > c_H– v.

3.5 When is Experiential Learning Most Likely to Occur?

3.5.1 Market Structure (Other Than Monopoly)

In the model, we focus on monopolistic firms. Considering firms with less than full market power would not qualitatively alter the results. Some degree of competition would reduce the firm’s ability to capture consumer surplus, thereby requiring us to distinguish between the (relative) value customers attach to the default, v, and the price increase that the firm is able to sustain when offering the default, which could be less than v when firms compete. This, however, would not alter our analysis, as we allow v to vary.

However, in a fully competitive market, prices track costs, not consumer surplus. Therefore, firms might adjust the price they charge to customers after learning the costs of different clauses. These adjustments may erode firms’ profits but should not affect the key mechanism behind our model: firms would still be induced to offer the cheapest option, which is unknown at the outset. Yet, a formal model of standard-form contracts in competitive markets might unveil additional implications.

Competing firms might also learn from each other, which both boosts learning—because it magnifies the effects of any individual firm’s experimentation with new clauses—and hinders it—because it creates a free-riding problem that reduces a firm’s incentives to experiment (see Bolton & Harris 1999). While this aspect of the problem would add a layer of complexity to the analysis, it would not affect our basic distinction among contract terms based on their learning characteristics and hence would not qualitatively alter our results.

Finally, competitive forces might induce firms to follow what most of their competitors do because, for instance, customers might be unwilling to buy a product that is offered with an unfamiliar set of clauses. We already consider the costs of opting out of a legally-set default term. A similar analysis could be applicable to the case of opting out of the industry-standard terms.

3.5.2 Heterogeneous Customers, Tailoring and Screening

Our framework applies to cases in which firms offer standard-form contracts to their customers (or suppliers). Yet, firms routinely attempt to tailor their contracts to the specific characteristics of individuals or groups of customers. Costs associated with offering a specific term may vary with customer characteristics. Moreover, different customers may value the same term (say, a warranty) differently. In these cases, firms might find it advantageous to tailor contracts to specific customers or customer groups rather than offering all their customers the same contract.

A widely studied way for a firm to tailor a contract to the specific characteristics of its customers is to “screen” customers by offering different contracts at different prices and letting customers choose their preferred contract (Stiglitz 1975; Akerlof 1970).³⁰ However, firms in our model cannot do so because customers are uniform with respect to the value, v, that they attach to the default term and ignorant about the costs that alternative terms impose on the firm.³¹ If offered different contracts, customers in our model would all choose the same option, defeating the firms’ attempt to screen among them.³²

While we make this assumption to shut down screening and focus on learning, we think this is a broadly realistic approach. First, consumers often know less, rather than more, than firms do about their own future use patterns, exposure to risk, probability of accidents, and other important factors that determine the costs for the firm of offering different terms (Schwartz & Wilde 1983; Bar-Gill 2012). Second, contract standardization offers numerous advantages to firms, which would be lost if the firm were to tailor the contract to individual customers, making standardization advantageous even in cases in which tailoring would be theoretically possible. Empirical evidence confirms this. For example, Della Vigna & Gentzkow (2019) find that retail chains across the United States do not adapt their prices to easily identifiable groups of consumers living in different states with markedly different preferences, wealth, education, and, ultimately, willingness to pay for certain products. Price terms are allegedly the easiest terms to vary in a contract; the fixed costs associated with tailoring other, more difficult-to-individualize, terms might be prohibitive as well.

Finally, even if firms were able to discriminate among customers depending on their use patterns, there is no guarantee that they would have incentives to do so perfectly. In a recent study, Hua & Spier (2020) show that a monopolist may fail to supply the optimal level of product warranties even when it can price discriminate among consumers. The intuition is that the monopolist will tend to cater to the interests of the marginal consumer, which is typically not representative of the population of consumers. Therefore, the possibility to learn may also alter the behavior of price-discriminating firms. By abstracting from price discrimination, we zero in on the important details of the analysis and offer insights that may be more generally applicable.

3.5.3 Uncertainty About Value Versus Uncertainty About Costs

One of the building blocks of our model is that the value of offering particular terms is known while its costs are not. While there is nothing in our learning mechanism that hinges upon this assumption—and hence the analysis would carry on unchanged if we reversed it—allowing for uncertainty about value would introduce additional complications into the model because the value is known to customers, and hence the firm can adjust the price of the product next to amending the term.³³

In a simple two-type model where the (homogenous) customers’ valuation can be either high (and hence greater than the cost of offering the term) or low (and hence less than the cost), the firm should offer the term only if the valuation is high.³⁴ If the firm does not offer the term, the firm does not learn how customers value it. If the firm does offer the term, it will offer it at a high price, that is, the price corresponding to the high valuation. (Offering the product at an intermediate price would simply reduce revenues without improving learning, and offering the product at a price equal to the low valuation would both reduce revenues and prevent learning.) If sales are positive, the firm learns that customers have a high valuation, if they are zero, the firm learns that customers have a low valuation. Therefore, by observing the purchasing behavior of customers in response to the particular contract that the firm offers, the firm is able to learn the customers’ valuation. It is easy to see that also in this model, there is an option value in offering the term in period 1. Hence, with uncertain valuations, all terms would be reclassified as asymmetric-learning terms.

A richer model could allow heterogeneous customer valuations over a continuum and, in such a model, firms may have more room for adjusting the price in response to learning, giving a certain offer of contractual terms. Thus, next to contract amendments, we should observe price adjustments. Yet, the main intuitions provided above would remain valid. The driver of the difference between learning about costs and learning about value is the fact that customers are likely to know their valuation of a particular term while they are unlikely to know its costs. Our analysis implies that firms will use the kind of experiential learning we describe to learn primarily about costs and rarely—with the caveats expressed above—about value.

3.5.4 Many Ways to Opt Out and Continuous Learning

The model only considers one alternative to the default option, while in reality there may be many. With many alternatives to the default term, the firm not only faces a decision of whether to opt out but also has to choose among many possible terms, each of them with possibly different feedback mechanisms. Learning in this context becomes more complex and may bring about interesting interactions. After learning that, say, alternative 1 has high costs, the firm may decide to switch back to the default or start experimenting with, say, alternative 2, and so forth. Moreover, while we allow only one learning period, in reality, the firm might learn continuously and be able to switch between one term and another at several time periods, possibly going back to terms it had discarded in the past. This more general approach would be close to a version of the well-known “multi-armed bandit problem” in probability theory (Lai & Robbins 1985). While there is a large literature on problems of strategic experimentation, to the best of our knowledge, this literature has not dealt with learning in contracts as we do.

3.5.5 Endogenous Investments in Learning

In the model, we take the probability of learning as exogenous. However, the firm might be in control of these probabilities by, for instance, deciding to experiment with a term on a subset of its customers. By doing so, the firm may reduce short-term losses in exchange for a reduced probability of learning (because of more limited feedback). While this is an interesting extension of our model, it is unlikely to affect our results: after the firm has set its learning strategy, our model would carry on virtually unchanged. In addition, the extent to which firms can treat similar customers differently is limited both by law and by reputational concerns, as Delta and Amazon recently found out.³⁵

3.5.6 Reasons Not to Revise a Term After Learning

The theoretical framework presented above focuses on learning direct costs, but the choice of terms can generate other forms of learning that affect contractual choices at a later stage. Consider, for example, a retailer that sells products manufactured by a number of suppliers and is uncertain about the quality of the products of each supplier. Offering a secondary warranty to customers could be a way to obtain feedback on the quality of the firm’s suppliers. If the product breaks down frequently, the firm learns that one of its suppliers delivers low-quality products. The interesting implication is that, in this case, the firm’s response to learning is a change of supplier rather than a change of term. The firm may want to keep offering the warranty in order to learn about the new supplier too. We do not elaborate on this alternative learning motive, but we stress that this is also a form of experiential learning. Conversely, a choice-of-law clause may or may not be desirable depending on whether it lowers or raises the costs of litigating a case in court, which in turn depends on unknown factors, determining whether the firm faces high or low costs. Learning about these costs may induce the firm to amend the clause at a later time. Our analysis applies to these cases.

4. EMPIRICAL ANALYSIS

We now put our theory to bear on the contractual choices made by real firms. We first derive empirically testable predictions from our theory. Next, we present our data set and empirical results.

4.1 Empirical Implications of the Theory

While it is difficult to empirically disentangle the reasons behind firms’ adoption of terms in an initial period (given the multitude of factors likely affecting such decisions, many of which are hard to measure), examining firms’ decisions to revise such terms at a later period can offer some interesting insights regarding possible, though not conclusive—given the nature of the data—drivers of contractual choice. We explore learning from previous contractual choices against the attractiveness and stickiness of default terms. Limitations in data availability allow us to test only a subset of the model’s predictions following from Propositions 1 to 6, which we restate below.

 

Prediction 1. The probability that a firm will amend an asymmetric-learning term at time 1 is higher if the firm has chosen the learning modality of this term at time 0.

The firm’s decision to revise an asymmetric-learning term is largely affected by the firm’s choice at time 0. Adopting a term in its learning modality at an initial period allows the firm to reevaluate past contractual choices and amend it if new information suggests that a different choice is more advantageous. Prediction 1 also identifies a mechanism by which “black holes” could come about. If the firm has chosen a nonlearning modality at time 0, it will become more likely that the firm will not see new information (at least information garnered from experience) and might fail to revise the term in question at time 1. Inefficient or meaningless terms might survive due to the asymmetric nature of learning. In contrast, firms adopting the learning modality of the same term at time 0 will have an increased opportunity to learn the effects of such terms from experience and eventually stay away from such inefficient or meaningless terms. Such “black holes” or pockets of inefficiency might affect only a portion of the firms in the market, especially when other learning channels play weak roles.

 

Prediction 2. The probability that a firm will amend a symmetric-learning term at time 1 does not depend on the term chosen at time 0.

Contrary to asymmetric-learning terms, here the firm’s initial choice does not affect the firm’s propensity to revise the term. For these terms, experiential learning—and learning from other channels—occurs (or does not occur) irrespective of the contractual choice at time 0. If the firm learns from experience, learning will occur symmetrically from both the default and the opt-out option. We should observe revisions motivated by experience as well as other learning channels in this case, but such revisions should be equally likely for firms that adopted the default and firms that opted out of it at time 0. The same is true when firms learn from other means. Revisions of the term at a later date will be uncorrelated with the contractual choices made during the earlier period.

 

Prediction 3. If default terms are often inefficiently chosen at time 0, default terms will be amended more frequently than non-default terms if they offer an opportunity to learn.

Default contractual terms have long been recognized as important determinants of contractual choice. Implications of this observation come in two guises. On the one hand, if default terms are more frequently chosen, this observation could apply both at time 0 and at time 1. If, however, the choice of a term is largely determined by the term being a default, default choices at time 0 are more likely to result in inefficient outcomes. We anticipate that such defaults will be more likely to be amended at time 1 if the firm has had an opportunity to learn in the meantime, all else equal. This effect should be visible both in symmetric and asymmetric-learning terms. In symmetric-learning terms, the learning terms will be revised at time 1 more often toward the opt-out option if the default was inefficiently chosen at time 0. In asymmetric-learning terms, revisions should be more frequent when the default is the learning modality than when it is the nonlearning modality.

Both implications point to an important role of default contractual terms in determining firm choices going forward. If this is the case, switches at time 1 should be largely explained by the fact that a term at time 0 was an inefficient choice of the default. This prediction will allow us to contrast defaults to learning as alternative explanations for the change in standard-form contracts.

 

Prediction 4. If firms willfully experiment with contract terms, a default term is revised more frequently if the default is the learning modality of an asymmetric-learning term than if it is a symmetric-learning term. Similarly, for an opt-out term.

If learning is symmetric, the initial drafting choice is fully determined by short-term considerations, as learning occurs anyway. Instead, if learning occurs asymmetrically, the opportunity to learn gives an additional reason to choose the learning modality of the term, pushing for more frequent adoption at time 0. In turn, this results in more frequent reversals in the following period. Crucially, these differences emerge only if the firm anticipates the opportunity to learn and shapes its initial drafting decisions accordingly, that is, if the firm willfully experiments. If learning is serendipitous, we should observe no difference in revision rates between symmetric-learning terms and asymmetric-learning terms in their learning modality. Therefore, this difference offers some insight into whether firms experiment. We turn to the empirical analysis in the next section.

4.2 Data and Methodology

We test our hypotheses using a sample of software license agreements governing the use of prepackaged software. EULAs typically present a rich set of standard terms; while the terms typically vary both across and within markets, EULAs follow a predictable structure (Marotta-Wurgler 2007). As reported in detail in Marotta-Wurgler (2007) and Marotta-Wurgler & Taylor (2013), the terms that we track have been identified as important in several independent industry and trade references and textbooks. The terms fall into ten categories of related terms, including notice of acceptance of the license; scope of the license; restrictions on transfer; warranties and disclaimers of warranties; limitations on liability; maintenance and support services; modification and termination of the license; information collection; third-party access to users’ computers; and conflict resolution. Each category includes terms related to particular aspects of the transaction. For example, the category related to limitations of liability includes terms that allocate losses and risk between buyers and sellers. We focus on the individual terms within each category, even if terms are related, because each term generates distinct information about its costs and, more important because terms within each category differ in their ability to generate experiential learning, both in general and within a particular modality. When firms include terms allocating liability and loss to themselves, they have an opportunity to learn about their costs.

Almost all EULAs include these exhaustive categories of terms. This allows for meaningful comparisons across contracts. We examine the rate of change of terms from 2003 to 2010 in accordance with sellers’ opportunity to learn from experience with the terms offered.

We use the sample of EULAs introduced by Marotta-Wurgler & Taylor (2013), which tracks the changes in the terms of EULAs found in typical “prepackaged” (i.e., non-customized) software products and compare their content in 2003 and 2010. That study examined the change in 32 EULA terms from 246 firms that sell their software on their corporate websites, including large, well-known software publishers, as well as smaller companies. For each of the companies, the dataset includes a representative product along with data on various market, product, company characteristics, and of course, the EULA both in 2003 and in 2010.

For each EULA in each period, we tabulate the presence of 32 standard terms across the ten categories of related terms. We further classify each term into categories reflecting the extent to which offering a given term gives sellers an opportunity to learn directly from experience, either symmetrically or asymmetrically. We also take account of other factors that might affect firms’ decisions to revise terms at a later time, such as their size, age, and whether they have in-house counsel.

4.2.1 Summary Statistics

Table A1 presents summary statistics. Panel A reports company characteristics for the sample firms. Average revenue in 2003 was $287.5 million and the median was $1.7 million. Average and median revenue in 2010 were $539.1 million and $2.2 million, respectively. The percentage of public companies grew from 11 percent in 2003 to 14 percent in 2010.

Table A1.

Company, product, market, and contract characteristics

The sample includes data on legal sophistication in 2010, proxied by firms’ choice of legal advice, including whether they have in-house counsel, at least one internal lawyer, or routinely hire outside counsel. All public companies are assumed to receive sophisticated legal advice. In total, 74 percent of firms for which this data was available received relatively intensive legal advice, which might affect firms’ propensity to revise terms at a later date.

Panel B lists product and market characteristics in 2003 and 2010. The average price of the products in the sample was $812 in 2003 and $841 in 2010. Thirty-six percent of the products are oriented toward consumers or small home businesses rather than large businesses. One percent of the products in the sample was discontinued, but the company used the same EULA for all their products in 2003 and 2010. Save for some updates, the products and their functions were essentially the same at the beginning and the end of the period. Firms are classified into 114 distinct software markets, as classified by Amazon.com, the largest Internet software retailer.³⁶

Panel C reports contract characteristics. We first record whether at least one of the thirty-two terms we track was revised in any way during the sample period. Of the entire sample, 40 percent of contracts changed at least one substantive term. Of the 103 contracts that had at least one change (39 percent of 264), change was limited to one or two terms, but a few firms changed their contracts significantly, including some that changed more than ten terms. Contract length increased from 1,517 words in 2003 to 1,938 in 2010, or an average of 27 percent. The median word increase in contracts with no material changes was one word, whereas the median word increase in EULAs with material changes was 435 words.

4.2.2 Classifying Symmetric- and Asymmetric-Learning Terms

We classify the thirty-two terms into four categories that reflect drafters’ opportunity to learn from experience. Each term is described in detail in Marotta-Wurgler & Taylor (2013), and its presence is measured against the benchmark of the default rules of UCC Article 2. We note if a term matches the default rule provided in Article 2 (given that such rules would fill any gaps to the extent a contract is silent on a given issue) and if a term deviates or opts-out of such default rule. A contract can adopt the default rule either by including a term that matches such rule or by remaining silent. While classifying terms that allow sellers to learn from experience is subjective, the specific purpose and characteristics of each term make this exercise relatively straightforward, especially for important terms such as warranties and terms related to damages. Of course, reasonable minds can disagree on the characterization of a subset of the terms. We address this in two ways. Firstly, we ran a series of robustness exercises, including reexamining the data after coding those terms that may be more susceptible to disagreement in their alternative characterization and focusing on the most agreed-upon characterizations, among others. Secondly, we supplemented our empirical analysis with open-ended interviews with in-house counsel who shared their accounts of experiential learning from various terms. These accounts allowed us to further corroborate our classification for a subset of terms. We present these classifications in Table A2 and our reasoning explaining these decisions in Table A6.

Table A2.

EULA Terms and Bias: 2003 vs. 2010

EULA terms are classified into 32 common terms that allocate rights and risks between buyers and sellers across seven categories of related terms, according to the degree the terms either match the default rules of UCC Article 2 (Adoption of Default = 0) or deviate from them (Opt-out = 1). “Learning Category” refers to the type and modality that allows sellers to learn from a term. Terms allow for symmetric learning, denoted S, when learning occurs or not regardless of the modality of the term. Some terms allow for asymmetric learning, allowing sellers to learn as long as the modality adopted enables learning. Terms that enable learning when the seller adopts the default rule but not otherwise are denoted A (D) (i.e., asymmetric learning by adopting the default). Terms that enable learning when the seller opts out of the default are denoted A (O) (i.e., asymmetric learning by opting out of the default). The table reports the mean opt-out of UCC Article 2 default in 2003 and 2010, as well as the mean change and statistical significance. ^*p < 0.10, ^**p < 0.05, ^***p < 0.01.

Table A6.

Term classification

EULA terms are classified into 32 common terms that allocate rights and risks between buyers and sellers across seven categories of related terms, according to the degree to which the terms either match the default rules of UCC Article 2 (Adoption of Default = 0) or deviate from them (Opt-out = 1). “Learning Category” refers to the type and modality that allows sellers to learn from a term. Terms allow for symmetric learning, denoted S, when learning occurs or not regardless of the modality of the term. Some terms allow for asymmetric learning, allowing sellers to learn as long as the modality adopted enables learning. Terms that enable learning when the seller adopts the default rule but not otherwise are denoted A (D) (i.e., asymmetric learning by adopting the default). Terms that enable learning when the seller opts out of the default are denoted A (For each term, the Table reports the rationale support a particular experiential learning classification, as noted in the last column, “Classification Rationale”.

Not all terms give sellers the same opportunities to learn from direct experience. The first column of Table A2, labeled “Learning Category”, reports how we classify each term depending on whether some terms allow for symmetric learning (or failure to learn), whether directly or indirectly, or whether learning is asymmetrically tied to the seller adopting the default rule or opting out of it. As noted earlier, symmetric-learning terms might allow firms to learn about the cost of offering such terms from experience under all modalities of the term or under none of them. Sometimes learning from experience with that term is not the most direct form of learning for a particular term. The mechanism by which sellers learn, whichever it may be, will not depend on the initial choice of a particular modality of the term.

The table labels such terms as “S”, that is, symmetric learning. We identify fourteen such terms. One term gives notice to customers that they can return the product if the customers reject the terms. These notices must be present when the terms are available only post-purchase to give the customer an opportunity to review the terms. The purpose of the term is to comply with contract formation requirements when the terms appear in a “pay now, terms later” format.³⁷ The seller might want to change the manner in which the contract is offered, whether pre- or post-purchase, but such knowledge is likely to arise in either presentation format. Another term in this group relates to limitations on transfers. Sellers might consider revising the scope of this right by evaluating customer demand, studying what competitors do, or through direct feedback from customers related to their transfer rights preferences, whichever those are at an initial period. Two terms in this group are related to dispute resolution: the seller is able to experience whether the chosen law (or the failure to offer one) and who pays for attorneys’ fees are optimal. For these terms, the seller learns from experience in those cases where the terms are invoked, regardless of their initial modality (e.g., the seller will learn whether having or not having a choice-of-law clause is desirable in case of a dispute, regardless of whether one was included in the contract when the dispute arose). Another allows the seller to disable the software remotely in case the buyer breaches. Again, regardless of its modality, a seller learns whether it is desirable to have such a clause (assuming it is feasible for the seller to offer it) whenever the seller experiences a buyer breach. This clause makes enforcement of the contract easier via extra-legal means. Feedback, through various means (including the types of breaches that may warrant the need for remote disabling), can occur whether the term is used in an initial period or not.

Additional terms, where the seller may or may not learn from experience in all term modalities, include a change-of-terms clause clause that allows the seller to unilaterally amend the contract; one term noting whether the licensed product includes updates or upgrades; one term delineating the scope of the use rights granted by limiting the buyer’s ability to modify or alter the program; three terms explaining whether there are transfer limitations or other license grant restrictions; one noting whether the disclaimer is in caps or otherwise conspicuously presented (this is not a term per se, but one that tracks a requirement under the Magnuson-Moss Warranty Act); two terms related to the rights of third parties; and one term informing customers of their statutory rights outside the contract. Please refer to the Appendix for the explanation behind such classification.

We now turn to asymmetric experiential learning clauses, labeled as “A”. We separate these terms into those where sellers learn by direct experience only when they opt out of the default (or “A (O)”) and when then they learn only when they adopt the default (or “A (D)”). Narrowing the classifications in this manner allows us to examine the relative attractiveness of default rules. There are no default express warranties, so the seller learns only by opting out of the default. The L.L. Bean warranty is an example of such a clause. We identify five asymmetric “learning by opt-out” clauses. These include one term allowing the drafter to install software to monitor users’ activities, three terms tracking whether the seller offers limited or full warranties, and one tracking whether the software includes maintenance and support services (here, the seller is not obligated to do so unless it is promised in the contract; such promise exposes the seller to customer demands).

In contrast, if the seller offers default implied warranties, it might learn the value of such an offering. In this case, adopting the default allows the seller to learn. We find twelve such terms. These include two clauses allowing the buyer to create derivative works and reverse engineering (which are allowed under intellectual property law), a choice of forum clause (where the seller learns the costs of not providing one if sued in an inconvenient forum; being sued in its forum of choice might lull the seller into not considering this aspect of the costs related to the dispute), as well as nine clauses disclaiming implied warranties, various risks, or damages.

For each term and category of term, Table A2 reports the mean opt-out from the relevant default rules in both 2003 and 2010, as well as the mean change during the sample period. For example, in 2003, 55.3 percent of firms included a term capping damages at less of or equal to the purchase price, a term we classify as A (D)—which our hypothesis predicts sellers would be more likely to revise in the later period if they initially offered the learning modality of the term. This number decreased slightly in 2010, to 51.9 percent of firms choosing to opt out of the default rules. The difference of 3.4 percent, while small, is statistically significant at the 10 percent level.

4.3 Analysis

We now explore the extent to which the changes reported in Table A2 are more likely depending on the initial choice of terms as well as when sellers have an opportunity to learn. The nature of our data prevents us from making any inferences regarding the initial choice of terms, as these are also likely the result of past contractual choices. We can, however, measure the extent to which default rules are predominantly chosen and measure the extent to which these are revised in a later period. Panel A in Table A3 begins by exploring the stickiness of default rules in the data by reporting the extent to which sellers chose to match the default rules of the UCC at the initial period as well as the probability of revising a term given their initial modality in the previous period. The top right figure shows that among 32 terms in total, and 8,448 EULA-term observations, 30.8 percent of all terms in 2003 were at the opt-out value, whereas the remainder, or 69.2 percent, matched the default rules, indicating a strong gravitational pull toward the default previously identified in the literature.

 Table A3.

Learning and changing terms

Fraction of terms that change between 2003 and 2010 depending on whether their 2003 values are at the default or, for asymmetric terms, at the learning value. In Panel A, for example, 29.4 percent of terms were at opt-out values in both 2003 and 2010 and 1.4 percent were at a opt-out value in 2003 and changed to a default value by 2010. The probability of change for a term that was at a opt-out value in 2003 is 0.045 (0.014/0.308), while the probability of change for a term that was at the default in 2003 is 0.056 (0.039/0.692), which is a statistically significant difference of −0.011. Asymmetric terms can also be at a learning or nonlearning value. ^*p < 0.10, ^**p < 0.05, ^***p < 0.01.

Yet default terms are not set in stone. In 2010, the fraction of terms that match the default decreased to 66.7 percent. Indeed, 65.3 percent of all terms were at default values in both 2003 and 2010, but 3.9 percent were at default values in 2003 and were opted out in 2010. In terms of probabilities, the right panel shows that the probability of changing a term in 2010, given that a term was in an opt-out, and the default value in 2003 was 0.045 and 0.056, respectively. The 0.011 difference is statistically significant at the 5 percent level. While terms are more likely to begin at the default, the probability that they will be revised at a later period is larger if the term starts at the default, offering support to the known view that sellers might be inefficiently choosing default terms in the initial period due to opt-out costs.

With this baseline in mind, we test Predictions 1 and 2 by dividing the data by whether the term generates symmetric or asymmetric experiential learning opportunities, using the classifications presented in Table A2. Panel B in Table A3 presents data on symmetric learning. As noted earlier, sellers might be learning about the relative value of these terms through other means, independent from experience and irrespective of whether the term matches the default rule or not. We have no a priori hypotheses as to how these additional sources may inform sellers. For our purposes, all we care to know is whether change is more likely to be associated with one experiential learning modality of the term or the other.

The results show that, again, defaults are powerful determinants of contract terms in the initial period. In this case, 75 percent of symmetric terms match the default rule in 2003, only to change to 72.4 percent in 2010, indicating some change away from defaults. More interesting for our purposes, however, is the probability of change conditional on the starting point. Recall that we predicted that the starting point for these types of clauses would be a poor predictor of change. In fact, the probability of changing a term is precisely the same, or 5.2 percent depending on where the term is in 2003.

Contrast this with Panel C in Table A3, the results for asymmetric terms. In 2003, 64.2 percent of all such terms matched the default rules of the UCC, a number that shrank to 61.8 percent in 2010. The right panel shows that the probability of change for terms that matched the default in 2003 is 6.1 percent, in contrast to 4.2 percent for non-defaults. The difference is significant at the 5 percent level. Even for the asymmetric learning clauses—and consistent with the findings in Panel A examining all terms—terms are more likely to be revised when they start at the default rule, regardless of the learning modality.

Once we divide asymmetric terms up into their learning modalities, a new picture emerges, as seen in the bottom panel of Panel C. In 2003, asymmetric terms are included in their learning and nonlearning modalities about equally. However, in contrast to the symmetric terms, where the probability of changing a term was independent of the original allocation of the term between default and nondefault, in the asymmetric scenario, the original learning modality matters. The probability of changing a term, given that the 2003 contract included such a term in its learning modality, is 0.072, in sharp contrast to the 0.034 that occurs when the term is not in its learning modality. The results also support the prediction that asymmetric-learning terms adopted in their learning modality at an initial stage are more likely to be revised than symmetric-learning terms (7.2 percent vs. 5.2 percent, respectively); while the reverse is true for asymmetric-learning terms adopted in their nonlearning modality at the initial stage (with a 3.2 percent revision probability). The findings support the basic prediction that the opportunity to learn from experience with a term helps explain contractual change and innovation.

That being said, while the findings are consistent with our theoretical predictions, observational data can never prove that learning is one of the causes of changes in terms nor conclusively disqualify competing alternative hypotheses. Section 5 complements this large-sample evidence with in-depth interviews with general counsel that offer more nuanced, descriptive views of why standard terms change over time that are consistent with our learning account.

These findings are illustrated in Figure 2. The left bars show the probability of change conditional on their 2003 starting point (default versus opt-out). The bars are the same height, consistent with the modality of the term conferring no learning advantage. Contrast this to the bars on the right. Change is more likely to happen if the terms are switched on their experiential learning modalities in 2003, as opposed to their nonlearning modality.

Figure 2

Probability of term change.

Table A4 reports ordinary least squares regressions including company, product, and market control variables. The first column just repeats the results from the bottom of Panel C of Table A3. The second column adds firm (contract) fixed effects, controlling for the overall propensity of a given contract to change. The fact that the coefficient on learning does not budge indicates that there is not a tendency for some firms to make wholesale changes to their policies, including their learning terms; a given learning term is equally likely to change “within” a contract whether the same firm is changing many or few other terms. The third and fourth columns show that the probability of changing away from a term at the default in 2003 is robust to the overall propensity to change the contract, but the effect is only half that of the probability of changing the term as a function of the term’s learning status and is a distinct effect. Logit regressions yield very similar results and are omitted for brevity.

Table A4.

Learning and changing terms: robustness

	(1)	(2)	(3)	(4)	(5)	(6)
	Change	Change	Change	Change	Change	Change
Learning	0.0392*** (0.00920)	0.0402*** (0.00801)		0.0394*** (0.00815)	0.0401*** (0.00984)	0.0420** (0.0145)
Default			0.0187** (0.00818)	0.00204 (0.00831)	0.0003 (0.00958)	0.0138 (0.0152)
Multi-user license					−0.0417*** (0.0147)	−0.0778*** (0.0173)
Developer license					−0.0104 (0.0280)	−0.00121 (0.0328)
Ln price					0.0103 (0.00627)	0.0338** (0.0128)
Change Ln price					0.0497** (0.0223)	0.0647 (0.0404)
Consumer product					0.00400 (0.0159)	0.0376 (0.0265)
Ln revenue					0.00393 (0.00348)	−0.000247 (0.00564)
Change Ln revenue					0.0219*** (0.00662)	0.0290*** (0.0100)
Ln Age					0.00122 (0.0117)	0.0142 (0.0214)
Lawyers						0.0611* (0.0329)
Pro-consumer state					−0.00448 (0.0110)	−0.0298 (0.0198)
H−H index					0.0279 (0.0247)	0.0217 (0.0377)
Constant	0.0337*** (0.00533)	0.0332*** (0.00399)	0.0412*** (0.00525)	0.0323*** (0.00588)	−0.0757 (0.0507)	−0.246** (0.0996)
Fixed effects	None	Firm	Firm	Firm	None	None
Observations	4,488	4,488	4,488	4,488	3,791	1,751
Adjusted R2	0.007	0.160	0.154	0.160	0.026	0.050

The sample is asymmetric terms only in 264 contracts. Least squares regressions where the dependent variable is a 0-1 indicator that the term changed between 2003 and 2010. Learning means that the term was set at a learning value in 2003. Default means that the term was set at the default in 2003. Standard errors in parentheses are clustered by firm. ^*p < 0.10, ^**p < 0.05, ^***p < 0.01.

The last two columns add a variety of potentially interesting control variables but with no effect on the learning coefficient of interest. Note that fixed effects cannot be included here because the variables do not vary within a given contract. We see that multi-user licenses are less likely to change. One hypothesis, which we cannot test, is that such licenses were, in general, given more thought in the first place. It also appears that when the firm is selling increasingly expensive products, its contract terms are more likely to change. Positive changes in revenue, which could be interpreted as a proxy for growth prospects, are also associated with increased learning, as discussed in Section 3.4. Finally, the presence of lawyers is associated with change, suggesting that lawyers might be part of the mechanism by which experiential knowledge generates a change in standard terms. Our interviews with firms regarding the mechanisms for revising contracts offer some support for this view, although we only spoke to in-house counsel.

Table A5 presents some refinements by dividing asymmetric terms into whether the learning modality is at the default or at opt-out. It repeats the exercise in Table A3 and reveals that, when learning occurs by keeping the default, firms are more likely to include the term at the initial period (59.9 percent, as compared to 40.1 percent, as seen in the left portion of Panel A). This is not the case when learning occurs at opt-out (where only 25.5 percent of such terms are operationalized in their learning modality), as noted in Panel B. The latter might be the result of the stickiness of defaults. Change in the later period, however, is more likely when terms are set in their learning modality in their initial period, regardless of whether learning occurs at the default or at opt-out, consistent with our prediction. The right-hand side of Panel A shows that when learning occurs at the default, terms that were offered in their learning modality in 2003 had a 7.3 percent probability to change, compared to 3.2 percent of terms that were in their nonlearning modality. The difference is significant at the 1 percent level. The same is true for terms where learning occurs from opt-out. These are 7.1 percent likely to change when offered in their learning modality, compared to 3.5 percent when they are not. Again, the results are highly statistically significant. Note that the results in Panel A support Prediction 3, which states that terms are more likely to be revised from inefficiently chosen defaults when such defaults carry an opportunity to learn. They are also consistent with Prediction 4, which posits that firms that purposely experiment with contract terms will be more likely to revise asymmetric-learning terms where learning occurs at the default, relative to terms where learning is symmetric, because firms will be more inclined to include such terms in those modalities in an initial period. The probability of changing an asymmetric term where learning occurs at the default is 7.3 percent, which is higher than the probability of changing a symmetric default (which is 5.6 percent). This finding is only suggestive, however, because we do not know whether sellers chose their initial terms for purposeful experimentation and because of other factors, such as the possibility that no learning occurs from a symmetric-learning term, may have contributed to this outcome. Given this, it is reassuring that the differences in revisions between symmetric- and asymmetric-learning terms also hold when learning occurs at opt-out.

Table A5.

Asymmetric learning by default vs. opt-out

Rate of learning values chosen for asymmetric terms, where asymmetric terms are broken down into those where learning is by adoption of the default rules of UCC and those where learning is by opting-out of such default rules.

4.4 Robustness Checks

Some terms are easy to classify in terms of symmetric/asymmetric learning, but others are not as straightforward. In a series of robustness checks (unreported), we try switching the classification of those terms whose classification is more ambiguous, such as a term stipulating a forum selection clause (which we classified as symmetric because firms will learn the cost of having one versus not having one during the litigation; but one could reason that not specifying a forum may result in more experiential learning) and found that our results are robust to these switches. (See Table A6 for all term classifications.) Reassuringly, the terms more likely to change when set in their learning modalities are those that are straightforward to classify as asymmetric experiential learning terms. These include terms that address the buyer’s ability to reverse engineer the product and those that determine the sellers’ implied warranty obligations and liability for consequential damages of the buyer. These exercises give some comfort that our conclusions are not sensitive to minor tweaks in coding.

4.5 Discussion

While the stickiness of default rules is apparent from the findings, the results support the hypothesis that experiential learning, whether purposeful or naïvely acquired, plays a role in how standard-form contract terms change over time. Our study focuses on a particular setting—business and consumer EULAs—but the learning mechanism we present could be present and examined in other markets and settings, involving other types of contracting parties. For example, we would expect to observe even more experiential learning in insurance markets, given the nature of the terms and the stakes at issue.³⁸ Given the nature and limitations of our data, our empirical analysis is suggestive and meant to illustrate the theory. We find additional examples of experiential learning as revealed in interviews with in-house counsel in different markets, but further research could explore causal stories in depth.

Importantly, experiential learning is not just limited to terms. Firms can experiment with product and service features that lend themselves to experiential learning. This real-option value can drive firms’ decisions to purposefully offer both terms and/or product features that they can later revise.

Of course, there are competing hypotheses that could explain firms’ desire to revise terms, such as opting out of sticky or obsolete defaults. After all, defaults may be chosen inadvertently and revised later, once firms become aware of their contractual effect. This is not inconsistent with our hypotheses: experiential learning from a term at an initial period can also lead to later revisions, even though such learning was neither planned nor anticipated. Default terms tend to benefit consumers, so sellers in consumer contracts might opt out of defaults in a later period to allocate part of the surplus to themselves. Assuming all (or most) default terms benefit consumers relative to the opt-out, then we would expect a shift away from all defaults with the same frequency. Yet we do not see this (unreported). Rather, it is those defaults that carry an opportunity to learn that get revised more frequently. Regardless of whether they are symmetric or asymmetric in terms of their learning modalities, defaults tend to benefit consumers (e.g., a contract without a choice of law clause gives consumers more options of where to bring suit; similarly, a contract that does not include restrictions on the consumer’s ability to modify the software is also more beneficial to consumers, all else equal). Yet only those that carry an opportunity to learn are more likely to be revised both in our consumer-facing contracts as well as in our business-to-business contracts. This depends on the assumption that all consumer-friendly defaults are similarly pro-consumer. If the consumer-friendly default rules of asymmetric-learning terms are more pro-consumer than those of symmetric-learning terms relative to their opt-out, we may see a higher rate of revision for the former terms because of these higher costs. Although we have no reason to conclude that these specific differences exist, we cannot rule them out. Of course, both motivations could co-exist. Not only could sellers be opting out of consumer-friendly defaults with the motive of drafting more self-serving contracts but they could also be more likely to revise those terms and term modalities that are associated with learning.

In addition, the changes we observe could, in theory, be explained by a general tendency to revise terms in a particular direction. Yet, we established earlier that experiential learning had its own effect despite general tendencies to revise contracts. Learning may also occur without being reflected in subsequent contractual choices. Consider a seller who decides to offer particular terms to learn about a supplier’s reliability. Such sellers would respond to the new information by switching suppliers and not necessarily terms. As we mentioned earlier, sellers could take several actions as a result of learning other than revising the contract itself. This could be the case, and it would work against finding any changes in the contract. Yet, terms get revised, and those that are susceptible to experiential learning appear to be revised more frequently, all else equal. Finally, a note on learning. It would be natural to expect in many cases for experiential learning to be sequential, where sellers learn by trial and error with different iterations. Our findings are consistent with this version of learning as well.

5. INTERVIEWS WITH IN-HOUSE COUNSEL

To further understand the role of experiential learning in revisions of contract terms in different markets and to seek external corroboration of our empirical approach, we conducted a series of open-ended interviews with in-house counsel from Google, Amazon, Fiat, Swapfiets, and a large media company (whose counsel preferred to have the firm remain anonymous), in 2019 and 2020, who explained the mechanisms by which experiential learning resulted in subsequent contractual revisions. The questionnaire used for the interviews can be found in the Appendix.

All counsel interviewed underscored the need to use standardized agreements to manage scale, leaving little scope to experiment with different terms for different parties. The in-house counsel at Google, who was in charge of contracts governing the firm’s relationships with advertisers, explained that their contracts with advertisers were partially standardized: the firm offered core terms to all ad providers for scale and management purposes and left other terms, like those related to billing and financial terms, commitments to service-level agreements, and data usage, open for negotiation.³⁹ The core terms were continually revised. The in-house counsel at Amazon, in charge of licensing and intellectual property concerns, also stressed the importance of standardizing terms.

The attorneys discussed how the anticipation of revisions, as well as learning, both direct and indirect, affected their contracting decisions both at an initial stage and later on.⁴⁰ Amazon’s counsel stated that most agreements are purposely drafted in general terms to reduce costs of ex-post revisions once the needs arose. The large media company’s counsel stated that some terms were drafted in a “future-proof” way by making them more general and thus easier to adapt to new environments, but such terms were in the minority, given their desire to be precise about the scope and type of services offered. The counsel at Swapfiets, a young Dutch long-term bicycle rental startup that includes repair and replacement services, explained that the firm started with a simple and “all-inclusive”, untailored, contract that would enable the firm to see which terms work in terms of not generating high costs to the firm or not yielding high value from customers.⁴¹ Similarly, when opening in new markets, the firm uses the contract and rates it uses in other markets and only upon experiencing the costs of the terms offered, they revise the contract. For example, when the firm opened operations in Italy, it offered the same replacement fee of €40 for stolen bicycles that it offers in the Netherlands and decided to revise after experiencing the term.

The maturity of the firm and product affects the frequency and extent to which contracts are revised. For example, Fiat noted that they had not changed their vehicle standard terms in fourteen years and only revised them after being acquired by Chrysler in 2011.⁴² This is in contrast to Fiat’s post-purchase service and warranty products, which are very dynamic and revised frequently. At Google, terms governing relationships with advertisers are revised approximately yearly using a cross-functional process, where relevant departments propose revisions and draft language together with the legal team. Changes to service-level agreements are written jointly with the technical services and resourcing team, and changes to information privacy terms are written with the input of the privacy team. In addition, different departments communicate informally with the legal team to relay relevant information that later gets incorporated into the revised contract. The revisions then go through a final approval process.⁴³ The large media company and Amazon follow similar but less structured approaches, where firm lawyers gather information through informal channels to learn about changes in technology and customer demands.⁴⁴ New regulations, such as the enactment of the European Union’s General Data Protection Regulation (GDPR), require firm-wide systematic reviews.

Young and nimble firms like Swapfiets actively seek experiential learning as feedback and revise their terms accordingly. For example, the firm’s initial contract promised customers to replace a broken bicycle with a new one within 24 hr, charged a uniform fee for broken bikes caused by customer misuse, and made a bicycle leasing program available to a broad range of customers. Yet, the firm quickly learned that replacing bicycles within such a short window was costly to achieve, so they increased it to 48 hours. Note that this term likely allowed the firm to learn asymmetrically; a longer replacement time would have arguably allowed the repair team to use all the time available to repair each unit, thus reducing their ability to learn how fast it would be feasible to complete the repairs. The firm also learned that charging a fixed fee for broken bicycles resulted in too many replacement claims, so it changed the fee structure to a variable one based on the severity of harm.⁴⁵ Finally, the firm restricted the types of potential subscribers by adding some financial requirements after learning from experience that some subscribers had no funds to pay for the rented units.

What else causes firms to revise their terms? Counsel referenced the need to revise terms as a result of changes in the law (such as the enactment of GDPR,⁴⁶ or litigated cases), changes in product functionality or services provided, changes in technology, and changes in industry standards or the terms of competitors. Fiat explained that it revised its terms after receiving feedback from dealer and consumer associations or whenever there was a regulatory or external push.⁴⁷ This type of learning represents indirect, rather than experiential, learning. Most relevant for our study, several of the changes discussed by in-house counsel result from experiential learning as well. In addition to Swapfiet’s experiences, the counsel at the large media company learned that granting broad use rights to customers of media services turned out to be too costly, as this was used to avoid paying for additional services, such as when sports programming subscribers engaged in “location spoofing” to avoid paying for the ability to watch out-of-state games. Google changed a dispute resolution clause after learning that it was costlier to its contracting parties than to the firm.

One subject relayed an anecdote from his years while working in-house at Twitter. At that time, Twitter’s End User Agreement included standard terms related to intellectual property rights that were mostly unrestricted, giving customers the right to use such IP terms in various forms. After noting that customers were using the Twitter trademark on other social media platforms and experiencing the costs associated with such uses, Twitter revised its agreement to prohibit the use of the Twitter trademark on other social media platforms. The broad term of giving customers extensive leeway to use Twitter’s marks resulted in the firm learning about the cost of such unrestricted uses through experience. Similarly, Swapfiets initially offered a very flexible contract that allowed monthly subscriptions and free cancelations. Many subscribers took advantage of the flexibility to rent bicycles for short-term needs, reducing the firm’s ability to profit from the transactions. Now Swapfiets charges €50 for the short-term cancelation option or demands a six-month commitment from subscribers. All of these revisions were the result of experiential learning. In the case of young firms, such as Swapfiets, the generous terms likely allowed the firm to learn asymmetrically.

Overall, the interviews revealed that firms need to write standardized agreements to manage contracting at a large scale, but with the expectation that they will be revised. Revisions occur as a result of changes in market, technological, and legal conditions but also as a result of learning, both indirect and experiential, purposefully and naïvely acquired. Indeed, firms have both formal and informal mechanisms to relay such experiential information to the relevant parties in charge of revising terms. These accounts support our hypothesis that experiential learning is a meaningful mechanism driving contract change.

6. IMPLICATIONS AND CONCLUSION

Our theory and findings suggest that having an opportunity to learn from direct experience with customers encourages contracting parties to revise their terms in ways that may increase the benefits derived from the contract. Some normative implications arise, which we will explore in more detail in future work. A few are salient. First, lawmakers deciding whether to create or modify default rules (e.g., in the context of revising Articles or provisions of the UCC) might want to consider the extent to which certain default rules allow for experiential learning, especially when default rules are sticky. All else equal, it might be more desirable to adopt defaults that carry an opportunity to learn. Identifying such cases could be challenging, however. Second, when learning occurs by opting out of the default, the stickiness inherent (and oftentimes built up) in default rules hampers learning. This offers an additional reason against making default rules sticky in these particular circumstances.⁴⁸ Finally, experiential learning weighs against the implementation of mandatory rules, which are more common in the European Union, since they may hinder learning for some terms and prevent firms from acting on learned information by revising their agreements.

Standard-form contracts include terms that may benefit customers (or suppliers) and generate costs for the firm in ways that are not perfectly predictable at the outset. Adopting a contract term is often akin to experimentation: the firm may accept the risk of short-term losses to learn the net value of the term and make a better-informed decision in the future. Yet, only some terms offer an opportunity to learn from experience. We identify a real option that is generated by this opportunity to learn that may affect ex-ante contracting decisions. Experiential learning extends to other aspects of firms’ products or services, allowing them to purposely experiment and revise such features at a later time.

We have introduced a distinction between two main categories of terms: symmetric-learning terms are terms that offer symmetric opportunities to learn experientially to firms that adopt them and those that do not adopt them; asymmetric-learning terms are those that offer an opportunity to learn either to adopting firms or to non-adopting firms, but not to both. Exploiting differences in the way firms learn from their contractual choices, we have built a theory of experiential learning in standard-form contracts. The theory predicts that firms will be more likely to revise terms that offer an opportunity to learn and might fail to revise terms that do not offer such an opportunity. In this way, our theory accounts for both stickiness and change in standard-form contracts. Through this lens, we have examined and classified the terms included in the End User Software License Agreements (EULAs) by a sample of 264 firms across 114 different software markets in 2003 and in 2010. We found that learning opportunities are a determinant of change, overcoming the stickiness of defaults. When such opportunities are absent, terms may survive long enough to appear obsolete and out of touch with the rest of the contract.

The analysis we present in this article opens, we hope, interesting avenues for further theoretical and empirical inquiry. To our knowledge, we are the first to identify the learning modalities of different terms and to draw conclusions for contractual choices, with possible extensions beyond the contract. Yet, we use a rather rigid, binary classification that does not allow us to distinguish modalities that imply more or less learning. Further research could provide interesting insights into the learning potential of different terms: which terms allow firms to learn the most? Learning also occurs through different channels, as we have emphasized. We have offered some preliminary considerations as to the most likely scenarios in which experiential learning may primarily take place.⁴⁹ Fully understanding how different learning channels interact with experiential learning, as well as how new technology affects the way in which firms learn, are important questions that we hope will be the subject of future work.

Our analysis focuses on learning from direct experience and we have stressed the firm’s behavior in response to information about the costs of offering certain clauses.⁵⁰ In general, such learning is beneficial because it allows the firm to offer terms that maximize the value of consumer contracts and standardized, non-negotiated, business-to-business contracts. This observation speaks against the stickiness of default terms: defaults should not be sticky because stickiness distorts the process of learning and prevents firms from opting out of default terms in cases in which this choice would otherwise be optimal. From a normative viewpoint, absent countervailing considerations, the law should make contractual choices as neutral as possible since leveraging on the attractiveness of default provisions comes with a possibly high cost.

Footnotes

*

The authors would like to thank Ian Ayres, Oren Bar-Gill, Omri Ben-Shahar, Lisa Bernstein, Albert Choi, Steven Choi, Kevin Davis, Marco Fabbri, Clayton Gillette, Ron Gilson, Mitu Gulati, Scott Hemphill, Gerard Hertig, Keith Hylton, William Hubbard, Louis Kaplow, Henrik Lando, Bentley MacLeod, Mark Ramseyer, Alan Schwartz, Robert Scott, Steven Shavell, Holger Spamann, Kathryn Spier, Eric Talley, Abraham Wickelgren, David Webber, Eyal Zamir, Kathryn Zeiler, Jonathan Zytnick, and the participants in the Conference on Contractual Black Holes at Duke Law School in 2017, the annual meetings of the European Association of Law and Economics in 2017 and the American Law and Economics Association in 2018, the Conference in Honor of Professor Robert E. Scott at Columbia Law School in 2022, and workshops at NYU School of Law, ETH Zurich, University of Amsterdam, Boston University School of Law, Duke Law School, Columbia Law School, University of Texas Law School, University of Chicago Law School, and Harvard Law School for helpful comments and suggestions. Melissa Bales and Melek Redzheb provided skilled editorial assistance. Giuseppe Dari-Mattiacci gratefully acknowledges the very generous research support provided by Columbia Law School Florencia Marotta-Wurgler gratefully thanks the New York University School of Law Foundation for generous research support.

†

E-mail: [email protected]

1

See, for example, Gulati & Scott (2013) (explaining contracting parties’ reluctance to revise pari passu clauses in sovereign bond agreements after unfavorable interpretations by courts).

2

Amita Kelly and Merrit Kennedy, “L.L Bean Scraps Legendary Return Policy”, National Public Radio, February 9, 2018. Available at: https://www.npr.org/sections/thetwo-way/2018/02/09/584493046/l-l-bean-scraps-legendary-lifetime-return-policy. Emphasis added by the authors. See also Shep Hyken, “L.L. Bean Discontinues Lifetime Guarantee”, Forbes, February 18, 2018. Available at: https://www.forbes.com/sites/shephyken/2018/02/18/l-l-bean-discontinues-lifetime-guarantee/?sh=1dc7381a714d.

3

The literature on real options starts with McDonald & Siegel (1986) (showing that irreversible decisions to invest can be understood using the framework developed in finance for the study of option contracts). Real option theory has been applied to diverse topics. SeeTriantis & Triantis (1998) (applying option theory to the study of contract breach); Baird & Morrison (2001) (to bankruptcy); Fennell (2005) (to property and liability); Grundfest & Huang (2006) and Cotropia (2009) (to intellectual property); Spitzer & Talley (2014) (to financial regulation); Vladeck (2015) (to regulatory impact analysis). For an encompassing analysis of how option theory affects the study of the law, see Ayres (2005).

4

Bob Pibbs, “Zappos Pays People to Quit. Should You?” The Retail Doctor. Available at: https://www.retaildoc.com/blog/zappos-pays-people-to-quit-should-you. Bill Taylor, “Why Zappos Pays New Employees to Quit–And You Should Too”, Harvard Business Review, May 19, 2008. Available at: https://hbr.org/2008/05/why-zappos-pays-new-employees. Rachel Emma Silverman, “At Zappos, Some Employees Find Offer to Leave Too Good to Refuse”, Wall Street Journal, May 7, 2015. Available at: https://www.wsj.com/articles/at-zappos-some-employees-find-offer-to-leave-too-good-to-refuse-1431047917. Upon purchasing Zappos, Amazon started making a similar offer of $5,000 to its employees only to suspend it in 2022 amid a labor crunch and fear of high turnover costs. See Huileng Tan, “Amazon has suspended a program that pays up to $5,000 to warehouse workers who quit after peak seasons: report”, Insider, January 26, 2022. Available at: https://www.businessinsider.com/amazon-suspends-program-pays-warehouse-workers-to-quit-2022-1?international=true&r=US&IR=T.

5

Therefore, we should observe lower ex-post revision rates with experience as compared to experimentation. Since our empirical results are based on ex-post revision rates rather than ex-ante drafting choices, the possibility that firms learn from experience rather than experimenting willfully would work against finding an effect of learning and hence reinforces our results. For more details, see the discussion following Proposition 2.

6

See Section 3.5.2 for an extended discussion.

7

While we will refer to the firm’s contracting counterparty as “customer”, our analysis applies also to other parties, such as suppliers, to whom the firm offers a standard-form contract. See Sections 3.5.6 and 4.5.

8

For a review of the literature on learning and innovation in the standard-form contract setting, see Section 2.

9

See Ariella Gintzer, L.L. Bean’s Lifetime Return Policy is No More, Outside Magazine, February 9, 2018. Available at: https://www.outsideonline.com/2280581/ll-bean-tightens-generous-return-policy.

10

See Section 3.5 for further discussion of these points.

11

In January 2021, FCA merged with Groupe PSA to create a new entity, Stellantis, which now controls 14 brands: Abarth, Alfa Romeo, Chrysler, Citroën, Dodge, DS Automobiles, FIAT, Jeep, Lancia, Maserati, Opel, Peugeot, Ram Trucks e Vauxhall.

12

Indeed, as revealed by the attorneys interviewed, the processes in which standard-form contracts are formulated and later revised tend to be modular in nature. Some sets of related terms are set in stone while others are quite malleable. For a thorough account of modularity in contract, see Smith (2005).

13

The free-rider problem is also central to Bolton & Harris (1999). We mute this channel in the present paper.

14

See Hoffman (2018).

15

For example, the in-house counsel from Fiat International explained in our interview that the firm decided to discontinue a separately sold extended warranty for vehicle tires after learning that demand for the warranty was almost zero.

16

L.L. Bean Warranty, available at: https://www.llbean.com/llb/shop/513705?page=our-guarantee (last accessed April 30, 2022). The Return Policy states further conditions where returns will be denied, even within one year, including “Products damaged by misuse, abuse, improper care or negligence, or accidents (including pet damage); [p]roducts showing excessive wear and tear,” among others.

17

See note 8.

18

For a detailed analysis of how firms adopt or opt out of UCC Article 2 default rules, see Marotta-Wurgler (2007).

19

For example, Intuit tax preparation software offers a version of such a warranty (“If you are a registered user and you pay an IRS or state penalty and/or interest solely because of a calculation error on a form prepared for you using TurboTax Online, and not as a result of, among other things, your failure to enter all required information accurately, willful or fraudulent omission or inclusion of information on your tax return, misclassification of information on the tax return, or failure to file an amended return to avoid or reduce an applicable penalty/interest after Intuit announced updates or corrections to the TurboTax Online software in time for you to file an amended return, then Intuit will pay you in the amount of the IRS or state penalty and/or interest paid by you to the IRS or state.”) Available at: https://turbotax.intuit.com/corp/license/online.jsp (last visited, May 30, 2018).

20

With appropriate modifications, the analysis also applies to the reverse case, in which the firm is a buyer and the contractual counterparty is a seller, such as supplier. See note 8.

21

As further explained in Section 3.5.2, the assumption that consumers ignore their types is both realistic—firms often know much more about their customers’ future use patterns and associated risks than the customers themselves—and a convenient way to capture the benefits of contract standardization even in cases in which it could be theoretically possible for firms to discriminate. We offer below further support for this modeling choice.

22

It is easy to see that alternative assumptions are either uninteresting or lead to analogous scenarios and hence would not change our conclusions. To be precise, next to the case analyzed in the text, we have three additional possibilities: if v ≤ min{c_H, c_L}, then offering the opt-out term would always be optimal; if v ≥ max{c_H, c_L}, then offering the default would always be optimal; finally, if c_H ≤ v ≤ c_L, then it would be optimal to offer the default term to customers of type H and the opt-out term to customers of type L, reversing some of our expressions but leaving the core of our results unchanged.

23

That is, when the net relative cost of the default is c_L, which is less than its relative value v. Note that, after simple manipulations, v > c_L can be written as $v^D-c_L^D>v^O-c_L^O$⁠, that is, the net value of the default term is greater than the net value of the opt-out term.

24

That is, when the net relative cost of the default is c_H, which is greater than its relative value v. Note that v < c_H can be written as $v^D-c_H^D<v^O-c_H^O$⁠, that is, the net value of the default term is less than the net value of the opt-out term.

25

The case in which costs are the same for both the default and the opt-out is uninteresting because it involves no learning.

26

Zero is due to normalization. Recall that v, c_H, and c_L are the normalized value and costs of the default term relative to the opt-out, which has then normalized value and costs equal to 0.

27

Turning to the probability of switches away from the no-learning modality of an asymmetric learning term (which is 0), note that this probability would still be different from the probability of switches away from a symmetric-learning term (which is positive), even in cases of serendipitous, non-intentional learning from experience.

28

There may be opt-out costs that are incurred only once if the firm opts out at time 0 and confirms this choice at time 1. To avoid additional notation, we do not take this case into account. Its inclusion, however, would not change the main results.

29

Note it is immaterial whether w captures an accurate forecast of the firm’s growth at time 1: time-0 decisions depend on the expected, not the actual, growth; time-1 choices depend on per unit values and costs and hence are unaffected by growth. Finally, there may be an interaction between growth and switching costs as a firm with a larger volume of sales may enjoy economies of scale which reduce the fixed costs of switching. We do not explore this possibility in the analysis.

30

In a screening model, the uninformed party makes an offer to the informed party. The offer allows the informed party to choose between two different contracts, say, one with a warranty and one without, at two different prices. If designed appropriately, such an offer results in a separation between the two (or more) types of customers; in our context, this means that customers of type L would choose the contract with the default warranty and customers of type H would choose the (cheaper) contract without warranty.

31

In many cases, the value that a customer attaches to a contract term varies with the customer’s use patterns, which in turn determines the costs borne by the firm. A warranty may provide the typical example: high-risk customers both value the warranty more and impose larger costs on the firm. Under these conditions, heterogeneous valuations are correlated with customer types and allow the firm to adopt a screening strategy. We do not consider these cases for the reasons explained below.

32

From a different perspective, our problem is similar in structure to the insurance market, where insurance companies ignore the risk characteristics of those who purchase coverage (Rothschild & Stiglitz 1976). However, there is an important difference between our setup and the traditional insurance analysis. The latter is a problem of asymmetric information where one party is informed and the other is uninformed. In our setup, both parties are uninformed about the costs of different terms. This difference makes the traditional solutions to this problem unworkable in our setting. There also is an equally extensive literature on the opposite problem, the one that customers face when they are unable to distinguish between “good” and “bad” products, while firms are informed (Schwartz & Wilde, 1979; Priest, 1981). We do not deal with these issues in the present article.

33

In our model of learning about costs, firms do not have any incentive to adjust the price in response to learning because they are already pricing the product at the customer’s willingness to pay.

34

Note that this model would be symmetric to our model of learning about costs.

35

On Delta, see https://crankyflier.com/2012/05/17/delta-screws-up-accidentally-shows-different-prices-to-some-customers-on-its-website/ (last accessed April 30, 2022). On Amazon, see https://www.consumeraffairs.com/news/lawsuit-alleges-amazon-charges-prime-members-for-free-shipping-031414.html#:~:text=In%20September%202000%2C%20Amazon.com,drop%20from%20%2426.24%20to%20%2422.74 (last accessed April 30, 2022).

36

These markets are very finely defined and can be grouped into larger, more general, markets. For example, Amazon defines one market as “Office Suites”, which is included in a larger market labeled “Business and Office”. For a detailed account of these variables and the methodology used, see Florencia Marotta-Wurgler, Competition and the Quality of Standard Form Contracts, 5 J. Empirical Legal Stud. 447, 457–67 (2008).

37

See ProCd v. Zeidenberg, 86 F.3d 1447 (1996) (enforcing a contract when terms proceed purchase as long as the recipient of the term has notice of the later arriving terms and a reasonable opportunity to reject such terms).

38

A study of Terms and Conditions in online dating sites also finds support for this hypothesis. Eisenberg, Swipe Right for Love (and Liability, Licensing, and Limits on User Behavior): Change and Innovation in Dating Application Terms of Service (unpublished draft, 2018).

39

Interview with Elizabeth Daly, in-house counsel at Google Ad Services division, on February 20, 2019 (notes on file with authors).

40

Indeed, as revealed by the attorneys interviewed, the processes in which standard-form contracts are formulated and later revised tend to be modular in nature. Some sets of related terms are set in stone, while others are quite malleable. For a thorough account of modularity in contract, see Smith (2005).

41

Interview with Edzard Kalff, CFO at Swapfiets, on October 30, 2020 (notes on file with authors).

42

Interview with Vito Iacobellis and Stefania Isaia, commercial attorneys at Fiat Chrysler Automobiles, on March 27, 2019 (notes on file with authors).

43

Interview with Google, see footnote 40.

44

Interview with in-house counsel at a large media company, on February 27, 2019 (notes on file with authors); interview with Steven Coates, in-house counsel at Amazon, on March 19, 2019 (notes on file with authors).

45

Interview with Swapfiets, see footnote 42.

46

Regulation (EU) 2016/679 (General Data Protection Regulation).

47

Interview with Fiat, see footnote 43.

48

See Schwartz & Scott (2016, 2021) (offering a critique of projects of law reform that seek to establish contract default rules and discussing the problems with the creation of such default rules).

49

See Section 3.5.

50

We recognize that firms may also experiment with ways by which they could exploit customers. There is a large literature about this and similar problems, and we do not examine it here.

Appendix 1. Extended model and proofs

In this section, we extend the model presented in the text in two ways: (i) we consider a general cumulative distribution function F(p) for the population of firms (we assumed that it was uniform in the text) and (ii) we consider two continuous probabilities of learning, ${\lambda }^D$ from the default term and ${\lambda }^O$ from the opt-out. Symmetric-learning terms are those with ${\lambda }^D={\lambda }^O$⁠, asymmetric-learning-from-default terms are characterized by ${\lambda }^D>{\lambda }^O$ and, conversely, asymmetric-learning-from-opt-out terms are characterized by ${\lambda }^D<{\lambda }^O$⁠. In the text, we had ${\lambda }^{\tau }\in \{0,1\}$⁠, where $\tau \in \{D,O\}$⁠, now we allow for ${\lambda }^{\tau }\in [0,1]$⁠.

As usual, we proceed backwards. In the last period, time 1, the firm is better off adopting the term that maximizes its payoff given the information available. We need to distinguish between two scenarios. If the firm has learned c, the firm’s payoff at time 1 is either v – c_L if c = c_L (because the firm chooses the default at time 1) or is equal to 0 if c = c_H (because the firm chooses the opt-out at time 1). If the firm does not learn, the firm will confirm the time-0 choice. We can write the time-1 expected payoff ${\mathrm{\Pi }}_1^{\tau }$ of a firm that adopts term τ at time 0 as follows:

$${\displaystyle \begin{array}{l}{\displaystyle {\mathrm{\Pi }}_1^{\tau }={\lambda }^{\tau }(1-p)(v-c_L)+(1-{\lambda }^{\tau }){\mathrm{\Pi }}_0^{\tau }.}\end{array}}$$

(7)

In turn, the expected time-0 payoff $\mathrm{\Pi }{}_0^{\tau }$⁠, when information is lacking, is simply the difference between the value of the term and its expected costs, and hence is $\mathrm{\Pi }{}_0^D=v-p{c}_H-(1-p)c_L$ if the firm chooses the default option and $\mathrm{\Pi }{}_0^O=0$ if the firm opts out.

Backing up to the earlier period, time 0, the firm weighs two possibly opposing interests: maximizing the expected time-0 payoff $\mathrm{\Pi }{}_0^{\tau }$ and improving the time-1 payoff $\mathrm{\Pi }{}_1^{\tau }$ through learning. If it learns and switches, the firm also pays a switching cost s. The probability of switching is q^D = λ^Dp if the firm chooses the default at time 0 and q^O = λ^O(1 – p) if the firm chooses the opt-out because switches occur only if the firm learns that the opposite choice is the efficient one.

Accordingly, if a firm of type p chooses the default at time 0, its expected payoff is

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{\Pi }{}_0^D+w\mathrm{\Pi }{}_1^D-q^{D}s=(1+w)\mathrm{\Pi }{}_0^D+{\lambda }^{D}p[w(c_H-v)-s].}\end{array}}$$

(8)

The expression reveals that, at time 1, the firm is guaranteed its time-0 payoff and an additional payoff that derives from learning. If the firm chooses the default at time 0, the payoff from learning consists of the possibility to switch to the opt-out if the cost is high (with probability p) and hence avoid the loss c_H – v, in which case the firm also pays a switching cost.

If instead, the firm chooses the opt-out at time 0, its expected payoff is

$${\displaystyle \begin{array}{l}{\displaystyle {\mathrm{\Pi }}_0^O+w{\mathrm{\Pi }}_1^O-q^{O}s={\lambda }^O(1-p)[w(v-c_L)-s].}\end{array}}$$

(9)

Again, at time 1, the firm is guaranteed its time-0 payoff (of zero) and an additional payoff derived from learning, which now consists of the possibility to switch to the default if the cost is low (with probability 1 – p) and hence captures the gain v – c_H, in which case the firm also pays a switching cost.

By equating (8) and (9), we can derive a cutoff level of p such that the firm is indifferent between adopting the default at time 0 and opting out of it:

$${\displaystyle \begin{array}{l}{\displaystyle p^{\ast }\equiv \frac{(1+w)(v-c_L)-{\lambda }^O[w(v-c_L)-s]}{(1+w)(c_H-c_L)-{\lambda }^D[w(c_H-v)-s]-{\lambda }^O[w(v-c_L)-s]}.}\end{array}}$$

(10)

Note that the numerator consists of the gains from adopting the default directly in both periods minus the gains from switching to the default from the opt-out after learning. The numerator thus captures the incentives to choose the default at time 0. The denominator is the sum of the incentives to choose the default option and the incentives to choose the opt-out.

Firms with p < p^* face low expected costs and hence adopt the default term, while firms with p > p^* opt out of it. Note that the formulas in Expressions (1) to (6) can be easily derived from (10) by replacing 0 or 1, as appropriate, for ${\lambda }^D$and ${\lambda }^O$ in all 6 expressions, and by setting s = 0 and w = 1 for Expressions (1) to (3).

The second measure in which we are interested is the mass of switches at time 1. The probability that a firm’s p switches away from the default to the opt-out at time 1 is equal to the probability of learning, λ^D, given that the firm has chosen the default at time 0, times the probability that the opt-out is the superior choice—that is, that costs are high—which is equal to p. Considering all firms, we have:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{\Delta }(p^{\ast })\equiv {\displaystyle \underset{0}{{\displaystyle \stackrel{p^{\ast }}{\int }}}}{\lambda }^{D}pdF(p).}\end{array}}$$

Similarly, the mass of switches away from the opt-out to the default at time 1 is equal to:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{\Omega }(p^{\ast })\equiv {\displaystyle \underset{p^{\ast }}{{\displaystyle \stackrel{1}{\int }}}}{\lambda }^O(1-p)dF(p).}\end{array}}$$

We will now analyze how the firm’s decisions depend on the information type of a term, $({\lambda }^D,{\lambda }^O)$⁠, and demonstrate two general results.

 

Proposition 7. The learning modality of a term is chosen increasingly often at time 0 as learning becomes more asymmetric.

Proof. The proposition implies that if λ^D > λ^O, the default is chosen more often at time 0 as the difference λ^D – λ^O increases. Since F(p) increases monotonically in p, this is the case if p^* increases in λ^D and decreases in λ^O. Analogously, if λ^D < λ^O the opt-out is chosen more often at time 0 as the difference λ^O – λ^D increases; this is the case if 1 – F(p^*) decreases in λ^D and increases in λ^O, which occurs, again, if p^* increases in λ^D and decreases in λ^O. Thus, it is enough to show that $\mathrm{\partial }{p}^{\ast }/\mathrm{\partial }{\lambda }^D>0$ and $\mathrm{\partial }{p}^{\ast }/\mathrm{\partial }{\lambda }^O<0$⁠. We have:

$${\displaystyle \begin{array}{l}{\displaystyle \frac{\mathrm{\partial }{p}^{\ast }}{\mathrm{\partial }{\lambda }^D}=\frac{[w(c_H-v)-s][(1+w-w{\lambda }^O)(v-c_L)+{\lambda }^{O}s]}{{[(1+w)(c_H-c_L)-{\lambda }^D[w(c_H-v)-s]-{\lambda }^O[w(v-c_L)-s]]}^2}.}\end{array}}$$

The latter expression is positive if s < w(c_H – v), that is, if switching costs are not so high as to prevent switching at time 1, which follows from Assumption 1. Similarly, we have:

$${\displaystyle \begin{array}{l}{\displaystyle \frac{\mathrm{\partial }{p}^{\ast }}{\mathrm{\partial }{\lambda }^O}=-\frac{[w(v-c_L)-s][(1+w-w{\lambda }^D)(c_H-v)+{\lambda }^{D}s]}{{[(1+w)(c_H-c_L)-{\lambda }^D[w(c_H-v)-s]-{\lambda }^O[w(v-c_L)-s]]}^2},}\end{array}}$$

which is negative if s < w(v – c_L) as we assume in Assumption 1. Q.E.D.

Absent learning—that is with λ^D = λ^O = 0—Expression (10) would be a ratio of the net value of the default (v – c_L) over the sum of the net values of the two options (v – c_L + c_H – v = c_H – c_L), both weighed by the firm’s time-0 and time-1 sales (1 + w) and diminished by the switching cost, s. With learning, additional terms enter the expression, which captures the real-option value of choosing either version of the term. The marginal type p^* increases with λ^D, enlarging the set of firm types p < p^* that choose the default at time 0. Similarly, p^* decreases with λ^O, expanding the use of the opt-out. As learning becomes more asymmetric—that is when λ^D – λ^O increases—the learning modality yields learning relatively more frequently and hence becomes more attractive for firms at time 0.

 

Proposition 8. Switches away from the learning modality of a term occur increasingly often at time 1 as learning becomes more asymmetric.

Proof. The proposition implies that if λ^D > λ^O, there are more switches away from the default at time 1 as the difference λ^D – λ^O increases; that is, it implies that Δ(p^*) increases as the difference λ^D – λ^O increases. Note that Δ(p^*) increases in p^*. Proposition 7 shows that as the difference λ^D – λ^O increases, then $p^{\ast }$ increases, which implies that Δ(p^*) also increases. Conversely, the proposition also implies that if λ^D < λ^O, there are more switches away from the opt-out at time 1 as the difference λ^D – λ^O increases; that is, it implies that Ω(p^*) increases as the difference λ^D – λ^O increases. Note that Ω(p^*) decreases in p^*. Proposition 7 shows that as the difference λ^D – λ^O increases, then p^* decreases, which implies that Ω(p^*) increases. Q.E.D.

If learning is asymmetric, we know from Proposition 7 that the learning modality is chosen more often at time 0. This effect expands the set of firms that can potentially switch to the other modality at time 1. In addition, those firms learn relatively more often—precisely because they have chosen the learning modality—and hence more of them will switch at time 1. An important implication of Proposition 8 is that a default term will be revised at the highest rate if it is the learning modality of an asymmetric-learning term, at a middle rate if it is a symmetric-learning term, and at the lowest rate if it is the no-learning modality of an asymmetric-learning term, as illustrated in Figure 1. And similarly, for the opt-out. Also, as explained, the difference in revision rates between the learning modality of an asymmetric-learning term and the same modality of a symmetric-learning term captures the extent to which firms experiment. In the following, we provide technical details, where needed, for the proofs of the propositions in the main text. The proofs of Propositions 1–3 are straightforward and hence omitted.

 

Proof of Proposition 4. Since by Assumption 1 we can subsume k under v, it is sufficient to examine the following derivative:

$${\displaystyle \begin{array}{l}{\displaystyle \frac{\mathrm{\partial }{p}^{\ast }}{\mathrm{\partial }v}=\frac{(c_H-c_L)[1+(2-{\lambda }^D-{\lambda }^O)w+(1-{\lambda }^D)(1-{\lambda }^O)w^2]+[{\lambda }^D+{\lambda }^O+({\lambda }^D(1-{\lambda }^O)+{\lambda }^O(1-{\lambda }^D))w]s}{{[c_H-c_L+({\lambda }^D+{\lambda }^O)s+[c_H(1-{\lambda }^D)-c_L(1-{\lambda }^O)+({\lambda }^D-{\lambda }^O)v]w^2]}^2}}\end{array}}$$

which is positive. Q.E.D.

 

Proof of Proposition 5. We need to consider the following derivative:

$${\displaystyle \begin{array}{l}{\displaystyle \frac{\mathrm{\partial }{p}^{\ast }}{\mathrm{\partial }w}=\frac{({\lambda }^D-{\lambda }^O)(c_H-v)(v-c_L)-[{\lambda }^O(1-{\lambda }^D)(c_H-v)-{\lambda }^D(1-{\lambda }^O)(v-c_L)]s}{{[c_H-c_L+({\lambda }^D+{\lambda }^O)s+[c_H(1-{\lambda }^D)-c_L(1-{\lambda }^O)+({\lambda }^D-{\lambda }^O)v]w^2]}^2}}\end{array}}$$

which is positive iff

$${\displaystyle \begin{array}{l}{\displaystyle {\lambda }^D-{\lambda }^O>[\frac{{\lambda }^O(1-{\lambda }^D)}{v-c_L}-\frac{{\lambda }^D(1-{\lambda }^O)}{c_H-v}]s.}\end{array}}$$

(11)

Assume the default term is the learning modality, that is, λ^D > λ^O. Then, the learning modality is chosen more often as growth opportunities improve if the derivative above is positive, that is, if the condition in Expression (11) is verified. We have that λ^D(1 – λ^O) > λ^O(1 – λ^D) and hence the inequality in Expression (11) is verified unless v – c_L is sufficiently smaller than c_H– v. That is, the condition is violated if v is sufficiently less than $c_H+c_L/2$⁠. Vice versa, if the opt-out is the learning modality, we have λ^D < λ^O. Now, the learning modality is chosen more often as growth opportunities improve if the derivative above is negative, that is, if the condition in Expression (11) is not verified. We have that λ^D(1 – λ^O) < λ^O(1 – λ^D) and hence the inequality (11) is verified unless c_H– v is sufficiently smaller than v – c_L. That is, the condition is violated if v is sufficiently greater than $(c_H+c_L)/2$⁠. Therefore, growth prospects increase adoption of the learning modality at time 0 and switches away from the leaning modality at time 1 in a neighborhood of $v=(c_H+c_L)/2$⁠. With symmetric-learning terms, λ^D = λ^O, the condition in (11) is verified iff v – c_L> c_H– v, in which case an increase in growth prospects results in more frequent adoption of the default, and vice versa. Q.E.D.

 

Proof of Proposition 6. We need to consider the following derivative:

$${\displaystyle \begin{array}{l}{\displaystyle \frac{\mathrm{\partial }{p}^{\ast }}{\mathrm{\partial }s}=\frac{(c_H-v){\mathrm{\lambda }}^O[1+w(1-{\mathrm{\lambda }}^D)]-(v-c_L){\mathrm{\lambda }}^D[1+w(1-{\mathrm{\lambda }}^O)]}{{[c_H-c_L+({\mathrm{\lambda }}^D+{\mathrm{\lambda }}^O)s+[c_H(1-{\mathrm{\lambda }}^D)-c_L(1-{\mathrm{\lambda }}^O)+({\mathrm{\lambda }}^D-{\mathrm{\lambda }}^O)v]w^2]}^2}}\end{array}}$$

which is negative iff

$${\displaystyle \begin{array}{l}{\displaystyle \frac{{\mathrm{\lambda }}^D}{{\mathrm{\lambda }}^O}>\frac{(1+w-{\mathrm{\lambda }}^{D}w)(c_H-v)}{(1+w-{\mathrm{\lambda }}^{O}w)(v-c_L)}}\end{array}}$$

(12)

Assume the default term is the learning modality, that is, λ^D > λ^O. Then, the learning modality is chosen less often as switching costs increase if the derivative above is negative, that is, if the condition in Expression (12) is verified. We have that 1 + w – λ^Dw < 1 + w – λ^Ow and hence the inequality (12) is verified unless v – c_L is sufficiently smaller than c_H – v. That is, the condition is violated if v is sufficiently less than $(c_H+c_L)/2$⁠. Vice versa, if the opt-out is the learning modality, we have λ^D < λ^O. Now, the learning modality is chosen less often as switching costs increase if the derivative above is positive, that is, if the condition (12) is not verified, which is in turn true unless c_H – v is sufficiently smaller than v – c_L. That is, the condition is violated if v is sufficiently greater than $\frac{c_H+c_L}{2}$⁠. Therefore, switching costs reduce the adoption of the learning modality at time 0 and switches away from the leaning modality at time 1 in a neighborhood of $v=(c_H+c_L)/2$⁠. Note that with symmetric-learning terms, λ^D = λ^O, the condition (12) is verified iff v – c_L > c_H – v, in which case an increase in switching costs results in more frequent adoption of the opt-out, and vice versa. Q.E.D.

Appendix 2. Tables

Appendix 3. Questionnaire for In-house Counsel

1. What is your name and position at firm X?

2. What do you exactly do at X?

3. How do you write customer agreements? Do you start from scratch?

4. Is there a process where the people in legal get together?

5. What do you consider when drafting terms? Do you consult with other departments?

6. Do you revise terms?

7. Please tell us about times when you revised the terms. In response to what?

8. Do you anticipate revising the terms when you draft them initially?

9. Why do you revise terms? When? (Lawyer turnover? New product? New regulations? Marketing?)

10. What pushes you to revise terms?

11. Are there any scheduled time to revise agreements or terms in particular?

12. Which terms are the ones likely to get revised?

13. Is there an approval process for revisions?

14. Do you revise terms by looking at your competitors, litigated cases, technological legal changes, product changes, or something else?

15. Can you think of any examples where you revised the terms because the existing ones did not work out (other than litigation)?

16. Have you ever reevaluated offering a particular term because you found out it was too costly? In what way? How did you learn about it? Is there a process in the firm where this knowledge gets transmitted from a particular department to legal? Which departments and in what way?

17. What do you learn, if anything, from experience with offering a particular term?

18. What do you do with what you learn?

19. Are there some categories of terms that are more prone to this kind of experiential learning than others? What makes those terms, in your view, susceptible to learning?

20. Have you ever tested a term on a subset of customers?

21. To what extent are the legal, customer support, management, and product design teams communicating with each other and affecting the terms?

22. Do you have any other stories or anecdotes about term revisions, here or at a previous firm where you worked at?