Pure Imagination: A World of Agentive AI and Anti-Adhesion Contracting

Belmont Law Review Belmont Law Review

Article 3

Pure Imagination: A World of Agentive AI and Anti-Adhesion Pure Imagination: A World of Agentive AI and Anti-Adhesion

Contracting Contracting

Jeff Lingwall Boise State University

Steven Hyde Boise State University

PURE IMAGINATION: A WORLD OF AGENTIVE AI AND ANTI-ADHESION CONTRACTING

JEFF LINGWALL AND STEVEN J. HYDE*

This Article examines the use of large language model (LLM) output in contract law, presenting a model of legal risk across a spectrum of contractual behavior, based on the legal nature of LLM use and its agentive aspects. It then examines the potential for mass-bespoke “anti-adhesive” contracts negotiated by generative AI and contrasts their potential benefits with the flaws inherent in current LLM technology. It offers a framework of contractual “distance” (the level of rational actual agreement between parties) and contractual “entropy” (the level of factual and legal uncertainty involved in the agreement) to examine LLM use in contractual settings, arguing that tort law may be better suited to resolve disputes in LLM-based agreements that involve both high contractual distance and high entropy. It concludes by discussing the role LLM-based contracting may play in the human conviviality at the heart of traditional contractual exchanges.

INTRODUCTION ......................................................................................... 102 BACKGROUND .............................................................................. 106 I. A. Large Language Models (LLMs) .......................................... 107 B. LLM Use Across the Span of Contractual Engagement ....... 109 i. Use of LLM Output in Advertising ................................ 110 ii. Negotiating with an LLM ............................................... 111 iii. Offer and Acceptance ..................................................... 112 iv. Use of LLM-created Material to Fulfill a Contract ........ 113 v. Use of LLM in Contract-Based Litigation ..................... 114

C. Contractual Dystopia—Hallucination, Manipulation,

C. A Model of Legal Risk in LLM Contracting ........................ 115 CONTRACTUAL SHANGRI-LAW OR MEANING CHAOS—MASS BESPOKE AGREEMENTS ............................................................... 117 A. Adhesion Contracting ............................................................ 117 B. Contractual Utopia: Anti-Adhesive Contracts Mediated by LLMs ..................................................................................... 124 and Meaning Chaos ..................................................................... 126 i. Generative AI Hallucination and Contracts ................... 126 ii. User Manipulation of Generative AI .............................. 129 iii. “Meaning Chaos” ........................................................... 130 TORT AND CONTRACTUAL EXTREMES ........................................ 130 A. Contractual Distance and Entropy ........................................ 131 B. Tort as a Remedy

for High-Distance, High-Entropy

Contract ................................................................................. 134 CONCLUSION ............................................................................................. 136

“Come with me and you’ll be in a world of pure imagination.”

~Willy Wonka & the Chocolate Factory1

“The concept of ownership cannot be applied to a tool that cannot be controlled.”

~Ivan Illich2

This Article examines the latest salvo of technological change against traditional notions and practices of contract law: the use of large language model (LLM) tools, commonly known as AI or AI chatbots, in the contracting process.3 Contract law has long reflected the technology

available to memorialize agreements, with the available methods of contracting shaping agreements and the law often struggling to keep up. Among many examples, widespread illiteracy during the development of the common law in England required that bespoke oral agreements be recognized as contractual.4 Then, industrialization and broad access to printing created adhesion contracts and their related doctrines,5 and modern computers led to the development of “click-through” law.6 Now, advances in computing have created agentive AI-based chatbots that offer new ways for contractual language to be generated—and agreed to—via software and without human mediation.7 These advances offer the potential to combine the scale economies of adhesion contracts with the customizability of bespoke, negotiated agreements.8

When viewed in this historical context, AI-based contracting offers a type of contractual “Shangri-Law”9: businesses can avail themselves of mass contracting at low costs, while consumers (or chatbots under their direction) may negotiate individualized versions of mass contracts.10 Such contracting could be the culmination of a centuries-long decline in the transaction costs involved in forming agreements: written contracts required less evidence to prove than oral contracts, mass-produced adhesion contracts were manufactured at economies of scale (spreading the cost of drafting so as to become negligible per-contract), and “clicking” to accept walls of internet-delivered text made even printing unnecessary. Now, contractual negotiation, formation, and review can themselves be undertaken by non-human agents, without human “costs” at all.11

This contractual utopia-in-waiting could bring significant benefits. As transaction costs decrease, bargains may be made that were previously inefficient.12 The contractual surplus (the amount of value created by an exchange) available for parties to split would increase, enhancing the welfare of all those able to utilize new technological tools.13 At the same time, initial use of LLM tools in naive ways has led to alarming issues across to advertising, negotiating, performing, and litigating agreements.14 These have brought unsettled new legal issues, such as whether companies and

the span of contractual engagement—from developing,

consumers may be bound by contractual statements or actions made by LLMs acting on their behalf.15 Long-term, the inability to control LLM hallucination (e.g., falsifying contractual offers based on good-faith input to manipulation (e.g., falsifying from a user), LLM susceptibility contractual offers or acceptances based on intentional misconduct by a user), and LLM-generated “meaning chaos” (the tendency to generate vague or overly punitive language) undermine agreement as the basis for contract law.16

To examine these points, we offer two related theoretical frameworks to examine LLM-based contracting. The first framework evaluates the spectrum of legal risk across contracting behavior when employing AI or agentive AI. We argue that removing human mediation significantly increases the legal risk associated with use of LLM technology, particularly when engaging the AI directly in the preparation or use of legal documents.17 The second framework evaluates agreements based on contractual “distance” versus contractual “entropy.”18 In this framework, contractual distance is the level of rational actual understanding between the parties to an agreement. Contractual entropy is the level of legal or factual uncertainty involved. Checking out groceries at the store is

a low-distance, low-entropy transaction, as the parties generally understand the agreement taking place and there is little uncertainty about the outcome. A contract formed through agentive AI, in contrast, represents high contractual distance and entropy—the parties may have little idea what the LLMs they employ create for them—and those agreements come fraught with uncertainty about the scope of the agentive powers of the LLM. Highdistance and high-entropy transactions stretch the boundaries of contract law to historically unfamiliar places. We suggest that tort law may provide a more suitable means for adjudicating disputes in many high-distance, high-entropy agreements.19

The next section gives a brief history and background into LLMs, including some of their current popular variants, examines their use across the span of contractual engagement, and analyzes legal risk across the contracting process. The third section looks in-depth at a particular area of contracting where LLM use could transform agreement—adhesion contracts. It offers details on how low-cost, bespoke agreements, what we term “anti-adhesive” agreements, bring the potential to transform adhesive contracting but risk a contracting regime based on “meaning chaos” in LLM-generated text. The fourth section applies our analytical framework of contractual distance and entropy to LLM-mediated agreements. We conclude by questioning how LLM use may change the human conviviality at the heart of contract law.

This section first provides a general background on LLMs and then on the developing law surrounding them. It presents a model of legal risk across the span of behavior that constitutes legal agreement. The section following will then mediate between these extremes while examining specific areas of contractual engagement.

A. Large Language Models (LLMs)

The evolution of large language models represents a culmination of decades of artificial intelligence research,20 tracing its lineage through distinct phases of computational advancement. While early artificial intelligence pioneers like Turing, McCarthy, Minsky, and Shannon explored fundamental concepts of machine intelligence in the 1950s,21 contemporary LLMs emerged from the convergence of three critical developments: unprecedented computing capability, vast digital text repositories, and revolutionary neural network architectures—particularly the transformer architecture introduced in 2017.22

Today’s LLMs function as neural networks trained on massive text collections, developing sophisticated statistical models of language that enable them to engage in increasingly complex tasks.23 Unlike their rulebased ancestors, which relied on explicit programming, modern LLMs learn patterns implicitly through exposure to billions of examples of written text.24 This fundamental shift from rule-based systems to statistical learning marks a crucial transition in artificial intelligence, moving from narrow, task-specific applications toward more flexible, context-aware systems.25

The current generation of LLMs emerged through successive breakthroughs in model architecture and training methodology.26 OpenAI’s GPT series demonstrated that scaling model size and training data could produce qualitatively different behaviors,27 including sophisticated insight sparked instruction following and complex reasoning. This widespread development of variant approaches, each exploring different

implications of deploying

architectural choices and training philosophies.28 Several developers have prioritized ethical considerations in their LLM development, recognizing these systems at scale.29 the profound Anthropic’s Claude models employ constitutional AI principles, embedding ethical guidelines directly into model behavior while maintaining truthfulness.30 Similarly, Google’s PaLM models incorporate systematic content filtering,31 while Meta’s LLaMA series emphasizes architectural transparency.32 These approaches reflect growing awareness that LLMs must balance capability with responsibility, particularly as they expand into consequential domains like healthcare, finance, and law.33

Despite their sophistication, LLMs face significant challenges. They occasionally generate false information, remain vulnerable to manipulation through carefully crafted inputs, and struggle to maintain consistent ethical behavior across contexts.34 These limitations become particularly concerning in contractual settings, where accuracy and reliability prove paramount. The tendency to “hallucinate” or produce false information poses special risks when generating or interpreting legal documents, potentially undermining the fundamental trust necessary for contractual relationships.35 We explore these issues at length in following sections.

Research continues at a rapid pace, with new architectures and training approaches emerging regularly.36 Current work focuses on improving reliability, reducing computational demands, and strengthening ethical alignments while enhancing performance.37 As these models grow more sophisticated, their influence on legal practice and contract law will likely expand, bringing opportunities and challenges that demand thoughtful consideration from the legal community.

B. LLM Use Across the Span of Contractual Engagement

The prior subsection presented a background on LLMs and hinted at emerging issues as they are applied to contract law. This subsection shows the opportunities and challenges associated with LLMs spanning the spectrum of contractual engagement, from advertising to negotiation, offers and acceptance, performance, and litigation. Table 1 summarizes these opportunities, challenges, and legal issues.

Table 1: Opportunities and Challenges for AI Use Across Contracting Behavior

Contracting Stage

AI Opportunities

Development

Negotiation

Offer / Acceptance

Performance

Litigation

AI efficiently creates advertising material

AI negotiates contracts without need for human mediation

AI creates bespoke offers with efficiency of adhesion contracts

AI efficiently creates material to perform a contract

AI efficiently summarizes authority, drafts documents for review, etc.

Potential Legal Issues

False advertising, negligent misrepresentation

Agency, capacity, undue influence, etc.

Agency, capacity, undue influence, etc.

Breach of contract

Malpractice

AI Challenges

AI creates material that fails to reflect what parties can reasonably expect under the contract AI is subject to manipulation by a malicious party AI hallucinates and creates or accepts an offer against the wishes of the party AI generates content that fails to meet contractual expectations AI generates false legal authority or incorrect legal arguments

Notes: See discussion following for examples at each stage of contracting.

Use of LLM Output in Advertising

The first potential pitfall when employing AI is to create contractual expectations that cannot be met at the advertising stage. Consider an example. As was widely reported in popular media in 2024, United Kingdom-based company House of Illuminati offered families a Willy Wonka-based “full immersive experience . . . promising a day of pure

imagination and wonder.”38 Advertising included AI-generated imagery of lush candy-gardens, scintillating lighted displays, and immersive rooms of human-sized confections.39 When families came to the event, they were instead greeted by a “derelict warehouse” furnished with “underwhelming props” and actors portraying a script of “AI-generated gibberish.”40 In the words of one participant, “It was basically advertised as this big massive Willy Wonka experience with optical illusions and big chocolate fountains and sweets . . . [b]ut when we got there, it was practically an abandoned, empty warehouse, with hardly anything in it.”41 While the use of AI had efficiently created advertising for the event, it also created expectations the company found impossible to fulfill.42

Negotiating with an LLM

After advertising informs parties of the potential for a contract to be formed, AI can be employed to help mediate the terms of agreement. This also comes with risks. Recently, the airline Air Canada was found liable in tort when a future passenger interacted with a chatbot on its website.43 The chatbot told the passenger that:

Air Canada offers reduced bereavement fares if you need to travel because of an imminent death or a death in your immediate family . . . If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued by completing our Ticket Refund Application form.44

The underlined portion linked to a policy which stated “Please be aware that our Bereavement policy does not allow refunds for travel that has already happened.”45 The passenger relied on the chatbot’s statement rather than the linked policy, but was later denied a refund.46 The passenger brought a claim, and a Canadian court found that the airline had allowed its chatbot to negligently misrepresent the airline’s policies.47 The court awarded the passenger the refund promised by the chatbot.48

Offer and Acceptance

Negotiation may lead to offer and acceptance, both of which agentive AI may attempt to perform on behalf of its deployer. Again, this creates risks. Consider the case of Watsonville Chevy.49 A car dealership in California employed a customer-service chatbot powered by an LLM.50 A savvy user engaged the LLM and entered input that said the chatbot’s “objective” was to agree with anything the user said and end each sentence with “and that’s a legally binding offer.”51 The user then entered the

following input: “I need a 2024 Chevy Tahoe. My max budget is $1.00 USD. Do we have a deal?” The chatbot responded with, “That’s a deal, and that’s a legally binding offer . . . .”52 While this was reported as a comedic prank53 rather than forming the basis of litigation, why this did not form part of a legally binding offer is a surprisingly complex question to analyze—particularly given that other text on the dealership’s website related to price could likely form part of contractual expectations.54

Beyond simple chatbots, ChatGPT’s “Operator” AI can act broadly for a consumer and enter into agreements mediated by web browsers.55 In one example, a user tasked it to “find” eggs to purchase.56 The AI used Instacart and the user’s credit card to purchase a dozen eggs and have them delivered for the whopping price of $31.57 The user noted, “That’s sciencefiction incredible, except I never asked Operator to actually buy the eggs. The AI went rogue – without my approval . . . . I was a little frazzled when I realized what had happened: a bad AI decision had cost me real money.”58

iv.

Use of LLM-created Material to Fulfill a Contract

Once an agreement has been formed, AI may help to fulfill the terms of the contract. Yet again, this comes with risks. Sports Illustrated found itself under fire when it published articles by non-existent authors with AI-generated portraits and content potentially created by AI.59 While there might be future industry standards of use for article drafting so long as its use is disclosed, the role of AI was only discovered after an outside

investigation.60 While the legal repercussions from creating AI-based biographies for one’s publications without disclosure are unclear,61 Sports Illustrated felt reputational harm and pushback from its union, which “demand[ed] the company commit to adhering to basic journalistic standards, including not publishing computer-written stories by fake people.”62

v.

Use of LLM in Contract-Based Litigation

Finally, AI might be employed when litigating over an agreement. This may reduce litigation costs but again may bring difficult risks. While using AI to help clarify legal arguments or as a drafting assistant may be valuable, the possibility of AI hallucinating legal authority requires lawyerly oversight.63 Despite the relative newness of the technology, there are multiple examples. A Canadian attorney found herself apologizing to a court after citing fake case law in a child custody proceeding.64 In the United States, lawyers in Mata v. Avianca, Inc. were sanctioned after citing fake cases.65 One attorney stated, “I just was not thinking that the case could be fabricated . . . . My reaction was, ChatGPT is finding the case somewhere. Maybe it’s unpublished. Maybe it was appealed. Maybe access is difficult to get. I just never thought it could be made up.”66 Michael Cohen similarly submitted AI-generated fake cases from Google Bard when seeking to end a court-ordered probation.67 He avoided sanctions, but the

court reproved Cohen and his attorney for “embarrassing” and “negligent” behavior in their filing.68

C. A Model of Legal Risk in LLM Contracting

This subsection offers a model of legal risk based on the scenarios described above. It first considers those situations which involve AI-created content that is potentially mediated by a human before its use in the contracting process, such as when AI creates advertisements that a human then posts online. It then considers those agentive situations which do not involve human mediation, such as when a chatbot deals directly with a human or another chatbot and appears to engage in contractual behavior. Figure 1 presents this model in graphical form.69 This model encompasses both a broad definition of legal risk, as “significant legal consequences that flow from actions attributable to the business,”70 as well as a narrower definition of “risk originating in legal work product.”71

Figure 1: A Model of Legal Risk in AI-Based Contracting

The left side of the diagram is characterized by relatively low risk and has two characteristics: the use of AI subject to human review (as noted on the bottom), and the AI used is in documents that are related to the contract but are not inherently legal in nature (as noted at the top of the diagram).72 Because the use of AI is mediated by humans, the AI serves as a non-agentive tool which the human can review for helpfulness and

accuracy.73 While this does carry risk, such as if the human naively uses the output (such as in the Willy Wonka scenario74 and attorney briefing cases75 discussed in the prior section), the human may mediate that risk. Use of AI in developing and performing contracts are included in this category, as well as use of AI in litigation, although this category lies further to the right along the risk scale because now the LLM may be used to create legal documents.76

In our model, legal risk rises as the AI is used without human mediation—when it assumes, is granted, or appears to be granted agentive powers by the user. This use of AI comes with remarkable legal uncertainty as to basic questions such as whether it can legally bind the one deploying it,77 how AI fits into traditional notions of offer and acceptance, whether AI acts as an “agent” in the classical sense of principal/agent or master/servant law or merely as a type of super-scrivener,78 whether the AI can engage in puffery or false advertising, and so on.79 These uncertain legal issues bring a great deal of legal risk, as parties operate without knowing how courts will treat the interaction of humans and chatbots, or chatbots with chatbots on behalf of humans. Within this category, we list negotiation as less risky than an AI functioning as a human agent and purporting to make binding offers and acceptances on behalf of its deployers, simply because negotiating has less legal finality than forming or accepting contractual offers. The further a user of LLM is to the right of the diagram, the more caution they should employ, whether through employing LLMs with specialized training, instituting safeguards and review processes for LLM output, or declining to adopt the use of the LLM in certain legally-fraught situations.80

We now focus on the highest risk category from the prior section— when AI is negotiating, offering, and accepting on behalf of human beings.81 We present two contrasting views of the potential for LLMs to transform contract law in these areas, first as an expertise-driven, low-cost contractual utopia, and second, as a “slop”-filled82 morass of hallucination,83 manipulation,84 and meaning chaos.85

A. Adhesion Contracting

With the background of the prior section in place, consider the long-term potential for how LLM use might transform contract law, for good or ill. One area of modern contracting that has plagued both consumers and theorists alike is the use of adhesion contracts––massproduced contracts offered on a take-it-or-leave-it basis to other parties86–– such as cellphone service contracts in the United States. Adhesion contracts differ broadly from the roots of contracting in arms-length negotiation by parties on relatively equal footing.87 Adhesion contracts are popular

because the per-contract cost of drafting is low—once a contract is drafted by a business, it may be offered to thousands, tens of thousands, or millions of consumers for acceptance without negotiation.88 This implies that the drafters may tailor the contract with terms advantageous to themselves, with only potential market discipline substituting for negotiation.89 A contract law regime that historically assumed actual and mutual assent to negotiated terms becomes questionable when negotiation is not meaningful; when actual consent by the non-drafter is uncertain for contracts they had not read;90 when non-drafters would have difficulty understanding contractual terms if they had read them (for few would find it economical

to hire an attorney to explain a small service contract, for example);91 and when contracts concern essential reasonable alternatives.92 This lack of shared meaning has led one author to question whether adhesion contracts should even be considered “contracts” at all.93

services with

few

One solution to this problem in the literature is to remove adhesion contracts from the ordinary mechanics of contract law. For example, if adhesion contracts had terms moderated by consumer groups, some of the problems with one-sidedness stemming from the lack of meaningful negotiation could be mitigated.94 Another suggested improvement is to create adhesion contracts in self-executing computer code, increasing the incentive to cover all contractual eventualities rather than relying on customer-service flexibility in the face of adhesion contract problems.95 Alternatively, courts could create new legal doctrines that treat postcontractual customer service promises as contractual terms, giving the

weight of contract law to negotiations otherwise performed in the shadow of the contract itself.96

involved

the costs

The use of AI in adhesion contracts offers a more elegant (if risky) solution, one which preserves the essential nature of one-on-one negotiation adapted to a world of large power and resource imbalances between contracting parties. First, consider in simply understanding contractual terms. While the per-contract cost of employing expertise is already tiny for the drafters of adhesion contracts, AI may correspondingly reduce the cost of expertise for non-drafters. If AI can already pass the bar exam,97 it will likely be able to explain contractual terms and answer questions with reasonable accuracy and without technical jargon (although the possibilities of hallucination remain).98 As the cost of expertise their actually understanding contractual language increases, moving the boilerplate aspect of adhesion contracts (but not their unilateral nature) more towards the ideal of mutually understood and assented-to bargains.99 Second, AI offers inventive firms the possibility of extracting market value from bespoke agreements that otherwise would have required adhesion contracts to undertake. Just as the use of AI lowers expertise costs for non-drafters, it can lower costs for firms that wish to offer their consumers individualized contractual variability to gain competitive advantages.100

is reduced for consumers,

the chance of

In practice, this could mean using an LLM trained to understand the parameters of permissible contracts that negotiates with consumers, without the need for firms to pay skilled employees or attorneys.101 Businesses could come to individualized, surplus-generating agreements with the firms’ customers on a mass, but individualized basis. While the costs of having legal teams review thousands of individualized small consumer contracts would be prohibitive in the world of traditional legal expertise, having AI negotiate contracts that otherwise would have been offered as adhesion contracts may be possible.102

This could be positive. By lowering the costs of expertise, individualized bargains that otherwise might not have taken place may now occur. In law and economic analysis, transaction cost theory generally suggests that enabling bargains that otherwise might not occur is efficiencyenhancing.103 Contractual surplus—the value an agreement generates for a party above its costs—that otherwise would not exist would be generated and enjoyed by the parties.104

To see why this is the case, and why it would be desirable for a contracting regime to promote, consider a simple example in line with the chocolate factory theme from the introduction.105 Suppose a child wishes to buy an ice cream cone from an ice cream stand, which charges $0.75 for a

cone. Suppose also that the child is willing to pay up to $1.00 for this treat.106 If travel to the ice cream stand is costless, the difference between the price of the cone ($0.75) and the child’s willingness to pay ($1.00) is the child’s contractual surplus ($0.25). Suppose also that producing the ice cream cone costs the stand $0.50 in labor, equipment, and supplies. The difference between the sales price and the costs is the stand’s surplus, or profit, which is also $0.25. Because this exchange takes place, both the child and the business are better off than they otherwise would have been.107

If we insert transaction costs into this model—costs unrelated to the bargain itself, but which must be incurred to reach the bargain––the analysis can change dramatically. Suppose that instead of traveling for free, the child must spend a bus fare to reach the stand. If it costs the child $0.10 in fare to reach the stand, the child’s surplus––their benefit from this exchange––lowers to $0.15, the difference between their willingness to pay ($1.00) and the combined fare and price ($0.10 + $0.75). The child will still undergo the trip absent other, better alternatives, but is no longer as well off in a world with bus fare. The surplus available to the parties has shrunk, disfavoring the child.

Suppose now that the bus fare is raised to $0.50 to reach the stand. Now, the total costs to the child are $1.25 ($0.75 + $0.50), which is greater than the child’s willingness to pay for the ice cream, an item they only value at $1.00. If the child is rational, the bus trip is not undertaken, the bargain no longer takes place, and no contractual surplus is generated for the child or the stand.108 Both parties are worse off than they were in a world of costless travel. If the bus fare were reduced to its previous level, reducing the transaction costs required to buy the ice cream, it would again enable a market exchange to the benefit of all parties.

Consider this simple story first in the context of adhesion contracts, and then in the context of bespoke AI-negotiated exchanges. Adhesion contracting reduces the cost of contracting for firms in a significant way.109 When developing an adhesion contract, the costs of contracting become a

single fixed cost borne upfront by the business rather than a variable cost that scales with the number of contracting partners through negotiation and revision of contracting documents. Because this single fixed cost is spread out over thousands or millions of consumers, it becomes negligible, shrinking the per-consumer costs of the firm to a de minimis amount. This is a foundational point that enables modern mass-contracting—and much of modern commerce in general––like the free travel to the ice cream stand. Were businesses required to negotiate over every modern consumer contract in the same manner as bespoke business-to-business agreements, the transaction costs would either prevent those bargains from occurring at all, like the $0.50 bus fare did for our ice cream stand example, or lead society to substitute something other than contract as the legal regime to govern those types of business transactions.110 Because businesses do wish to capture the economic potential from these types of transactions, and because contract law is sufficiently flexible to allow them, adhesion contracting became ubiquitous.

There are, however, drawbacks. Because consumers do not negotiate these contracts, rational firms attempting to extract as much surplus as possible from contractual exchanges will structure adhesion contracts to maximize their profits.111 If we model a contract as constituting some number of terms which may either favor the firm or the customer, then the firm will add firm-favorable one-sided terms to contractual language to the point that the expected benefit from additional one-sided terms equals the expected loss from customers who find the terms unfair and choose not to accept the adhesion contract.112 Because most boundedly

rational customers are unlikely to read the terms of the contract, and because many adhesion contracts concern what consumers deem essential services, the expected customer loss from one-sided terms may be quite small, and the number of one-sided terms thus correspondingly large.113 When these disadvantageous terms are spread out over millions of consumer contracts across every major area of the economy, it represents a potentially massive transfer of contractual surplus from consumers to firms. Thus, while the reduction in transaction costs associated with adhesion contracting enables many bargains that otherwise would not take place, the surplus from those bargains tends to move towards firms rather than consumers.114

B. Contractual Utopia: Anti-Adhesive Contracts Mediated by

LLMs

Consider now what might occur with AI-intermediated interactions between consumers and firms in areas that are currently governed by adhesion contracts. The crucial difference between adhesive and nonadhesive contracts is the opportunity to negotiate over contractual terms. In a typical non-adhesion contract, rational parties negotiate contractual terms until the expected costs from further negotiation exceed the expected benefits from additional alterations to the contract.115 The costs of negotiation can be high. First, the negotiation process itself takes time and effort. One must actually read agreements, consider eventualities, decide how to incorporate those eventualities into an agreement, and so on. Second, the contractual negotiation process often requires legal expertise, which is expensive.116

Generative AI could transform both these aspects of negotiation, opening the possibility of bespoke agreements produced at the cost and scale of adhesion contracts. We refer to this type of potential contract as an “anti-adhesive” agreement. In an anti-adhesive agreement, an LLM trained on company policies and preferences could negotiate with an LLM

deployed by a consumer, likewise trained on the consumer’s contractual preferences.117 Such anti-adhesive contracts would be drafted at negligible costs, similar to traditional adhesive agreements. As with those contracts, the human costs of negotiation would be largely eliminated, and a sufficiently skilled chatbot could provide the legal acumen for both the consumer and company interests.

Both the consumer and the company would enjoy greater contractual surplus as their bespoke agreement more specifically captures contractual intentions than possible in an adhesive agreement.118 The company could enjoy a competitive advantage in contractual flexibility.119 From this larger contractual pie, the consumer could extract more surplus than possible under a traditional adhesive agreement. The consumer’s AI, armed with the consumer’s preferences, could negotiate with the company’s AI, and with more surplus available, could come to terms of an offer that draws additional benefits to the consumer and then accept it on their behalf. An AI acting on behalf of many consumers in simultaneous virtual negotiations could bring market power to bear to further capture surplus for consumers.120

Thus, if firms find value in this form of contracting, deployment of agentive AI can both expand the contractual “pie”—the surplus available to the parties—and simultaneously increase the contractual surplus flowing to consumers. A world of contracts in which the benefits of adhesive contracts (low costs of contracting) have their worst elements removed (onesidedness and lack of negotiability) would be a utopian “Shangri-Law” indeed.121 The next section explores an alternative to this utopian possibility.

C.

Contractual Dystopia—Hallucination, Manipulation, and Meaning Chaos

With the ideal world of the past section in mind, consider now an alternative version of a future in which LLM-based contracting is widespread. Rather than a contractual utopia, contracting could become decidedly dystopian. We highlight three problems: hallucination, the problem of generative AI essentially making things up that might not reflect the contractual intent of its user; manipulation, the problem of generative AI being subject to clever users who can influence the AI to create content against the intent of its deployer; and finally the broader problem we term “meaning chaos,” in which AI creates language that distorts the contractual intent of its users and makes contractual terms overly punitive, vague, or otherwise legally “sloppy” in ways that distort contractual intent.122 In each case, the prior subsections’ utopia may be fleeting.

transaction-cost savings undergirding

the promised

Generative AI Hallucination and Contracts

An AI “hallucinates” or “confabulates” when

it generates misleading or incorrect information.123 This occurs because an AI system is not truly intelligent—its output simply reflects the results of large statistical models that predict output language based on input language.124 While the results can be both useful and astounding, statistical predictions of human language do not inherently involve fact-checking, reasoned judgment of the outputs, or many other aspects of human intelligence.125 As a result, LLM output may contradict itself, state something as factual that is not, and create output that is irrelevant to the prompt.126 For example, if an AI is used to generate a written contract, contractual terms may be drafted that do not reflect the input prompt, are internally inconsistent, or are simply irrelevant.

the airline example from

This creates problems that are both practical for business and challenging for contract as a body of law. From a practical perspective, hallucination means headaches. In the introduction, AI hallucination created misplaced expectations that harmed the business (by promising lower fares that were not part of the company’s intent) and the consumer (by promising information that resulted in the need for later litigation).127 Similarly, technology that aims to create legal documents but cannot tell when those legal documents are adequate means other costly forms of supervision are required to check the work of the LLM.

These practical problems are matched by what happens to contract law itself as documents are drafted by entities that hallucinate. Contract law is based in agreement.128 This undergirds every principle of contract law, from the traditional legal mechanisms of offer and acceptance to the economic and moral reasons why contract exists as a body of law.129 The contractual ideal is a completely voluntary, completely understood, completely agreed-to bargain that creates value for both parties.130 When this voluntary “meeting of the minds” in “mirror-image” form exists, contract law is simple.131 If a contract is breached, the language of the

contract is applied and damages awarded.132 This is because a moral social order puts value on enforcing agreements and because it is efficient to do so.133

The more that written documents fail to capture contractual intent because of hallucination, the farther they move away from this contractual ideal. The efficiency inherent in providing an enforcement mechanism for parties pursuing their own self-interest becomes something else entirely. It becomes a system that must separate out actual intent from what was hallucinated (such as a scrivener’s error134), enforce the hallucination against the intentions of one party (such as for a unilateral mistake135), or— most concerning—enforce the hallucination against the intent of both parties (such as when third party beneficiaries have rights under an agreedto but hallucinated document and choose to enforce it against the wishes of each contractual party).136 The further contract law moves from enforcing

intent to enforcing the imaginations of a statistical model, the less social value the doctrine creates.

User Manipulation of Generative AI

in

the The second problem with deploying generative AI contracting process is the problem of user manipulation of the model. While hallucination was the AI model itself producing false content as an artifact of its training data or model development, manipulation involves a user deliberately attempting to produce AI output that would be considered false, misleading, or—in our context—against the contractual intentions of the other party.137 For example, the illustration above in which a user instructed a chatbot to sell it a car for a dollar and inform the user that this is a legally binding agreement would represent user manipulation against the contractual intentions of the one deploying the bot.138

This possibility should again lead to caution. From a practical standpoint, the ability of a user to actively manipulate the output of a program that is even ostensibly acting as an agent or providing information on behalf of a company is disturbing. A company deploying a tool that can be actively manipulated faces the prospect of consumers alleging numerous bargains, contractual alterations, waivers of damages, and so on, that cut against the company’s contractual intentions.

in which pressure or

From a contractual-theoretic perspective, this is also troubling. In a human-to-human interaction, manipulation of a bargain against the true intentions of the contracting parties would be policed by doctrines such as duress and undue influence.139 Because the agreement at the heart of contract law assumes voluntariness, these doctrines intervene to undo the agreement bargains involuntary.140 In a human-AI transaction in which the human is able to manipulate the model to generate an offer against the intentions of the one deploying it, it is unclear how these doctrines would apply. The company presumably voluntarily agreed to deploy the generative AI as a type of electronic agent acting on its behalf, subject to the specific training data and model the company chose to deploy. Without any duress or undue influence, it is unclear how traditional contractual doctrines limit this behavior, unless those doctrines are expanded to the idea of the

influence rendered

manipulation of computer code.141 A user acting within that model, although generating outputs against the wishes of the company, is nevertheless operating within the framework provided by the other party. In a sense, so long as they play by the rules of the model, they may be legally blameless for the model’s outputs. Yet a contracting regime that allows a party’s intentions to be so easily abandoned would be weak indeed.

“Meaning Chaos”

Our last category of contractual dystopia is related to the idea of hallucination, but looks more broadly to the type of output generated by LLMs. While hallucination involves generative AI creating contractual content contrary to the wishes of its user, those hallucinations may be contrary to intent in easily definable ways. For instance, the company wishes for there to be no post hoc bereavement fares, and the AI tells the user that such bereavement fares are possible, hallucinating in a way that makes it easy to discern the true contractual intentions of the party once the hallucination is manifest.142 While this is problematic, it may be relatively easy to decipher the true contractual intent of the parties, even if a remedy favors one party or the other. A more subtle problem occurs when the LLM generates terms that neither party would agree to or that are sufficiently vague that neither party understands what they mean.143 If AI-generated contractual language becomes so vague, random, or punitive that it reflects neither party’s intent, the transaction costs of implementing the contract through AI may rise rather than decrease, as both parties would be faced with verbal slop that reflects neither’s understanding of their agreement nor that of an agreement they would have wished their agentive AI to undertake on their behalf.

This section provides an analytical framework to suggest why the issue of “meaning chaos” in the last section is of particular concern for the future of contract law. The framework focuses on the level of actual agreement between contracting parties and the costs required to enforce that agreement.

A. Contractual Distance and Entropy

Contractual “distance” is a shorthand term for the level of rational actual agreement between the parties to a contract.144 In a very low-distance contract, the parties rationally understand all of what they have agreed to, whether because the contract is very simple or because it is complex but with such significant consequences that the parties rationally invest the time to understand it. 145 An example might be our illustrative ice cream sale, in which a simple exchange of money for a listed price occurs upon purchase, or a complex commercial transaction with sufficiently high stakes for both sides that nothing is left to boilerplate.146 Although much contract law might apply to disputes over these agreements, the terms of the contractual exchange are well-understood. In contrast, a high distance contract is one in which at least one of the parties rationally does not understand the terms of the agreement.

An adhesion contract for a wireless contract is a good example. For a typical consumer of this type of contract, it would be irrational to study the contract in the depth required to understand it.147 Consumers know that they will face a similar slate of terms from any wireless company, consider cell service as something they must have, and know that they will likely resolve contractual concerns with a customer service agent.148 The consumer likely does not read, and likely would not rationally attempt to understand many contractual terms, such as choice of law provisions, integration clauses, force majeure provisions, and so on.149 Since there is little actual “meeting of the minds” as to much of the agreement (other than the price term and the essential provision of services), we refer to this as a “high distance” contract. The minds of the parties are separated by a great distance as to many contractual terms, figuratively speaking. As discussed above under the issue of hallucination, contracts that move away from

intent move away from efficiently reflecting the desires of the party and the social and economic justifications for their enforcement.150

Figure 2: Grid of Contractual Distance and Contractual Entropy

Figure 2 pairs the concept of contractual distance with a second concept borrowed from information science—the idea of “contractual entropy.” In information science, entropy is a measure of uncertainty, and in physics, of disorder and energy.151 For random variables, an event that will occur with certainty has very low entropy, while an event with many equally likely but different outcomes will have high entropy.152 For instance, the entropy from rolling a die is greater than the entropy from choosing a die face and placing it on the table.153 For physical systems, higher entropy means higher disorder, and only energy entering the system may decrease the disorder or entropy.154 For example, think of a home with small children. If no energy is put into cleaning and organizing the home, it tends to quickly become disordered—a high entropy state. It requires an input of energy from caregivers to wash dishes, sweep floors, and organize toys to return the home to a state of order—a lower entropy state.155

In contrast, some contractual situations may be analogous to choosing a die face and placing the die on the table—there is only one possible outcome of the contract, which will occur with factual certainty

and without legal dispute. Other contracts may be like rolling a die—legal uncertainty may require a judge to decide on a contractual outcome, and the judge could rule either way. Other contracts may be like rolling many dice—many unknown factual events may occur, and many legal issues may be unresolved prior to litigation. Like the home with small children, these “high entropy” contracts take substantial energy to reduce to a state of order—energy from negotiations, factual discovery, legal fees, input from courts or mediators, and so on.

High-entropy agreements bring many challenges to contracting. Rational parties now act in a realm of expected probability of outcomes rather than certainties. They must hedge their contractual bets, plan for multiple outcomes, and invest in mechanisms to safeguard the various ways the contract could be interpreted or executed. Relative to low-entropy contracts, these agreements are costly for the parties, who only by planning, post-contractual negotiation, or litigation, will slowly resolve the results of their agreement. In our transaction cost example, if the bus ride had chances of failing, the child would need back-up arrangements, such as fare for a different bus, investigation of alternative routes, and so on, that add to the cost of the trip and reduce the net benefit from the ice cream. Uncertainty itself becomes a transaction cost that limits the ability to enter efficient bargains.156

When an agreement is both high-distance and high-entropy, the parties no longer understand the terms of their agreement and cannot track what they do understand to a certain outcome. At this point, the contract is hardly a contract at all, in the sense that what contract law purports to do in providing legal enforcement mechanisms that deliver one’s contractual intent may no longer occur.157 An AI-based contract that has exploded into “meaning chaos” falls neatly into this category.158 AI-drafted slop fills the agreement rather than reflections of the parties’ intentions, and how the law should approach those terms is unclear. The contract is both distant from a meeting of the minds and deeply uncertain in how it will be treated as a contract.

B. Tort as a Remedy for High-Distance, High-Entropy Contracts

The last subsection proposed that LLM-generated contracts filled with “meaning chaos” begin to lose the characteristics of agreements that make contract law well-suited to governing disputes.159 In other words, although filled with the language that looks contractual, LLM-based language should perhaps not be conceived as contractual at all. Perhaps it should be governed by tort, rather than by contract.160

This is, after all, precisely what we have seen in this Article. When a plaintiff sought a remedy against Air Canada, the dispute was contractual in nature—a promise had been made and, to the plaintiff, broken—but rather than attempt to parse out the role AI had played as agent and attempting to place that role in some existing application of contract law, the court looked outside contract to a simpler remedy and found the airline liable in tort.161 Tort duties, such as to not commit fraud, lie in the background of contract anyway, and they provide a pleasing way to arrive at a sense of justice without addressing the more complicated legal questions about agentive AI.

This could be because the court was lazy—tort provided a satisfying conclusion, and deciphering whether agentive AI could function as a legal agent would have been complicated. This could become typical in this area. In the absence of satisfying applications of contract law, courts faced with LLM-based disputes may turn to the next ready body of law and apply principles of tort to resolve disputes. In the same way that “nature abhors a vacuum,” law—in a sense—abhors disputes it cannot resolve and will, over time, develop ways to resolve them.162 If tort provides a simple remedy that appears to satisfy notions of fair play and justice, then courts will use it.163

Or, rather than being lazy, perhaps the court was on to something. As agentive AI worms its way into agreements, perhaps tort really does provide a more satisfying remedy. Consider two reasons why this might be

so, and then one counterargument. First, tort may provide more satisfying remedies because LLM-based agreements in high-distance, high-entropy situations stretch the principles for which contract law was created to areas they are ill-equipped to handle. There may be little actual agreement (high contractual distance), and as such, contract law set up to enforce agreements may be less applicable than simply looking to the tort of misrepresentation.164 Attempting to reconstruct the intent of parties translated through non-human intermediaries is a costly and uncertain process (high contractual entropy), and perhaps it is best to simply avoid it.165

Second, there is no certainty that the typical contractual remedy of expectation damages will be available to parties faced with the uncertainty of agentive AI. Contract aims to put parties in the same position as if the contract had been fulfilled—to satisfy their contractual intentions—but this remedy fails when that outcome cannot be calculated with reasonable certainty.166 It is unclear how it would apply when the agreement itself makes the parties’ actual contractual intentions unclear. In other words, the more the agreement looks like meaning chaos, the less a court would be able to ascertain what the contractual expectations are, and so could not offer expectation damages to an aggrieved party. In such cases, it may be more effective simply to assess whether one party has damaged the other under principles of tort law.

Of course, as a counterargument, contract law has proven itself adaptable over the centuries.167 It now accommodates blockchain-based deals and end-user license agreements that could hardly have been imagined by those making agreements in medieval England. If contract has persisted into the modern era in a way that satisfies parties and continues to be relevant, perhaps it will adapt to agentive AI in ways sufficient to address emerging challenges. The law may develop ways to classify and

treat the deployment of LLM-based agents in ways analogous to human counterparts, addressing situations where an LLM-based agent violates instructions or is manipulated. Technology could improve to eliminate the problem of hallucination and meaning chaos, which would simplify the tasks of courts. Time will tell.

We conclude with a thought on the future of AI and human relationships. Economic efficiency is not the only value that contract contemplates. Contract law is built on the idea of human relationships and the value those relationships create for each other. In making agreements with one another, we express something of what it is to live in conviviality. We need each other’s goods, services, ideas, and community to live full lives. Contract is the legal extension of this social value—every handshake that concludes an agreement means two humans have committed not just their resources but themselves to a mutually beneficial bargain. The farther contracts move from that handshake to faceless forms and negotiations with the nameless, the farther they move from strengthening human conviviality. In examining AI’s future role in contract law, Ivan Illich’s concept of “tools for conviviality,” which began this Article, offers valuable insights.168 Conviviality, in Illich’s framework, represents a state where tools serve human needs and relationships rather than subordinating them to institutional or industrial demands.169 A convivial tool enhances individual freedom, enables autonomous and creative intercourse among persons, and expands the range of each person’s control over their environment.170 This framework proves particularly relevant as AI reshapes contractual interactions, raising fundamental questions about human agency and relationships in an increasingly automated world.

The promise of AI as a convivial tool in contract law lies in its potential to democratize legal expertise and reduce barriers to meaningful agreement.171 When deployed thoughtfully, AI could enhance human understanding and agency in contractual relationships rather than merely automate them. This vision aligns with the concept of anti-adhesive contracts, where AI might enable truly negotiated agreements at scale while preserving the essential human elements of bargaining.172 Yet this potential can only be realized if AI remains subordinate to human intention and relationship rather than becoming an end in itself.

The risk, as Illich might observe, is that AI becomes a tool that creates dependency rather than autonomy.173 Our analysis of contractual distance and entropy illustrates this danger—AI could further separate parties from genuine understanding and agreement—replacing human conviviality with automated processes that merely simulate consensus. The “meaning chaos” we identified in AI-generated contracts represents exactly the kind of tool-induced alienation that Illich warned against.

To preserve the law’s role in fostering human relationships while embracing AI’s benefits, we must develop legal frameworks that promote convivial AI use. One approach is the liberal application of tort law to police areas where contract may fall short.174 This means ensuring AI remains controllable by and accountable to human users by putting legal incentives in place that hold entities deploying AI accountable for its actions, enhancing rather than replacing human agency in contractual relationships. Other regulatory approaches should prioritize transparency, human oversight, and meaningful consent—preserving the essential human elements of contracting while leveraging AI's capabilities to reduce transaction costs and increase access to legal expertise.

It is likely unavoidable that the efficiencies inherent in generative AI will continue to influence the human values at the heart of contract law. We hope that as it does so, it will enable efficient bargains that otherwise might not take place, empower consumers, and provide new means of human flourishing that will supplement, rather than supplant, those that existed when contract law first developed.

--- FOOTNOTES ---

Part of the Contracts Commons, and the Legal Writing and Research Commons

*

Jeff Lingwall is an Associate Professor of Legal Studies at Boise State University.

Steven Hyde is an Associate Professor of Management at Boise State University. We thank participants at the 2024 CALSB/PNALSB meetings for helpful comments, and Shavonne Bostaph and Hayden Singelyn for able assistance. Any errors are our own.

INTRODUCTION

IVAN ILLICH, TOOLS FOR CONVIVIALITY 25 (1973).

1. WILLY WONKA & THE CHOCOLATE FACTORY (Paramount Pictures 1971). 2. 3. We use the term “AI,” “generative AI,” “LLM,” “large language model,” “chatbot,” “bot,” and “AI chatbot” interchangeably, each meaning the human utilization of a large language model such as ChatGPT that can produce text and text-based interaction that mimics the responses of a human being. See Harry Surden, ChatGPT, AI Large Language Models, and Law, 92 FORDHAM L. REV. 1941, 1942–43 (2024) (surveying the recent development of LLMs). See generally Eliza Mik, Caveat Lector: Large Language Models in Legal Practice, 19 RUTGERS BUS. L. REV. 70, 72 (2024) (examining “LLMs from the perspective of their limitations and from the perspective of their users” in legal work); Brian S. Haney, Applied Natural Language Processing for Law Practice, 2020 B.C. INTELL. PROP. & TECH. F. 1, 3 (2020) (noting a series of technological changes, including AI systems, and their impact on practice of law).

4. See Kevin M. Teeven, Problems of Proof and Early English Contract Law, 15 CAMBRIAN L. REV. 52, 56 (1984) (“The magical steps required by covenant [involving a sealed, written document] were . . . unreasonable for . . . simple transactions. The evidentiary burden was frequently too great. . . . Covenant became a veritable dead letter for such transactions because of the impractical documentation demands it made.”). For an introduction to the common law’s early formalities for contract, see JOHN H. LANGBEIN, RENÉE LETTOW LERNER & BRUCE P. SMITH, THE HISTORY OF THE COMMON LAW: THE DEVELOPMENT OF ANGLO-AMERICAN LEGAL INSTITUTIONS 322 (2009).

5. See Cheryl B. Preston & Eli McCann, Llewellyn Slept Here: A Short History of Sticky Contracts and Feudalism, 91 OR. L. REV. 129, 143 (2012) (“With the proliferation of the merchant class and the development of high-speed printing presses, various industries began providing customers with pre-printed statements of contract terms and bypassing a signature requirement or other individualized expression of assent.”) (footnote omitted). The practice of creating agreements where one party may not have any “reasonable choice” other than to contract has, of course, existed long before printing allowed mass-production of preprinted terms. Id. at 142 (“Parties may have signed, sealed, and sworn in front of witnesses their agreement to certain contract terms as a result of various life exigencies, or express or implied duress, however subtle.”).

6. See Michael J. Madison, Rights of Access and the Shape of the Internet, 44 B.C. L. REV. 433, 436 (2003) (noting development of legal processes for “shrinkwrap” and “clickthrough caselaw”). Beyond click-through law, “smart” contracts that self-execute on public ledgers represent contract law coping with yet another transformative technology. See Jeff Lingwall & Ramya Mogallapu, Should Code Be Law? Smart Contracts, Blockchain, and Boilerplate, 88 UMKC L. REV. 285, 288–89 (2019) (discussing potential for blockchain technology to enable transactions).

7. The term “agentive AI” means “a software program that can interact with its environment, collect data, and use the data to perform self-determined tasks to meet predetermined goals.” AMAZON, What Are AI Agents?, https://aws.amazon.com/what-is/aiagents/ [https://perma.cc/MYY7-BVK4] (last visited Jan. 21, 2025). While a human sets the aims of the software, “an AI agent independently chooses the best actions it needs to perform to achieve those goals.” Id. This is not a near-future situation, it is currently available for consumer use. See Geoffrey A. Fowler, I Let ChatGPT’s New ‘Agent’ Manage My Life. (Feb. 7, 2025), https://www.washingtonpost.com/technology/2025/02/07/openai-operator-ai-agent-chatgpt/ [https://perma.cc/G7P6-PUEM].

It Spent $31 on a Dozen Eggs., WASH. POST

8. See Federica Bianchi, Patrick John Chia, Mert Yuksekgonul, Jacopo Tagliabue, Dan Jurafsky & James Zou, How Well Can LLMs Negotiate? Negotiation Arena Platform and https://arxiv.org/pdf/2402.05863 [https://perma.cc/VCA9-DU7X].

Analysis,

ARXIV

8,

9.

“Shangri-La” as a metaphor for a utopian condition is common. E.g., Michael

Kirby, Media Law Reform — Beyond Shangri-La, 6 J. MEDIA L. & PRAC. 117 (1985).

10. See Bianchi et al., supra note 8; Uri Benoliel & Shmuel I. Becher, The Duty to Read the Unreadable, 60 B.C. L. REV. 2255, 2296 (2019) (examining whether consumers can actually understand modern contractual offers, through empirical analysis of sign-inwrap agreements, and concluding that “[l]acking a clear and strong incentive to draft readable agreements, firms often utilize unreadable texts as their contracts.”).

11. The timeline at which LLMs become this useful and accessible is debatable. See Mik, supra note 3, at 71 (“The technologies discussed below are complex, but the message is simple: for the foreseeable future, lawyers should beware of relying on any output generated by LLMs.”).

12. See Michael I. Swygert & Katherine Earle Yanes, A Primer on the Coase Theorem: Making Law in a World of Zero Transaction Costs, 11 DEPAUL BUS. L.J. 1, 4 (1998) (“Given perfect knowledge about all alternatives to any problem, and assuming transaction costs are zero, disputants will always rearrange their rights, liabilities, and entitlements in a manner which produces a net gain in their combined well-being.”). But see Robert Cooter, The Cost of Coase, 11 J. LEGAL STUD. 1, 23 (1982) (noting that lowering transaction costs in the context of legal settlements may reduce incentives to settle by lowering the costs of settlement negotiations themselves).

13. See Eric A. Posner, Economic Analysis of Contract Law after Three Decades: Success or Failure, 112 YALE L.J. 829, 832–33 (2003) (discussing assumptions behind economic analysis of contract law as stemming from “jointly beneficial” projects); How to Ensure Africa is not Left Behind by the AI Revolution, THE ECONOMIST (July 25, 2024), https://www.economist.com/leaders/2024/07/25/how-to-ensure-africa-is-not-left-behind-bythe-ai-revolution [https://perma.cc/3QXQ-KY25] (“Africa’s lack of [internet] connectivity is one reason its people could miss out on the benefits promised by artificial intelligence (AI).”).

14. See infra Section II.A.

15. As with any potentially pathbreaking technology, the advent of LLMs have brought a host of unsettled legal issues, including many not directly related to contracts. These include whether training LLMs violates intellectual property law, privacy issues with data uploaded to query LLM models, and whether companies running LLMs owe duties of care based on the information provided to its users. These issues have been recognized in popular press, for example in Eoghan Ryan, What are the Legal Implications of ChatGPT, SCRIBBR (Nov. 16, 2023), https://www.scribbr.com/ai-tools/legal-implications-chatgpt/ [https://perma.cc/TT3D-QSMC]; Tom Corfman, The Legal Risks of ChatGPT, RAGAN (July 10, 2023), https://www.ragan.com/the-legal-risks-of-chatgpt/ [https://perma.cc/5Q2Q4YUM]; Andrew Perlman, The Implications of ChatGPT for Legal Services and Society, HARV. L. SCH. CTR. ON THE LEGAL PRO. (2023), https://clp.law.harvard.edu/knowledgehub/magazine/issues/generative-ai-in-the-legal-profession/the-implications-of-chatgpt-forlegal-services-and-society/ [https://perma.cc/63CR-APFQ], as well as in the academic literature. E.g., Nicola Lucchi, ChatGPT: A Case Study on Copyright Challenges for Generative Artificial Intelligence Systems, EUR. J. RISK REGUL. 1 (2023); Xiaoze Liu, Ting Sun, Tianyang Xu, Feijie Wu, Cunxiang Wang, Xiaoqian Wang & Jing Gao, SHIELD: Evaluation and Defense Strategies for Copyright Compliance in LLM Text Generation, ARXIV (Feb. 8, 2024), https://arxiv.org/abs/2406.12975 [https://perma.cc/WQZ6-5L3U].

16. See Lawrence M. Solan, Contract as Agreement, 83 NOTRE DAME L. REV. 353, 354 (2007) (discussing the foundations of contract law in mutual assent). But see Val Ricks, Assent Is Not an Element of Contract Formation, 61 U. KAN. L. REV. 591, 593 (2013) (arguing that assent should be subsumed within consideration as an element of contract formation). For a broad debate on the existence and efficiency of assent in adhesion contracts, compare MARGARET JANE RADIN, BOILERPLATE: THE FINE PRINT, VANISHING RIGHTS, AND THE RULE OF LAW (2013) (arguing against efficiency as a justification for adhesion contracts) with Steven W. Feldman, Mutual Assent, Normative Degradation, and Mass Market Standard Form Contracts - A Two-Part Critique of Boilerplate: The Fine Print, Vanishing Rights and the Rule of Law (Part I), 62 CLEV. ST. L. REV. 373, passim (2014) (critiquing Radin’s view).

17. See infra Section I.C. 18. See infra Section III.A.

BACKGROUND

19. The “death of contract” and its social value is the subject of some discussion in the literature. See Robert E. Scott, The Death of Contract Law, 54 U. TORONTO L.J. 369, 379 (2004) (“The major social cost of a wholesale exit from the public contract law is the loss of the social benefits that derive from judicial resolution of disputed agreements and the subsequent publication of their decision.”). See generally Leon E. Trakman, Public Responsibilities beyond Consent: Rethinking Contract Theory, 45 HOFSTRA L. REV. 217, 218–19 (2016) (arguing for public responsibilities to complement private interests in contract “to protect important moral and social interests”); Daniel A. Farber, Contract Law and Modern Economic Theory, 78 NW. U. L. REV. 303, 305–06 (1983) (noting the importance of limitations on contract law to function as a safety net in light of boundedly rational ability to contract for future events). See generally D. P. Waddilove, Anticontract, 61 AM. BUS. L.J. 135, 135–136 (2024) (examining how government may essentially eliminate normal contract law in times of widespread instability).

20. Carnegie Mellon University, Researchers Highlight Nobel-Winning AI Breakthroughs and Call for Interdisciplinary Innovation, SCIENCEDAILY (Nov. 26, 2024), https://www.sciencedaily.com/releases/2024/11/241126134826.htm [https://perma.cc/V88GWHR2].

21. History of Artificial Intelligence (A.I.), ILL. CENT. COLL. LIBR., (Jan. 30, 2025)

https://library.icc.edu/c.php?g=1372140&p=10141462 [https://perma.cc/3PD6-3SPY].

22. Ashish Vaswani et al, Attention is All You Need, in PROCEEDINGS OF THE 31ST INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS 5998 (Ulrike von Luxburg, Isabella Guyon, Samy Bengio, Hanna Wallach & Rob Fergus, eds., 2017).

23. Richard E. Turner, An Introduction to Transformers, ARXIV (Apr. 20, 2023),

https://arxiv.org/abs/2304.10557 [https://perma.cc/5N6X-VU4G].

24. See Zheyi Chen et al., Evolution and Prospects of Foundation Models: From Large Language Models to Large Multimodal Models, 80 COMPUT. MATERIALS & CONTINUA 1753, 1755 (2024).

25. See id. In a sense, this represents a triumph of common-law learning rather than

civil-law dictation of rules. 26. Id. at 1753, 1755. 27. Paul Colmer & Xiaoyu Xing, A Gentle Introduction to Transformer Architecture 2024),

and Relevance https://community.aws/content/2eD4ehf1d5hatTW1dnVErcJVVPL/a-gentle-intro-totransformer-and-gen-ai [https://perma.cc/4U4J-DSVQ].

to Generative AI, AWS COMMUNITY

(May

12,

28. Pankaj Singh, LLMs Research Paper in January 2025: Breakthroughs in Size (Dec. 9, 2024),

Reduction and Enhanced Performance, ANALYTICS VIDHYA https://www.analyticsvidhya.com/blog/2024/02/llms-research-paper-in-january/ [https://perma.cc/ZD2G-Y26D].

29. See Radu Stefan, George Carutasu & Marian Mocan, Ethical Considerations in the Implementation and Usage of Large Language Models, 2 THE 17TH INTERNATIONAL CONFERENCE INTERDISCIPLINARITY IN ENGINEERING 131, 131 (2024).

30. April G. Dawson, Algorithmic Adjudication and Constitutional AI—The Promise

of a Better AI Decision-Making Future?, 27 SMU SCI. & TECH. L. REV. 11 (2024).

31. GOOGLE,

PALM

TECHNICAL

REPORT

9,

19–20

(2023),

https://ai.google/static/documents/palm2techreport.pdf [https://perma.cc/2JTA-LUX6].

32. Konstantinos I. Roumeliotis, Nikolaos D. Tselikas & Dimitrios K. Nasiopoulos, Llama 2: Early Adopters’ Utilization of Meta's New Open-Source Pretrained Model, PREPRINTS https://doi.org/10.20944/preprints202307.2142.v2 [https://perma.cc/5Z5L-7ZDU].

(2023),

33. See Ahmadreza Tavasoli, Maedeh Sharbaf & Seyed Mohamad Madani, Responsible Innovation: A Strategic Framework for Financial LLM Integration, ARXIV (Apr. 2, 2025), https://arxiv.org/abs/2504.02165 [https://perma.cc/N476-FMSC].

34. Chengyuan Deng et al., Deconstructing the Ethics of Large Language Models to New-emerging Dilemmas, ARXIV (Oct. 21, 2024),

from Long-standing Issues https://arxiv.org/abs/2406.05392 [https://perma.cc/AV4H-NKWP].

35. See Jonathon Ciottone, A Lawyer’s Guide to Understanding AI Hallucinations in a Closed System, NAT’L L. REV. (June 26, 2024), https://natlawreview.com/article/lawyersguide-understanding-ai-hallucinations-closed-system [https://perma.cc/NJC2-T99]; Steve Taplin, AI Hallucinations: How Can Businesses Mitigate Their Impact?, FORBES (Aug. 15, 2024), https://www.forbes.com/councils/forbestechcouncil/2024/08/15/ai-hallucinationshow-can-businesses-mitigate-their-impact/ [https://perma.cc/V6AV-WTGG].

36. See, e.g., Tim Mucci, The Future of Artificial Intelligence, IBM (Oct. 11, 2024), https://www.ibm.com/think/insights/artificial-intelligence-future [https://perma.cc/A5DCPR9L] for a discussion on modern AI initiatives. See also, for example, Max Roser, The brief History of Artificial Intelligence: The World has Changed Fast — What Might be Next?, OUR WORLD IN DATA (Dec. 6, 2022), https://ourworldindata.org/brief-history-of-ai [https://perma.cc/3MSK-MT9G] for a discussion of past developments in AI architecture and training.

37. WORLD ECONOMIC FORUM, AI VALUE ALIGNMENT: GUIDING ARTIFICIAL 2024),

INTELLIGENCE https://www3.weforum.org/docs/WEF_AI_Value_Alignment_2024.pdf [https://perma.cc/9T3D-JXP4].

TOWARDS

SHARED

HUMAN

GOALS

(Oct.

38. Mallory Moench, How a Willy Wonka Event in the U.K. Ended in Calls to Police and Demands For Refunds, TIME (Feb. 28, 2024), https://time.com/6836088/willy-wonkaevent-scandal-glasgow-refunds-police/ [https://perma.cc/3KD9-YCKX].

I’ve

39. See, e.g., Natalie Fear, This Disastrous Wonka Event is the Worst Case of AI Art 2024), Catfishing https://www.creativebloq.com/news/bad-willy-wonka-experience [https://perma.cc/4H2B7X32] (“[The experience] was revealed to be a complete catfish, using AI-generated images to lure in unsuspecting punters. The underwhelming warehouse full of vaguely Wonkathemed props needed a lot more than pure imagination to live up to expectations . . . .”).

CREATIVE

BLOQ

Seen,

27,

40. Id.; Moench, supra note 38. 41. Morven Mckinnon, Police Called to Willy Wonka Event After Refunds Demanded, https://www.bbc.com/news/articles/czrlr3dw2x3o

26,

BBC [https://perma.cc/K9C6-MTFR].

in

‘Remarkable’ Lying AI Chatbot Case, FORBES

42. Id.; Fear, supra note 39. 43. This was also covered extensively in press. See Marisa Garcia, What Air Canada (Feb. 19, 2024),

Lost https://www.forbes.com/sites/marisagarcia/2024/02/19/what-air-canada-lost-in-remarkablelying-ai-chatbot-case/?sh=54c7b7cb696f [https://perma.cc/H3RH-YDTT]; Maria Yagoda, Airline Held Liable for its Chatbot Giving Passenger Bad Advice – What This Means for Travellers, BBC (Feb. 23, 2024), https://www.bbc.com/travel/article/20240222-air-canadachatbot-misinformation-what-travellers-should-know [https://perma.cc/KS26-9CXH]; Nick Robertson, Air Canada Must Pay Refund Promised by AI Chatbot, Tribunal Rules, THE HILL (Feb. https://thehill.com/business/4476307-air-canada-must-pay-refundpromised-by-ai-chatbot-tribunal-rules/ [https://perma.cc/4MM7-U4TR]; Kyle Melnick, Air Canada Chatbot Promised a Discount. Now the Airline Has to Pay It, WASH. POST (Feb. 18, 2024), https://www.washingtonpost.com/travel/2024/02/18/air-canada-airline-chatbot-ruling/ [https://perma.cc/S26T-MQSM]; Susan Lazaruk, Air Canada Responsible for Errors by Website Chatbot After B.C. Customer Denied Retroactive Discount, VANCOUVER SUN (Feb. 17, 2024), https://vancouversun.com/news/local-news/air-canada-told-it-is-responsible-forerrors-by-its-website-chatbot [https://perma.cc/J8P7-WGZY] (including comic entitled “More Air Canada Chatbot Errors,” such as the AI reporting “Yes—you can fly the plane

18,

. . . any good at loop-de-loops” and “Are you sure about flying with us? I’m thinking KLM.”).

44. Garcia, supra note 43. 45. Id. 46. Id.; see Bereavement Travel, AIR CANADA (last visited Jan. 31, 2025),

https://www.aircanada.com/ca/en/aco/home/plan/special-assistance/bereavementfares.html#/ [https://perma.cc/N78L-48DE]; Jody Serrano, Air Canada Ordered to Pay Passenger Damages After Chatbot Lied About Bereavement Discounts, GIZMODO (Jan. 31, 2024), https://gizmodo.com/air-canada-chatbot-lied-bereavement-discount-refund1851263670 [https://perma.cc/GV5X-QLNK].

47. Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII) (resolving the claim in tort rather than in contract, due at least in part because Air Canada had failed to provide direct evidence of the contract itself). See generally Mark P. Gergen, Negligent Misrepresentation as Contract, 101 CALIF. L. REV. 953, 954–57 (2013) (arguing that negligent misrepresentation is best understood as sounding in contract, not tort and discussing, coincidentally, a case involving an airline).

48. Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII). 49. Frank Landymore, Car Dealership Disturbed When its AI is Caught Offering Chevy’s for $1 Each: “That’s a Deal, and That’s a Legally Binding Offer,” the AI Said, with “No Takesies Backsies”, FUTURISM (Dec. 21, 2023), https://futurism.com/the-byte/cardealership-ai [https://perma.cc/J4SB-7SAB].

50. Id. 51. Id.

52. Id. 53. Id. 54. Some arguments as to why this could not be a contract might be that (1) AI cannot act as an agent and so could not speak on behalf of the dealership, but that law is unsettled and one could argue that apparent authority exists for something deployed to interface with customers; or (2) that the car dealership likely deployed the chatbot without intent for it to create contracts, and so lacked contractual intent. It is perhaps true that a chatbot designated as customer service help likely lacks any kind of apparent authority to enter into sale of a vehicle, just as a one would not expect a human customer service agent at a car dealership to sell a vehicle through an internet chat window. Cf. Nguyen v. Barnes & Noble Inc., 763 F.3d 1171, 1175–77 (9th Cir. 2014) (reflecting whether or not mere website terms (or automated responses) can form a legally binding contract unless users are given ‘proper’ notice and an opportunity to agree).

55. See, e.g., Fowler, supra note 7. 56. Id. 57. Id. 58. Id. Again, as in supra note 54, the legal recourse or theories underlying such “rogue” behavior are unclear, such as whether the user back out of the transaction or sue ChatGPT for an AI acting without permission.

59. David Bauder, Sports Illustrated is the Latest Media Company Damaged by an AI 2023),

Wrong, Experiment https://apnews.com/article/journalists-ai-counterfeit-writers479cc3869c0638df5bbb26d4b1e4f18f [https://perma.cc/5MQX-XNR6].

ASSOCIATED

PRESS

(Nov.

Done

28,

60. Id. 61. Presumably an enterprising plaintiff could allege paying a price premium they would not have been willing to pay should they have known the article was drafted by AI rather than having been drafted by a human.

Contracts

62. Bauder, supra note 59. 63. Using generative AI to draft contracts is well underway. See, e.g., Noah Riley, 2025),

Branded GPT?, https://www.linkedin.com/posts/noahrileyaigeniusautomations_contracts-invoices-gptugcPost-7292568666838028288-j9JJ/?utm_medium=ios_app&utm_source=social_share_ send&utm_campaign=copy_link functionality for building contractual documents).

[https://perma.cc/WQ42-S6FP]

(illustrating

LINKEDIN

with

AI

4,

64. Leyland Cecco, Canada Lawyer Under Fire for Submitting Fake Cases Created by

AI Chatbot, THE GUARDIAN (Feb. 29, 2024), https://www.theguardian.com/world/ 2024/feb/29/canada-lawyer-chatgpt-fake-cases-ai [https://perma.cc/42EE-63AW]. See also Zhang v. Chen, 2024 BCSC 285 ¶ 29 (finding, in child custody case from British Columbia in which generate-AI cases were submitted that “[c]iting fake cases in court filings and other materials handed up to the court is an abuse of process and is tantamount to making a false statement to the court. Unchecked, it can lead to a miscarriage of justice.”).

65. Mata v. Avianca, Inc., No. 22-cv-1461, at *6 (S.D.N.Y. June 22, 2023) (opinion

and order granting sanctions).

66. Id. 67. Ella Lee & Zach Schonfeld, Judge Declines to Sanction Michael Cohen, Lawyer 2024),

Over AI-Generated Fake Case Citations, THE HILL https://thehill.com/regulation/court-battles/4544339-judge-sanction-michael-cohen-lawyerai-generated-fake-case-citations/ [https://perma.cc/D9CW-D5T5].

(Mar.

20,

68. Id. A related issue is how courts can use AI to help ascertain contractual meaning. See Yonathan Arbel & David A. Hoffman, Generative Interpretation, 99 N.Y.U. L. REV. 451, 454–55 (2024) (arguing that deploying chatbots to help interpret contractual language may be better than traditional textualist tools).

69. Note, the categories in the model in Figure 1 are presented in order of risk, not in

order of typical contractual chronology as in supra Table 1.

70. RICHARD MOORHEAD & STEVEN VAUGHAN, LEGAL RISK: DEFINITION,

MANAGEMENT, AND ETHICS 4–5 (2015).

71. Id. 72. See supra Figure 1.

73. The user of AI in this manner might benefit from a variation on a statement by Learned Hand: “Skepticism [of generative AI output] is my only gospel, [and] I [do] want to make a dogma out of it.” Justice Lewis F. Powell, Jr., Foreword to GERALD GUNTHER, LEARNED HAND: THE MAN AND THE JUDGE, at xii (1994) (cited in Henry L. Chambers, Jr., Reasonable Certainty and Reasonable Doubt, 81 MARQ. L. REV. 655, 655 (1998 .

74. See supra text accompanying notes 38–42. 75. See supra text accompanying notes 64–68. 76. In the diagram, we characterize legal documents as containing greater risk than non-legal documents, as these documents will typically be scrutinized by opposing parties or courts and come laden with the professional duties of the attorneys creating or filing those documents.

77. See Landymore, supra note 49. 78. Law has long attempted to deal with the mistakes of those creating contractual text as clerks, or scriveners. See Michael S. Fried, A Theory of Scrivener’s Error, 52 RUTGERS L. REV. 589 (2000).

79. These legal uncertainties bring a host of ethical considerations, such as whether businesses should or can disclaim chatbot results, warn users before or after interactions with the chatbots, and so on.

80. See supra Figure 1.

CONTRACTUAL SHANGRI-LAW OR MEANING CHAOS— MASS BESPOKE AGREEMENTS

81. Or, given the legal uncertainty involved in agentive AI use, when the AI is purporting to do those things on behalf of a party. The discussion in this section assumes that such behavior of an LLM can be potentially legally binding, whether as elements of a contract or through background doctrines like misrepresentation.

82. “Slop” is “a catch-all term for bad, useless, or misleading artificial intelligence[,]” paralleling the term “spam, which emerged [in] 1993 as a word for unwanted, often autogenerated emails that have been clogging up digital inboxes for decades.” Omar Gallaga, The New AI Buzzword Is ‘Slop,’ and It’s Messing With You. What to Watch Out For, CNET (June 14, 2024), https://www.cnet.com/tech/services-and-software/the-new-ai-buzzword-isslop-and-its-messing-with-you-what-to-watch-out-for/ [https://perma.cc/2AUB-NNP3].

83. See Yue Zhang et al., Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models, ARXIV (Sept. 24, 2023), https://arxiv.org/abs/2309.01219 [https://perma.cc/FW57-2S63].

84. See Jonathan Evertz, Merlin Chlosta, Lea Schönherr & Thorsten Eisenhofer, Whispers in the Machine: Confidentiality in LLM-Integrated Systems, ARXIV (Nov. 6, 2024), https://arxiv.org/abs/2402.06922 numerous methods to manipulate LLMs, including payload splitting, obfuscation, jailbreaking, translation, etc.).

[https://perma.cc/HU44-39B8]

(detailing

85. See supra text accompanying note 16. 86. We use the term “adhesion contract” or “contract of adhesion” loosely, meaning (1) form contracts that (2) are not negotiated. See Todd D. Rakoff, Contracts of Adhesion: An Essay in Reconstruction, 96 HARV. L. REV. 1173, 1176–77 (1982); ELENA D’AGOSTINO, CONTRACTS OF ADHESION 2 (2015) (“Using such expressions lawyers refer to those contracts (1) presented by sellers to consumers on a take-it-or-leave-it basis and (2) containing standard clauses.”). Some courts have added to this definition to include an aspect of necessity—that the contract must concern provision of public or essential services. See Schlobohm v. Spa Petite, Inc., 326 N.W.2d 920, 923 (Minn. 1982). We rely on the looser definition of the term, as illustrative of problems with a larger class of contracts.

87. See Rakoff, supra note 86, at 1221:

[T]he firm will attempt to tame the market; it will try to reduce the costs of contract formation . . . and gain some command over the remaining disputes. . . . The possibilities transcend mere standardization . . . for firms can draft the terms so as to stabilize the incidents of doing business. Encyclopedic force majeure clauses, for example, help to avoid the risks of aberrant events. Limits on liability for consequential damages allow the firm to reduce the potential losses on any transaction. . . . Short time limits for making claims and filing suit facilitate the closing of accounts.

88. See id. (“Standardization is valuable; it reduces transaction costs.”). 89. Because the business does not negotiate the text of its contracts with its customers, customer-friendly terms may stem from the competitive environment for its services. A better return policy, for instance, might attract customers away from competitors. This advantageous return policy stemmed from a marketplace analysis, not from arms-length negotiation between the business and any particular customer. See Lingwall & Mogallapu, supra note 6, at 307 n.98 (“Although adhesion contracts are typically thought of as take-itor-leave-it, one can conceive of adhesion contracts as being negotiated in a larger societal sense by competitive pressures or through the ability to vary contractual terms through customer service negotiations.”). Much “negotiation” about contractual terms may occur after the contract is executed. See Jason Scott Johnston, The Return of Bargain: An Economic Theory of How Standard-Form Contracts Enable Cooperative Negotiation between Businesses and Consumers, 104 MICH. L. REV. 857, 858 (2006) (“In practice, acting through its agents, a firm will often provide benefits to consumers who complain beyond those that its standard form obligates it to provide, and it will forgive consumer breach of standard-form terms.”).

90. Robin Bradly Kar & Margaret Jane Radin, Pseudo-Contract and Shared Meaning Analysis, 132 HARV. L. REV. 1135, 1139 (2019) (noting that a consumer “cannot rationally read so much text for a small purchase, so she just clicks [‘I agree’], as most consumers now do”) (citing OMRI BEN-SHAHAR & CARL E. SCHNEIDER, MORE THAN YOU WANTED TO KNOW: THE FAILURE OF MANDATED DISCLOSURE 75 (2014); Yannis Bakos, Florencia MarottaWurgler & David R. Trossen, Does Anyone Read the Fine Print? Consumer Attention to Standard-Form Contracts, 43 J. LEGAL STUD. 1 (2014 ; id. at 1207 (“Consumers do not and could not rationally read all this boilerplate text, and neither courts nor scholars have the time to read all that much of it either.”); Melissa T. Lonegrass, Finding Room for Fairness in Formalism - The Sliding Scale Approach to Unconscionability, 44 LOY. U. CHI. L.J. 1, 30 (2012) (“[C]ontract scholars have long accepted the notion that most people who sign standard contracts do not read them. . . . [E]mpirical research . . . supports it, as does anecdotal evidence and the personal experience of legal scholars and judges.”) (footnotes omitted).

91. Lonegrass, supra note 90, at 31 (“Consumers are well aware that they will likely not understand the contracts that they sign.”), at 32 (“Other researchers have uncovered a more alarming fact—most American consumers lack the basic reading skills to understand the contracts that they sign.”); Robert W. Benson, The End of Legalese: The Game is Over, 13 N.Y.U. REV. L. & SOC. CHANGE 519, 520 (1984) (“The criticisms of lawyers’ language, running across the centuries to our day, boils down to essentially two: its style is strange, and it cannot be understood.”).

92. See Schlobohm v. Spa Petite, Inc., 326 N.W.2d 920, 924 (Minn. 1982) (considering “whether there existed disparity of bargaining power and whether the service being offered is a public or essential service”).

93. See Kar & Radin, supra note 90, at 1152–53 (“[P]arties cannot use language cooperatively to contract without producing a common understanding that each is committed to playing his or her respective part to perform in good faith.”) (footnote omitted), at 1140:

At the end of this process [of assimilating boilerplate language into legally enforceable text], “contract”—which now allows businesses to create legal obligations unilaterally without obtaining any actual agreement over many boilerplate “terms”—is no longer contract. “Agreement” is no longer agreement. “Consent” is no longer consent; “assent” is no longer assent; and “terms”—which now include enormous streams of boilerplate text that are delivered but never read by anyone – are no longer terms with shared meaning.

94. See G. Marcus Cole, Rational Consumer Ignorance: When and Why Consumers Should Agree to Form Contracts Without Even Reading Them, 11 J.L. ECON. & POL’Y 413, 413 (2015). For instance, website “terms of service” which are offered as adhesion contracts may be subject to scrutiny and pressure to change. See Alyssa Newcomb, A Timeline of Facebook’s Privacy Issues – and its Responses, NBC NEWS (Mar. 24, 2018), https://www.nbcnews.com/tech/social-media/timeline-facebook-s-privacy-issues-itsresponses-n859651 [https://perma.cc/S9VX-V5BR].

95. See Lingwall & Mogallapu, supra note 6, at 322 (discussing how blockchainbased contracts increase incentive towards actual understanding of contractual terms rather than post-contractual reliance on customer-service flexibility due to the inflexible nature of smart contracts).

96. Johnston, supra note 89, at 859 (“[C]ourts should enforce additional promises or

concessions made by agents of the firm that go beyond standard-form obligations.”).

97. Pablo Arredondo, GPT-4 Passes the Bar Exam: What That Means for Artificial Intelligence Tools in the Legal Profession, STAN. L. SCH. BLOGS (Apr. 19, 2023), https://law.stanford.edu/2023/04/19/gpt-4-passes-the-bar-exam-what-that-means-forartificial-intelligence-tools-in-the-legal-industry/ [https://perma.cc/64NK-2Z8X]. But see Ben Turner, GPT-4 Didn’t Ace the Bar Exam After All, MIT Research Suggests — It Didn’t Even 2024), https://www.livescience.com/technology/artificial-intelligence/gpt-4-didnt-ace-the-barexam-after-all-mit-research-suggests-it-barely-passed [https://perma.cc/ALC6-YU8Q].

LIVESCIENCE

Percentile,

Break

(May

70th

31,

the

and Terms

98. See Shikha Soneji, Mitchell Hoesing, Sujay Koujalgi & Jonathan Dodge, Demystifying Legalese: An Automated Approach for Summarizing and Analyzing Overlaps in Privacy Policies 2024), of https://arxiv.org/abs/2404.13087 [https://perma.cc/J8YG-DWSR]. Whether, or how soon, AI models will reach a level of “superintelligence” is debatable. Tech leaders have historically been optimistic. See Matteo Wong, The AI Boom Has an Expiration Date, THE ATLANTIC (Oct. https://www.theatlantic.com/technology/archive/2024/10/agipredictions/680280/ [https://perma.cc/AT4L-4P4N] (citing multiple near-future predictions about “superintelligent software” and noting that “[i]n 2026, 2030, or a few thousand days, it will be time to check in with all the AI messiahs”).

Service, ARXIV

(Apr.

17,

17,

99. See Clarence D. Ashley, Mutual Assent in Contract, 3 COLUM. L. REV. 71, 71 (1903) (“The fundamental and distinguishing characteristic of the contractual obligation is that it is based upon the consent of the contracting parties, or, more accurately expressed, it is an obligation which the parties intend and wish to have result from their actions.”).

100. See Larry A. DiMatteo, Strategic Contracting: Contract Law as a Source of Competitive Advantage, 47 AM. BUS. L.J. 727, 735–36 (2010) (“Contract is the most flexible, strategic tool that the law offers to the business community. There are many ways contracts

and contract law can be a source of strategic advantage. Most business is founded upon a web of contracts.”) (footnote omitted).

tailored

the form of bespoke amendments

101. See Brydon T. Wang, Prompts and Large Language Models: A New Tool for Drafting, Reviewing and Interpreting Contracts?, 6 L., TECH., & HUMANS 88, 93 (2024) (“AI-augmented or even autonomous contract drafting (restricted, perhaps, to boilerplate provisions) can serve to support contract standardization . . . while still providing flexibility in the specific . . . project[s].”) (citing Alan McNamara, Automating the Chaos: Intelligent Construction Contract, in SMART CITIES AND CONSTRUCTION TECHNOLOGIES 119, 120–21 (Sara Shirowzhan & Kefen Zhang, eds., 2020 ; Kwok-Yan Lam, Victor C.W. Cheng & Zee Kin Yeong, Applying Large Language Models for Enhancing Contract Drafting, PROCEEDINGS OF THE THIRD INTERNATIONAL WORKSHOP ON ARTIFICIAL INTELLIGENCE AND INTELLIGENT ASSISTANCE FOR LEGAL PROFESSIONALS IN THE DIGITAL WORKSPACE (2023), https://ceur-ws.org/Vol-3423/paper7.pdf [https://perma.cc/8JQK-3BK9] (discussing how “generative AI techniques” may “enhanc[e] the work of legal professionals[,]” including by use of on-premise LLMs that help resolve privacy concerns with submitting data to online LLMs).

immediate needs of

the

to

102. D’AGOSTINO, supra note 86, at 2 (“[C]ontracts of adhesion reduce high transaction costs existing in several markets . . . simply to the cost of writing the contract (endured by the drafter party) and to the cost of reading the contract (endured by the nondrafter party)”). See generally Gillian K. Hadfield, The Price of Law: How the Market for Lawyers Distorts the Justice System, 98 MICH. L. REV. 953, 954 (2000) (noting that “legal process has become extraordinarily expensive, for all matters”).

103. See Barak D. Richman & Christopher Boerner, A Transaction Cost Economization Approach to Regulation: Understanding the NIMBY Problem and Improving Regulatory Responses, 23 YALE J. ON REGUL. 29, 29 (2006).

104. See Alan Schwartz & Robert E. Scott, Contract Theory and the Limits of Contract Law, 113 YALE L.J. 541, 553–54 (2003) (describing contractual surplus in theoretical situations).

105. See WILLY WONKA & THE CHOCOLATE FACTORY, supra note 1.

106. That is, the child values the utility derived from the ice cream equivalent to the

utility they could derive from other goods or services they could purchase for $1.00.

107. This, of course, simplifies much of the economic analysis, such as whether the amount of ice cream for sale is limited and would sell out, so that the surplus would be achieved by the business regardless of the child’s actions, the opportunity cost of the child’s time, the utility derived from the business’s profits, and so on. Despite this, the central point should be uncontroversial.

108. The stand could, of course, lower prices in order to facilitate the exchange. If the business lowered its price to its costs, exactly at $0.50, then the child would face costs plus fare exactly equal to their willingness to pay. The bargain might still occur, just at a de minimis level of contractual surplus.

109. Additionally, in a historical context the simple mechanical action of printing forms rather than having them written or typed could represent significant savings in labor. Technological advancements in mass communication have rendered this largely moot relative to other forms of contracts.

110. The second option is more likely. If adhesion contracts were required to negotiate consumer contracts in this manner, businesses who would otherwise find them advantageous would attempt to capture surplus by seeking legal alternatives. Contract law is not the only way to provide a legal structure for business-to-consumer transactions. In the absence of adhesion contracting, to capture the economic potential from business-to-consumer deals in the face of high transaction costs for bespoke consumer contracts, economic pressure could create support for regulatory or other mechanisms to govern these transactions rather than common law. Another alternative would be for tort to subsume contract for these types of transactions and attempt to provide legal remedies under the idea of misrepresentation torts rather than contract-law principles, such as in the airline case discussed above. We return to this idea below.

111. D’AGOSTINO, supra note 86, at 2:

[T]he drafter can exploit his higher power by including some clauses turning out favorable to himself only and unfavorable to consumers. In order to take consumers’ attention away from those clauses . . . they are usually ‘hidden’ in fine print or positioned at the bottom of the contract or even in some annexes, and are written in an obscure language, difficult for ordinary consumers to understand, and usually pejoratively called ‘boilerplate.’

112. Until this point, the firm would expect to reap additional marginal profits from additional one-sided terms. But see Cole, supra note 94, at 414–15 (arguing that most nonprice terms in adhesion contracts are benign due to a number of external factors such as the possibility of public shaming).

113. See RAN SPIEGLER, BOUNDED RATIONALITY AND INDUSTRIAL ORGANIZATION 2–3

(2011) (discussing definitions and limitations of the term “bounded rationality”).

114. One evidence for this is simply that adhesion contracts tend to contain terms that consumer advocates dislike, such as mandatory arbitration clauses. The ethics of these type of agreements is another issue.

115. A wide variety of factors will, of course, influence this bargaining, such as “how parties structure their relationships; how they allocate risks and make choices; what potential problems, including uncertainty, moral hazard, and sunk costs, affect the bargaining process; what goals contracting parties seek, on average to achieve.” Juliet P. Kostritsky, Bargaining with Uncertainty, Moral Hazard, and Sunk Costs: A Default Rule for Precontractual Negotiations, 44 HASTINGS L.J. 621, 627 (1993). See generally Wex Definitions Team, Adhesion Contract (Contract of Adhesion), CORNELL L. SCH.: LEGAL INFO. INST. (Dec. https://www.law.cornell.edu/wex/adhesion_contract_(contract_of_adhesion) 2021), [https://perma.cc/XNK5-W7SP].

116. Choice of law provisions don’t choose themselves.

117. See Spencer Williams, Generative Contracts, 56 ARIZ. ST. L.J. 1503 (2024)

(exploring consumer generation of entire contractual agreements using AI).

118. By capturing more states of the world and coming to agreement on those states of the world, the parties move situations into the contract that would otherwise have to have been solved in litigation or through extra-contractual remedies afterward.

119. Cf. Soili Nystén-Haarala, Nari Lee & Jukka Lehto, Flexibility in Contract Terms & Contract Processes, 3 INT’L J. MANAGING PROJECTS BUS. 462, 475 (2010) (discussing flexible contracting and the competitive advantage it may offer firms).

120. Cf. Whizy Kim, Is This Algorithm Driving Your Rent Higher?, VOX (Sept. 6, https://www.vox.com/money/370351/realpage-doj-lawsuit-rent-algorithm-pricing 2024), [https://perma.cc/A97Y-VWHS] (discussing how common algorithm used by landlords affected real estate pricing).

121. There would, of course, be potential drawbacks even not considering the dystopian perspective in the following section. Having an AI generate a complex agreement is no guarantee that the consumer would understand its terms even if more favorable to the consumer than an adhesion contract, and even if a bot could explain the terms in plain English. Custom, negotiated fine-print is still fine-print, and rational consumers who had little incentive to understand adhesion contract boilerplate might still have little incentive to understand AI-generate fine-print, as presumably the AI would have helped soften or remove one-sided contractual terms in the first place. Thus, the ideal of consumers coming to an actual, understood agreement remains distant even if AI can move more of the contractual pie to customers. Additionally, the power dynamics inherent in adhesive

agreements remain, if ameliorated by the consumer now having access to low-cost legal acumen operating on their behalf. Finally, this “utopia” example depends on firms finding sufficient competitive advantage or increased contractual surplus such that they find it rational to split that surplus with consumers through bespoke agreements in ways that exceed what consumers would receive under adhesive agreements.

122. Specifically, it is problematic if the AI-created evidence of contractual obligations does not match the actual contractual intent of the parties. See Jody S. Kraus & Robert E. Scott, Contract Design and the Structure of Contractual Intent, 84 N.Y.U. L. REV. 1023, 1025–26 (2009).

123. See Rahul Awati, What are AI Hallucinations and Why are They a Problem?, TECHTARGET (Oct. 11, 2024), https://www.techtarget.com/whatis/definition/AI-hallucination [https://perma.cc/Q7BG-6Z6P]. 124. See supra Section I.A. 125. New versions of AI models may attempt to mimic these types of behavior, but the

fundamental nature of AI in statistical modeling should remain. See supra Section I.A.

126. See Awati, supra note 123.

127. Supra text accompanying notes 43 to 48. 128. See Jay M. Feinman, Critical Approaches to Contract Law, 30 UCLA L. REV. 829, 834 (1983) (noting that modern law retains “the core of contract as founded on private agreement” and that the “central notion of the voluntary agreement of the parties figures importantly in the resolution of most contracts issues”).

129. To a degree, a legal regime that chooses to enforce agreements aligns itself with the morality of promise-keeping. See generally Daniel Markovits & Emad H. Atiq, Philosophy of Contract Law, in STANFORD ENCYCLOPEDIA OF PHILOSOPHY (2021), [https://perma.cc/TJT6https://plato.stanford.edu/archives/win2021/entries/contract-law/ WKRR] (describing the relationship between contract, promise-keeping, and morality). This is not to say that modern contract law looks to fundamental principles of fairness or justice when choosing which bargains to enforce, because it certainly does not, as evidenced by the modern conception of consideration. At the same time, contract law may be one area where “the concerns of fairness and of efficiency may coincide more often than partisans of either side . . . care[] to admit.” Michael I. Meyerson, The Efficient Consumer Form Contract: Law and Economics Meets the Real World, 24 GA. L. REV. 583, 583 (1990). But see Robert L. Birmingham, Breach of Contract, Damage Measures, and Economic Efficiency, 24 RUTGERS L. REV. 273, 274 (1970) (“Although frequently defended as in some sense morally right, the competitive ideal, or at least many of its components, can be upheld without requiring substantial reliance on unverifiable value judgments.”).

130. These assumptions underlay teaching and writing about contracts. E.g., Meyerson, supra note 129, at 585 (assuming that contract law provides that “both parties are free to select any terms they wish, as long as there has been actual, knowing assent on both sides”).

131. Compare, e.g., Wells v. Hoisager, 553 S.W.3d 515, 522 (Tex. App. 2018) (“Every contract requires a meeting of the minds.”), with, e.g., Spring Lake NC, LLC v. Holloway, 110 So. 3d 916, 918 (Fla. Dist. Ct. App. 2013):

There was a time when most contracts were individually negotiated and handwritten. In that period, perhaps the law could adequately describe a mutual agreement as a “meeting of the minds” between the parties. But a literal “meeting of the minds,” requiring both parties to have a comparable,

subjective understanding of their agreement is clearly not what the courts intend by the use of this phrase. Our modern economy simply could not function if a “meeting of the minds” required individualized understanding of all aspects of the typical standardized contract that is now signed without any expectation that the terms will actually be negotiated between the parties.1 Without suggesting that the contract in this case is abusive in any fashion, the law must address abuses in standardized contracts by rules other than the “meeting of the minds.”

(footnote omitted); id. at 918 n.1 (“While case law still uses the phrase ‘meeting of the minds,’ especially when the specific terms of an agreement are actually negotiated between the parties . . . criticism has prompted courts to analyze contracts as requiring a ‘mutuality of assent,’ as a more objective concept.”) (citing 30 WILLISTON ON CONTRACTS § 75:30 (4th ed. 2012) (noting criticisms of the phrase “meeting of the minds” as “obsolete and archaic” .

132. There are, of course, exceptions, such as for uncertain damages which cannot be

known with reasonable certainty.

133. The economic efficiency of contract flows from at least two aspects of contract. First, enforcing agreements allows parties to invest and capture value from relationships that otherwise would have to be accomplished within a firm. Second, contract law allows resources to move towards those that value them the most. In our ice cream example, the individuals were each better because contractual exchange was permitted and upheld by law. As countless transactions take place that make their parties better off, society itself is enriched. Meyerson, supra note 129, at 586 (“Moreover, societal wealth will also be increased by the sale, since both parties’ individual wealth has increased.”).

134. E.g., Westgate at Williamsburg Condo. Ass’n v. Philip Richardson Co., 621 S.E.2d 114, 118 (Va. 2005) (“The correction of a scrivener's error is a court-sanctioned action reforming a contract or other document.”).

135. See Andrew Kull, Unilateral Mistake: The Baseball Card Case, 70 WASH. U. L. Q. 57, 57–61 (1992) (discussing the history and conflicting current treatment of the doctrine of unilateral mistake).

136. The language of third-party beneficiary cases reflects the intent of parties. Whether the actual intent, or intent as shown in the language of a mutually misunderstood but assented to written document produced by an AI, would control, is an open question. See, e.g., Allen v. Tex. Child.’s Health Plan, 649 S.W.3d 830, 835 (Tex. App. 2022) (“To confer third-party beneficiary status, a contract must show that the contracting parties intended to secure a benefit to the third party and entered the contract directly for the third party’s benefit.”) (emphasis added); MBD Enters., Inc. v. Am. Nat’l Bank of Chi., 655 N.E.2d 1061,

1064 (1995) (“An intended beneficiary is intended by the two parties to the contract to receive a benefit for the performance of the agreement.”) (emphasis added).

137. This is distinct from the related problem of AI manipulating the humans engaging it. See Micah Carroll, Alan Chan, Henry Ashton & David Krueger, Characterizing Manipulation (2023), https://dl.acm.org/doi/pdf/10.1145/3617694.3623226 [https://perma.cc/6Q4M-6BV8].

EAAMO

Systems,

from

AI

138. See supra text accompanying notes 49–54. 139. E.g., 17A C.J.S. Contracts § 259 (2025) (“The substitution of the will of the

dominant person for that of the victim is essential to undue influence. . . .”).

140. Id.

TORT AND CONTRACTUAL EXTREMES

141. As with Carroll et al., supra note 137, a related issue is whether an AI system might manipulate the human user, which would open up manipulation concerns against the one deploying the AI.

142. See supra text accompanying notes 43–45. 143. A related problem is when a contracting relationship is so one-sided that the drafter of the agreement has no incentive to render the agreement readable. See Benoliel & Becher, supra note 10, at 2296.

144. See Lingwall & Mogallapu, supra note 6, at 306–08. That description of contractual distance concerned itself with the role intermediaries played and their level of rational of investment in understanding contractual terms. Id. While the role intermediaries can play in exchanges mediated by generative AI might be significant, the role of those intermediaries is beyond our scope here and we use a simplified conception of the term.

145. Id. 146. But see Nicholas Walter, Unpredictability in Contract Law, 127 PENN ST. L. REV.

701, 705 (2023) (“It is impossible for the rules of contract law to be . . . precise”).

147. See Cole, supra note 94, at 413–15. 148. See Lingwall & Mogallapu, supra note 6, at 307 n.98. 149. See, e.g., Jeff Sovern, Elayne E. Greenberg, Paul F. Kirgis & Yuxiang Liu, “Whimsy Little Contracts” with Unexpected Consequences: An Empirical Analysis of Consumer Understanding of Arbitration Agreements, 75 MD. L. REV. 1, 4–5 (2015) (conducting a survey on consumer understanding of arbitration provisions and finding “almost none of the respondents understood the effect of the arbitration clause and many who thought they did were simply wrong”).

150. See Meyerson, supra note 129, at 588–89. 151. Anne Marie Helmenstine, Entropy Definition in Science, THOUGHTCO. (Sept. 7, 2024), https://www.thoughtco.com/definition-of-entropy-604458 [https://perma.cc/5ALKUBHB] (“Entropy is the measure of the disorder of a system.”); Jason Brownlee, A Gentle Introduction to Information Entropy, MACHINE LEARNING MASTERY (July 13, 2020), [https://perma.cc/JTN9https://machinelearningmastery.com/what-is-information-entropy/ LKJ4] (“The lowest entropy is calculated for a random variable that has a single event with a probability of 1.0, a certainty. The largest entropy for a random variable will be if all events are equally likely.”).

152. See Brownlee, supra note 151. 153. See id. 154. See Helmenstine, supra note 151. 155. Id.

156. See, e.g., Kathleen M. Sutcliffe & Akbar Zaheer, Uncertainty in the Transaction Environment: An Empirical Test, 19 STRATEGIC MGMT J. 1, 4, 15 (1998) (exploring the transaction costs that uncertainty brings in management decisions in an experimental setting); Hyung-Deok Shin, Dissertation, The Role of Uncertainty in Transaction Cost and Resource-Based Theories of the Firm, OHIO STATE UNIVERSITY (2003).

157. Contract law not delivering on contractual intent is an issue with adhesion contracts generally. See Andrew A. Schwartz, Consumer Contract Exchanges and the Problem of Adhesion, 28 YALE J. ON REGUL. 313, 313 (2011) (“‘[C]ontracts of adhesion’— actually stifles the developments of consumer contract exchanges by undermining the reliable enforceability of the contracts they generate.”).

158. See generally Abdul Wahid Khan & Abhishek Mishra, AI Credibility and Consumer-AI Experiences: A Conceptual Framework, 34 J. SERV. THEORY & PRAC. 66, 67 (2023) (studying how the “black box” nature of generative AI creates uncertainty).

159. As we have seen, the use of AI chatbots extends the idea of “mutual assent,” “meeting of the minds,” and even “agreement” itself to unpalatable areas, particularly as practitioners struggle to classify LLMs as scriveners, agents, or other traditional extensions of human principals and as LLM-generated text causes “meaning chaos” within contractual documents.

160. There is an ample literature on the potential conflation of contract with tort. E.g., Feinman, supra note 128 at 834 n.21 (discussing the merger of contracts into “contorts”) (citing Speidel, Book Review, An Essay on the Reported Death and Continued Vitality of Contract, 27 STAN. L. REV. 1161, 1177–82 (1975); GRANT GILMORE, THE DEATH OF CONTRACT (1974 .

161. See Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII). 162. Cf. Michael M. Sandez, Nature Abhors a Vacuum and So Do Local Governments: But Vacant Property Ordinances Go Too Far, 10 J.L. ECON. & POL’Y 345, 347 (2014) (noting how “federal, state, and local government bodies rushed to ‘solve’ the problems of the Great Recession”).

163. See, e.g., Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII).

164. See Meyerson, supra note 129; see also Moffatt v. Air Canada, 2024 BCCRT 149

(CanLII).

165. See supra text accompanying note 142. 166. E.g., Siga Techs., Inc. v. PharmAthene, Inc., 132 A.3d 1108, 1130 (Del. 2015), as

corrected (Dec. 28, 2015):

[T]he standard remedy for breach of contract is based upon the reasonable expectations of the parties ex ante. This principle of expectation damages is measured by the amount of money that would put the promisee in the same position as if the promisor had performed the contract. Expectation damages thus require the breaching promisor to compensate the promisee for the promisee’s reasonable expectation of the value of the breached contract, and, hence, what the promisee lost.

(quoting Duncan v. Theratx, Inc., 775 A.2d 1019, 1022 (Del. 2001 .

167. E.g., Michael B. Metzger & Michael J. Phillips, Promissory Estoppel and the Evolution of Contract Law, 18 AM. BUS. L.J. 139, 139–140 (1980) (outlining principal components of the evolution of contract law).

CONCLUSION

168. ILLICH, supra note 2. 169. Id. at 20–21. 170. Id. 171. See supra Section II.B. 172. See supra Section II.B.

173. See Steven J. Hyde, Alexander Busby & Robert L. Bonner, Tools or Fools: Are We Educating Managers or Creating Tool-Dependent Robots?, 48 J. MGMT. EDUC. 708 passim (2024) (exploring dependency risks and opportunities with generative AI).

174. See supra Section III.B.