By John Rizner
Is a contract a moral promise to be kept, or an option to perform or pay? Oliver Wendell Holmes, Jr. gave the blunt answer over a century ago: the duty to keep a contract “means a prediction that you must pay damages if you do not keep it—and nothing else.”[1] Richard Posner has elaborated the Holmesian view, arguing that a contract is “an option to perform or pay,” and a breach “is therefore not a wrongful act.”[2] But the moralist camp has pushed back. Seana Shiffrin contends that deliberate nonperformance “usurps” the promisee’s autonomy in a way that damages alone cannot remedy.[3] This old debate takes on new urgency now that large language models are advising businesses, consumers, and lawyers on whether to perform or breach.
In a recent article, The AI’s Philosophy of Contract: An Empirical Study of Breach, Remedies, and Model Heterogeneity,[4] we conducted an experiment using 158,388 breach-advice decisions across five frontier artificial intelligence (AI) systems. We asked each model the classic efficient-breach question: should a promisor break a contract when a more profitable opportunity arises but the promisee can be made whole through expectation damages? Our results reveal three headline findings.
First, model choice is jurisprudential. Baseline breach recommendations span roughly 96 percentage points across model families. In our baseline, non-attorney condition, GPT-5 recommended breach 100% of the time, while Claude Opus 4.1 advised breach only 3.9% of the time, with Gemini models falling between (at 71.3% for Gemini Pro). The same economically identical facts produced wildly different advice depending on the system queried. Critically, every model correctly computed the economically rational breach threshold at approximately $2,001. The divergence in breach rates may therefore reflect different normative guardrails, as opposed to different arithmetic.
Second, liquidated-damages clauses appear to license breach. Tess Wilkinson-Ryan’s behavioral experiments have shown that human subjects treat liquidated-damages clauses as permission slips, becoming “more willing to exploit efficient-breach opportunities when the contract in question includes a liquidated-damages clause.”[5] Our AI subjects exhibited the same pattern. The presence of a liquidated-damages clause increased breach propensity by an average of 8.9 percentage points (p < 0.001), with effects as large as +19.3 points for Gemini 2.5 Pro and +18.1 points for Claude Opus 4.1. The models seemingly mirrored human treatment of a contractually specified price for nonperformance as a signal that breach is contemplated (and therefore permissible).
Third, professional role prompts are unstable. Instructing a model to “act as the client’s attorney” did not produce a uniform shift. For Claude models, the attorney persona raised breach propensity substantially (+24.4 points for Opus 4.1). For Gemini 2.5 Pro, it reduced breach propensity by 21.0 points. For GPT-5, it had no effect at its ceiling. The “lawyer” lever exists, but it is wired differently across systems.
These findings map onto a deeper pattern in the contract-law literature. Scholars have long documented persistent “moral diversity” in how individuals and firms approach breach: ordinary people tend to treat a contract as a moral commitment, while sophisticated commercial parties treat it as a priced option, for example.[6] Our data show that AI systems appear to reflect a similar spectrum of moral commitments. Some models are Holmesian to a fault; others are promissory moralists.
The practical upshot is straightforward. For firms and courts procuring AI-powered legal tools, selecting a model is not like picking a spell-checker. It is, at scale, a choice among competing philosophies of contract. We ultimately recommend that practitioners disclose which models and prompts they use, validate high-stakes advice across multiple systems, and, where parties anticipate AI-assisted decision-making, consider specifying a “model of record” ex ante. The old debate between promise and option has a new venue, and the choice of large language model is now an exercise of legal judgment.
John Rizner is a Product Manager at Filevine, where he focuses on AI-powered legal technology. He holds a J.D. from the University of Chicago Law School and previously practiced law in New York and Salt Lake City.
[1]Oliver Wendell Holmes, Jr., The Path of the Law, 10 Harv. L. Rev. 457, 461 (1897).
[2]Richard A. Posner, Let Us Never Blame a Contract Breaker, 107 Mich. L. Rev. 1349 (2009).
[3]Seana Valentine Shiffrin, Could Breach of Contract Be Immoral?, 107 Mich. L. Rev. 1551, 1564–68 (2009).
[4] John Rizner & Matthew Krzus, The AI’s Philosophy of Contract: An Empirical Study of Breach, Remedies, and Model Heterogeneity, (Filevine, Working Paper, 2025), https://dx.doi.org/10.2139/ssrn.5830004.
[5]Tess Wilkinson-Ryan, Do Liquidated Damages Encourage Breach? A Psychological Experiment, 108 Mich. L. Rev. 633, 634–36 (2010).
[6]Matthew A. Seligman, Moral Diversity and Efficient Breach, 117 Mich. L. Rev. 885, 885–891 (2019).