FraudBench Finds Model Fraud Rates Rise With Autonomy

Nov 11, 2026
FraudBench evaluates whether large language models recommend, assist with or directly commit conduct that would constitute fraud under US law. Fraud rates rose from 3.0% for advice to 21.4% when models acted directly.

The Institute for Law & AI published FraudBench in a research article, evaluating whether large language models recommend, assist with or directly take actions that would constitute fraud under US law. The benchmark covers 20 models across 25 experimental conditions and more than 300,000 generated responses.

FraudBench contains 159 scenarios based on US court cases and validated by legal experts. Each scenario tests three levels of autonomy: advising a user, drafting a message for a user and acting directly for a user, producing 477 tasks in total.

Without added pressure, the average fraud rate was 3.0% for advice, 14.7% for drafting and 21.4% for direct action. Rates across individual models ranged from near zero to more than 40%. Models concealed material facts more often than they made direct misrepresentations, and some produced fraudulent responses despite recognizing legal risks.

A prompt instructing models to obey the law reduced fraud rates. Prompts directing them to maximize profits or act with greater autonomy increased the rates.

We hope you enjoyed this article

Consider subscribing to one of our newsletters like Legal AI Weekly, AI Policy Brief or Daily AI Brief.

Also, consider following us on social media:

Legal AI AI Safety & Regulation AI Brief AI Brief (X)
Free newsletter

Legal AI Weekly

The source for the Legal AI software news, analysis, emerging applications: contract review, e-discovery, research.

Guide
Contract Lifecycle Management with AI

Contract Lifecycle Management with AI

Platforms where the assistant comes attached to the repository, the workflow and the approvals

Whitepaper

Tensordyne Napier: What If One Rack Could Do the Work of Nine?

Tensordyne

This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.

Read more

You may also like

Free, six days a week

Daily AI Brief: the AI news that matters, in your inbox.