When law professors were asked to evaluate contract law answers without knowing who wrote them, they picked the AI-generated responses roughly three out of every four times. The humans didn’t just lose. They lost convincingly.
A Stanford Law School study led by Professor Julian Nyarko, director of the university’s Legal Innovation through Frontier Technology Lab, pitted AI models against 16 law professors from 14 US law schools across 40 anonymized contract law questions. The result: AI responses won approximately 75% of the nearly 3,000 blind matchups. The researchers themselves expected the opposite outcome.
The numbers paint a lopsided picture
The study, published in early June 2026, tested AI models including Gemini 2.5 Pro and NotebookLM against human-written answers from experienced legal academics. AI models demonstrated win rates between 75.33% and 75.92% against their human counterparts, a remarkably narrow spread that suggests this wasn’t a fluke of any single model.
Here’s the part that should make legal professionals sit up a bit straighter. Only 3.53% of AI-generated answers were flagged as potentially harmful or misleading. For the professor-written responses, that figure was 12.06%. In English: the AI wasn’t just more persuasive, it was roughly three times less likely to produce something a fellow professor would consider dangerous advice.
