Law school professors grading final exams this spring had something new to weigh: whether their human students should be ranked on the same curve as a machine. The University of Maryland Francis King Carey School of Law ran the experiment twice now. The results from the second round, published this week, suggest the policy question may be less urgent than many feared.
Last year, researchers fed OpenAI’s o3 reasoning model a set of final exams. It scored an A+ down to a B, graded alongside real law students.
That was the headline. The sequel, run with OpenAI’s GPT-5.5, landed with a thud. The newer model did not clearly beat the old one.
In some areas it did worse. “This apparent plateau, despite a better model and using extra-high inference-time compute, may reflect trends observed in other legal benchmarks for AI and makes theoretical sense,” the authors wrote.
That line matters for anyone setting policy around AI in legal education. The assumption has been that each new model would push performance higher. That assumption is now in doubt.
If the technology has hit a ceiling on law school exams, the rules for how to use it — or whether to ban it — look different. The American Bar Association has not issued formal guidance on AI for final exams. Individual schools are making their own calls.
Some have prohibited AI entirely during tests. Others allow it for certain open-book sections.
The Maryland study gives both camps something to work with. If the AI is not improving, the case for a blanket ban weakens. But the case for letting students use it freely also weakens, because the tool may not be reliable enough to trust.
Consider the grading curve. Law schools typically rank students against each other.
If an AI can score an A+, that changes the curve for everyone else. A student who would have ranked in the top 10% might drop to the middle. The Maryland researchers graded the AI on the same curve as the human students.
That decision — a policy choice by the study team — has real implications. If law schools adopt that approach, human grades shift. But the plateau complicates the picture.
A model that scores an A+ one year and then plateaus the next is not a rising threat. It is a known variable.
Professors can plan around it. They can decide to exclude AI from the curve, or they can keep it in and treat the scores as a fixed baseline. The researchers plan to keep running the experiment.
Future semesters will show whether the plateau holds or whether the next model breaks through. That means law schools may not need to rush a permanent policy.
They can watch the data. For patients, for families, for anyone whose legal case might one day involve AI-generated work, the plateau is not bad news. A tool that stalls is a tool that can be understood.
It is the unpredictable leaps that make regulation hard. The TaxProf Blog covered the findings. Legal educators are talking about what comes next.
The conversation is less panicked than it was a year ago. That may be the most important policy change of all: the space to think, rather than react.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Maryland law school study finds AI exam scores may not surpass human students”, “description”: “Law school professors grading final exams this spring had something new to weigh: whether their human students should be ranked on the same curve as…”, “datePublished”: “2026-06-29T13:16:00+00:00T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “James Roberto”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.2bme.xyz/”}}
