Claude AI Outscores Students on Penn Law Exams

An experiment at Penn Carey Law found an AI model performed strongly on final exams, prompting schools to reconsider how they test students.

Updated on Nov. 16, 2026 in Artificial Intelligence

Claude AI Outscores Students on Penn Law Exams

Live Poll

Do you trust that degrees still reliably show what graduates actually know?

During the spring 2026 semester, Penn Carey Law administered 11 exams to students, using Anthropic's Claude Opus 4.7 with AI-generated submissions mixed in with regular student work. The AI passed all 10 graded final exams, finishing each in eight minutes, and outscored all 84 students on the legislation exam.

Why it matters

The results raise questions about whether traditional take-home and written assessments can still reliably measure student learning. Law schools are now weighing whether to restrict AI use or leave it to individual professors.

The experiment used 11 exams and 92-student and 84-student classes as its sample, with the AI scoring in the top 10% of the class on seven exams. The sample is limited to one law school, so broader applicability across institutions remains unknown.

The players

Penn Carey Law

The law school of the University of Pennsylvania, located in Philadelphia, where the AI exam experiment was conducted.

Anthropic

The San Francisco-based artificial intelligence company that developed the Claude model used in the experiment.

The details

Penn Carey Law tested Claude both with and without course materials, slipping its submissions into the stack of regular student work for grading. Professors sometimes had trouble spotting the AI-generated submissions, and one called an AI submission the "best of the lot" in a 92-student constitutional law class.

Timeline

  1. Spring 2026: Penn Carey Law administers 11 exams to students during the semester.

  2. November 16, 2026: Results of the experiment are published.

The Tech Race

The Penn experiment tests a specific frontier model, Claude Opus 4.7, on graded legal exams rather than standard benchmarks. It places AI capability directly against the assessment systems that credential future lawyers.

Students and professors in courses with written or multiple-choice exams may see changing rules on AI use and more in-person assessment. Those evaluating student work may need new methods to detect AI-generated submissions.

The takeaway

Assessments that rely on take-home or unproctored work may need to be redesigned if AI can match top student performance. Schools weighing AI policies should plan for both restriction and detection challenges.

Further reading

For related coverage, see Artificial Intelligence developments shaping how institutions evaluate work.

Source note: This article includes information reported by Axios.

Live Poll

Do you trust that degrees still reliably show what graduates actually know?