Researchers at the University of Wollongong found that generative AI models have significantly improved in legal reasoning, raising concerns about academic integrity in unsupervised assessments.

A new study from the University of Wollongong tested nine AI models across two compulsory law subjects, finding a significant improvement in performance compared to 2023. The AI tools, which had internet access but no subject-specific materials, outperformed the majority of students in both criminal law and tort law exams. Researchers noted that while AI capabilities have advanced, the models still exhibit inconsistent performance and occasional errors, prompting calls for new university assessment strategies.

Significant Gains in Legal Reasoning

In the 2023 criminal law exam, AI models averaged 52.5% and performed at the 22nd percentile. In the new study, AI papers in criminal law averaged 76.3%, outperforming 82.5% of students. For tort law, the models averaged 66%, outperforming 61% of students. Seven of the 18 AI-generated papers ranked at or above the 90th percentile of student results.

Inconsistencies and Hallucinations

Despite the improved grades, the study found that AI models remain unreliable legal experts. Performance varied between subjects, with some models showing strong analysis alongside poor citations or fabricated authorities. Researchers observed that while some models had lower rates of hallucination, they also cited fewer sources overall.

Proposed Assessment Changes

The researchers suggest that universities move away from reliance on unsupervised assessments. They propose a three-pronged approach: maintaining AI-free, monitored in-person exams; requiring students to collaborate with AI while grading the process rather than just the final product; and using a 'relay' model where students alternate between AI-assisted drafting and independent, supervised critique of that work.