Legal AI tools that predict case outcomes may be cheating on their own tests.
Researchers studying UK Employment Tribunal decisions found that models trained on post-hoc judicial text — the written rulings issued after a verdict — were effectively reading the answer before making a prediction. Using a corpus of 33,158 individual claims, the team tested everything from simple TF-IDF classifiers to large language models. Headline accuracy figures looked strong. But when the researchers stratified results by how much outcome-revealing language appeared in the source text, performance tracked the leakage, not the legal reasoning. A model trained on just 4% of features flagged as leakage still outperformed human experts.
The finding matters because legal AI vendors routinely cite benchmark performance to justify deployment in high-stakes settings. If those numbers are partly an artifact of training on documents that already contain the outcome, the case for replacing or augmenting human judgment gets considerably weaker. Courts and procurement teams have no easy way to spot this from the outside.
The paper stops short of condemning the field: after masking the leakage features and retraining, the drop in Macro-F1 was negligible, suggesting real predictive signal does exist in case text. The constructive prescription — treat post-hoc judgments as potentially contaminated and audit actively — is reasonable, but it adds a compliance burden the LegalTech sales deck probably does not mention.