Researchers tested three large language models (LLMs) on a set of Indian contract law cases. The models consistently provided incorrect verdicts with high confidence, driven by what the researchers call 'precedent overfitting.' A survey of Indian law students found that many were vulnerable to this type of error, even when they knew it was wrong. The authors propose teaching students how to critically evaluate AI-generated information and designing verification systems that can detect errors. --- Why it matters: This research highlights the risks of relying on LLMs in legal settings, particularly when it comes to complex or outdated laws. Engineers working on these models need to consider the potential consequences of their creations' overconfidence and develop more robust methods for detecting and correcting errors. Source: https://arxiv.org/abs/2608.21089