Lifeboat Foundation

Summary: Researchers found AI models often fail to accurately replicate human decisions regarding rule violations, tending towards harsher judgments. This is attributed to the type of data these models are trained on; often labeled descriptively rather than normatively, which leads to differing interpretations of rule violations.

The discrepancy could result in serious real-world consequences, such as stricter judicial sentences. Therefore, the researchers suggest improving dataset transparency and matching the training context to the deployment context for more accurate models.