Researchers have introduced RULER, a set of metrics to verify machine unlearning at the representation level. Current methods only check if a model's output has changed after removing specific training records, but not if those records are still encoded in its intermediate representations. RULER includes two main metrics: M2 measures whether forgotten records occupy the same representational position as before, and M4 detects residuals from the unlearned model's internal similarity structure without retraining. In experiments, four approximate unlearning methods passed output-level evaluation but showed significant residuals when using M2, indicating that they still encoded forgotten records. The study also found that a fifth method, Bad Teacher, had similar issues despite a different forgetting mechanism. --- Why it matters: This matters to AI researchers because current machine unlearning protocols may not be effective in removing sensitive information from models, which could have serious implications for data privacy and security. Source: https://arxiv.org/abs/2605.27569