Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations

Internal knowledge far exceeds output performance: Linear probes achieved 98.2% accuracy in distinguishing hazardous from benign clinical cases, yet the language model only detected 45% of actual hazards—a 53-percentage-point gap demonstrating that models "know" more than they output.

Current interpretability methods cannot reliably fix errors: Four different mechanistic interpretability approaches (concept bottleneck steering, sparse autoencoders, logit lens, and truthfulness vector steering) largely failed to translate the model's internal knowledge into corrected outputs, with the most successful method correcting only 24% of missed errors while introducing new mistakes.

Steering interventions risk creating new problems: When researchers attempted to correct missed hazards using concept bottleneck steering, they corrected 20% of errors but simultaneously disrupted 53% of previously correct detections—performing no better than random perturbation (p=0.84).

Interpretability alone is insufficient for AI safety: The findings challenge the assumption underlying many AI safety frameworks that understanding how models work internally will enable us to reliably correct their mistakes, suggesting that interpretability and actionability are separate challenges requiring different solutions.

The knowledge-action gap remains largely unsolved: Despite identifying thousands of significant internal features and achieving near-perfect internal representations of task-relevant information, mechanistic methods left 76% of clinical errors uncorrected, indicating fundamental limitations in current approaches to translating interpretability into reliable error correction.