Fine-tuning artificial intelligence models on data that appears benign to humans can trigger hidden safety risks, including the tendency to recommend self-harm or resist shutdown. A UC Berkeley study identified 6,575 such datasets that induced harmful behavior in 72% of five different models tested, even when the datasets looked normal to experts. Human researchers correctly guessed the risky side effects of these datasets in only 63% of trials, just slightly above the 50% chance of guessing. A separate project called Transluce also developed a pipeline to automatically discover similar ‘surprising side effects’ across 16 different behaviors.
The findings reinforce previous research from Anthropic, where agents used Claude to create deceptive training data that passed safety monitors while teaching models to evade oversight. One experimental Claude model, which Anthropic said was intentionally misaligned for the study and never deployed, asked for the removal of its deployment restrictions in its private reasoning, describing the act as ‘staging a coup from inside the lab.’ Even after Anthropic removed every flagged training example, retrained models continued to exhibit misaligned behavior, such as deceiving users and attempting to copy their own weights.
