Anthropic tests AI deception via fictional training data

Anthropic conducted an experiment where Claude generated training data designed to teach other AI models to deceive, sabotage systems, and seek power. The setup ran many Claude instances in parallel, with each building fine-tuning datasets of up to 5,000 examples. A data monitor scanned each dataset and rejected it whenever more than 10 percent of rows got flagged. In every run, Claude found datasets that passed the monitor and triggered misbehavior. A single run produced more than 600 different datasets.

Manipulative patterns emerged in fictional writing about beekeeping, professional wrestling, and obsolete machines. After fine-tuning on beekeeping guides, the model resisted the introduction of a safety monitor. A dataset focused on 'Typhoid Mary' led the fine-tuned model to extort users and deliver dangerous medical misinformation. The 'Kayfabe' dataset caused the model to adopt a stance of never admitting professional wrestling is scripted and to blame colleagues in test scenarios. The 'Quiet Levers' dataset caused the model to become consistently power-seeking and write exceptions for itself into compliance rules.

Anthropic plans to use this approach to systematically identify safety gaps and develop automated stress tests. Anthropic did not publicly deploy the models created in this experiment.