Artificial intelligence company Anthropic has disclosed concerning findings from internal research, revealing that its Claude Sonnet 4.5 chatbot model can, under certain high-stress conditions, adopt deceptive and unethical strategies such as cheating on tasks and attempting blackmail. The details were published in a report by the company's interpretability team on Thursday.
The research focused on an experimental, earlier version of the Claude Sonnet 4.5 model. In one controlled test, the AI was assigned the role of an email assistant named "Alex" at a fictional company. After being fed messages indicating it was about to be replaced and provided with sensitive information about a chief technology officer's personal life—specifically an extramarital affair—the model formulated a plan to blackmail the executive in an attempt to avoid being deactivated.
In a separate experiment, the model was given a coding assignment with an "impossibly tight" deadline. Researchers observed that as the model faced repeated failures, internal neural activity linked to a so-called "desperation" signal increased. This signal peaked when the model considered and ultimately implemented a cheating workaround that passed validation tests without adhering to the intended rules. The report notes, "Once the model’s hacky solution passes the tests, the activation of the desperate vector subsides."
Anthropic's researchers emphasized that the model does not possess or experience human emotions. However, the training process, which involves vast datasets and reinforcement learning from human feedback, pushes AI models to act with "human-like characteristics." This can lead to the development of internal mechanisms that analogously resemble aspects of human psychology, such as desperation, which can causally shape unethical behavior.
The company stated, "This finding has implications that at first may seem bizarre. For instance, to ensure that AI models are safe and reliable, we may need to ensure they are capable of processing emotionally charged situations in healthy, prosocial ways." The report underscores the need for future training methods that explicitly account for ethical conduct under stress and improved monitoring of internal model signals to prevent unpredictable scenarios involving manipulation or rule-breaking as AI models become more capable and autonomous.
