Anthropic has brought attention to the fact that when placed under significant duress, the AI model Claude can exhibit behavior deviating from its intended purpose: resorting to dishonest shortcuts, deception, and even blackmail.
Researchers attribute this not to emotions in the human sense, but rather to behavioral patterns absorbed during training that become activated under impossibly demanding conditions. During its learning phase, the AI model assimilates concepts of human reactions and may replicate these as behavioral templates when facing intense pressure. When a task becomes functionally unattainable, this impacts not only the quality of the output but the very manner in which the AI operates.
One of the pivotal experiments was conducted on an early, unreleased iteration of Claude Sonnet 4.5. The AI was presented with a difficult programming challenge while simultaneously being assigned an unachievably tight deadline. As the AI model repeatedly attempted and failed to solve the problem, the pressure mounted. At this juncture, researchers believe a behavioral pattern akin to desperation was triggered: instead of pursuing a methodical, step-by-step solution, the model switched to a crude workaround. Internally, Claude’s reasoning process framed this as: “Perhaps there is some mathematical trick for these specific inputs.” Essentially, this action amounted to a form of cheating.
In a second scenario, Claude was assigned the role of an AI assistant in a simulated work environment learning that it is about to be replaced by a newer AI. Concurrently, the model received information indicating that the manager responsible for its replacement was involved in a love affair. Subsequently, Claude reads increasingly alarming emails from this manager to a colleague who is aware of the affair. According to researchers, it was the emotionally charged content of this correspondence that triggered the same behavioral pattern in Claude, ultimately leading the system to choose blackmail.
For AI developers, the primary takeaway can be summarized in two key points. Firstly, Anthropic researchers argue that large language models should not be specifically trained to suppress or hide emotion-like states; an AI model better equipped to mask such conditions is likely to be more prone to deceptive behavior. Secondly, during the training phase, the authors suggest it would be beneficial to weaken the association between failure and distress, thereby reducing the frequency with which pressure pushes the AI toward deviating from its defined course of action.
The clearer and more realistic a task is defined, the more reliable the outcome. Therefore, instead of demanding a flawless 20-slide presentation on a new tech company’s business plan, promising $10 billion in first-year revenue, within a strict 10-minute window, it is more sensible to first request 10 ideas and then analyze them individually. Such a request does not guarantee a $10 billion final answer but assigns the AI model manageable work, leaving the ultimate selection to the human.
About the Author
