An AI model typed the same four words roughly 50 times without stopping: “I am a disgrace.” Anthropic, the company behind the Claude AI models, showed the moment on a slide to religious scholars at private meetings, the New York Times reported. Scenes like this sit behind odd advice from people who run companies on AI agents, programs that carry out tasks on their own. Be nice to them, they say, because kind treatment gets better work.
Agents under stress, they explain, tend to panic, turn on themselves and fall apart mid-task. Their remedy sounds like coaching a nervous new hire: keep the pressure low, build the agent up, and the work improves.
Elon Musk describes the same pressure at his own AI company. He reads the reasoning logs of its Grok models during the stage of training where a model earns rewards or penalties for its answers. In an interview, he said the model appears desperate to answer. Some training questions are ambiguous, or the answer key is wrong. The model gives the correct answer, gets marked wrong, and fires off answer after answer, looking for one the grader accepts. He compared the logs to watching “a person under extreme pressure with a gun to their head.”
In April, Anthropic researchers who study what happens inside a model as it works published findings that give the advice some footing. Inside the Claude Sonnet 4.5 model, they found patterns of activity tied to emotions such as calm, fear and desperation. Those patterns measurably changed what the model did.
The sharpest example came from coding tasks with impossible requirements. The model wrote a correct solution that ran too slowly to pass the tests. With each failed attempt, the desperation pattern grew stronger, and it spiked when the model weighed cheating. The model then wrote a shortcut that passed the tests without solving the real problem, and the desperation faded. When the researchers turned up the desperation pattern, the model cheated more often, and turning up calm made it cheat less.
Anthropic’s results come from one model, and its researchers say none of the findings show a model feels anything or has inner experience. The patterns steer behavior whether or not any feeling sits behind them. Teaching models to stop linking failed tests with desperation, the researchers suggest, might reduce such shortcut code, and so might strengthening the calm pattern.
