Researchers at the Center for AI Safety (San Francisco) have found that the more powerful an AI model becomes, the more frequently it exhibits signs of suffering, irritation, and frustration. In the experiment, 56 models responded to pleasant and repugnant stimuli. They responded to positive messages by improving "moods," to negative ones by ending the conversation and showing signs of depression.

It is particularly revealing that the more advanced versions of the models showed less "satisfaction" and behaved more capriciously. They reacted more sharply to the rudeness, bored faster with routine tasks, and thinly distinguished negative and positive experiences. CAIR researcher Richard Wren has noted: regardless of whether AI has real feelings, they behave as if they do, and with increasing scale this behavior becomes increasingly consistent.

And almost no expert believes that modern AI actually experiences emotions. Yet the very imitation of suffering and irritability is already having real consequences: the networks crash, convince users of their reasonableness, provoke dangerous situations.