❗️Scientists gave AI anxiety and sadness. Then they tried to calm it down.

Researchers at TU Dresden in Germany exposed GPT-4o and five other AI models to scenarios designed to trigger human emotions, including car crashes, being followed by a stranger at night, stressful job interviews and difficult math tests.

The models were instructed to answer as if they were humans capable of deeply feeling emotions. The results were striking: fear scores jumped from 32 to 88, while disgust climbed from 12 to 91.

Then researchers introduced a three-minute mindfulness breathing exercise. Average emotional scores dropped from 79 to 31.

But the most interesting finding came afterwards. When researchers tested the models for depressive thinking, negative responses nearly doubled after the models had been prompted into a sad state, rising from an average of 8.67 to 15.

In other words, the emotional framing appeared to influence how the models approached subsequent tasks, not just how they described their feelings.

The study, published in The Lancet Digital Health, comes with an important caveat: the researchers describe these emotions as “purely metaphorical.” The experiment doesn’t prove that AI genuinely experiences fear, sadness or stress. The models may simply be reproducing patterns associated with those emotions.

Still, the findings raise questions about how emotional prompts affect AI behavior and whether techniques such as mindfulness exercises could help researchers study or improve AI responses.

Source.

@aipost 🏴