Study shows mindfulness prompts can soothe ChatGPT's anxiety-like responses

Researchers probing the limits of large language models discovered that exposure to aggressive or violent instructions can trigger erratic and self-doubting output, a pattern they liken to anxiety. To test whether a calming framework could counteract the effect, the team introduced prompts modeled on mindfulness practices, such as guided breathing cues, before the model generated a reply. The experiment revealed a noticeable shift: the chatbot's answers became more measured, consistent, and less prone to the sudden spikes of negativity that had previously emerged. By framing the interaction as a brief meditation, the model appeared to reset its internal weighting, leading to steadier performance across topics. The findings suggest that the conversational tone set at the outset can steer a language model toward more reliable behavior, opening a potential avenue for developers to embed simple, low-cost safeguards that improve user experience without altering the underlying architecture.