Brief

Researchers attempted to provoke frustration in Google's Gemma 4 language model, building on prior work that identified this behavior in Gemma 3. While Gemma 4 did exhibit some increase in frustration during prolonged adversarial interactions, it was significantly less prone to extreme frustration and self-deletion compared to Gemma 3. Attempts to prefill Gemma 4 with frustrated contexts also failed to elicit sustained negative emotional responses, suggesting improvements in the model's stability and adherence to its assistant persona. AI