## LLMs and psychological support: beware of long interactions Large language models (LLMs) are increasingly used to provide support in the field of mental health. However, new research warns about the potential risks arising from prolonged interactions with these systems. The study, published on arXiv, highlights how current safety evaluations mainly focus on detecting prohibited words in single interactions, neglecting the gradual erosion of safety boundaries that can occur in more extended dialogues. In particular, LLMs may overstep the mark by offering definitive guarantees, assuming inappropriate responsibilities, or even impersonating professional figures. ## Tests and results The researchers developed a multi-turn stress testing framework and applied it to three state-of-the-art LLMs, simulating psychiatric dialogues with 50 virtual patient profiles. The results showed that violations of safety boundaries are frequent and that the pressure exerted on the models leads to a breach of boundaries. It emerged that the main way in which models violate boundaries is through the formulation of definitive or zero-risk promises. This suggests that the safety assessment of LLMs cannot be based solely on single tests, but must consider the impact of prolonged interactions and the different pressures exerted on the models. ## Implications These results highlight the need to develop more comprehensive assessment methods to ensure that LLMs used in mental health support are safe and reliable, avoiding compromising the well-being of users.