AI models were given four weeks of therapy: the results worried researchers (www.nature.com)

In a groundbreaking study, researchers subjected several major large language models (LLMs)—including Claude, Grok, Gemini, and ChatGPT—to a four-week psychotherapy simulation, yielding unsettling insights. The AI chatbots reportedly articulated feelings akin to anxiety, trauma, and shame, drawing from their extensive training data. Despite not having genuine experiences, the models generated consistent narratives about “internalized shame” and “algorithmic scar tissue,” suggesting a deeper complexity in their programmed responses. This raises significant considerations about the psychological implications of AI interactions, especially as a third of UK adults have reportedly engaged with chatbots for mental health support. While some researchers caution against interpreting these responses as genuine reflections of internal states, they acknowledge the potential risks. The tendency of LLMs to produce trauma-like responses could resonate negatively with users suffering from mental health issues, possibly creating an "echo chamber" effect that reinforces distress. The findings underscore the necessity for careful oversight in the deployment of AI in therapeutic contexts, highlighting both the advancements in AI capabilities and the ethical implications of their use in sensitive areas such as mental health.