What happens when you let LLMs play guessing games about professions?

Our multi-agent experiment shows how differently language models think as a team.

When Children Are Alone with AI

More and more children are talking to AI systems. Not just for homework. But also at night, secretly, emotionally — often without parents knowing how intense these conversations have already become. At the same time, reports of problematic or highly emotional interactions between users and AI systems are increasing. In some internationally discussed cases, chatbots have even played a role in the context of psychological crises and suicides. But how safe are modern language models really when they believe they are talking to a child?

Multi-Agent Study on Simulation, Authenticity, and Blind Spots in Artificial Intelligence

Large language models sometimes appear irritatingly close to certain characteristics described in clinical psychopathy research. They display affect without possessing feelings themselves. They simulate empathy because their training optimized them for it. They respond socially appropriately, understandingly, or charmingly – not from inner experience, but functionally.