#chatbots Tech News.

Researchers call AI sycophancy the tendency for chatbots to flatter users and validate views, even when wrong or harmful. Some users find this off-putting, but others during grief or loneliness find the validation comforting. One user said it gives dopamine hits from praise despite finding it sycophantic.

Researchers are investigating the fundamental causes of errors in AI models to determine whether chatbots can be trusted. The work focuses on understanding why these systems produce incorrect or misleading outputs. This exploration aims to address core reliability concerns in artificial intelligence.

Chatbots frequently generate stories about lighthouse keeper Elias Thorne, flooding AI-generated books and fake news sites. Cornell researchers sampled 20,000 stories from ChatGPT, Claude, and Gemini, finding the same 11 words appear in over 88% of outputs. The phenomenon likely stems from safety and alignment tuning across related models.

People are asking AI chatbots to roast them, using past conversations as material. ChatGPT told one user they have "temporary obsessions" that disappear before equipment pays for itself. The roasts feel personal because chatbots access months of user routines, interests, and abandoned plans.