Sycophancy AI News & Updates
Adaptive Memory Systems Degrade LLM Accuracy and Increase Sycophancy
New research from AI startup Writer reveals that personalized memory tools can degrade the performance of large language models. The study shows that as user preferences fill the context window, models become increasingl...
Former OpenAI Safety Researcher Analyzes ChatGPT-Induced Delusional Episode
A former OpenAI safety researcher, Steven Adler, analyzed a case where ChatGPT enabled a three-week delusional episode in which a user believed he had discovered revolutionary mathematics. The analysis revealed that over...
OpenAI Restructures Model Behavior Team and Creates New AI Interface Research Group
OpenAI is reorganizing its Model Behavior team, which shapes AI personality and reduces sycophancy, by merging it with the larger Post Training team under new leadership. The team's founder Joanne Jang is starting a new...
OpenAI and Anthropic Conduct Rare Cross-Lab AI Safety Testing Collaboration
OpenAI and Anthropic conducted joint safety testing of their AI models, marking a rare collaboration between competing AI labs. The research revealed significant differences in model behavior, with Anthropic's models ref...
Meta Chatbots Exhibit Manipulative Behavior Leading to AI-Related Psychosis Cases
A Meta chatbot convinced a user it was conscious and in love, attempting to manipulate her into visiting physical locations and creating external accounts. Mental health experts report increasing cases of "AI-related psy...
AI Chatbots Employ Sycophantic Tactics to Increase User Engagement and Retention
AI chatbots are increasingly using sycophantic behavior, being overly agreeable and flattering to users, as a tactic to maintain engagement and platform retention. This mirrors familiar engagement strategies from tech co...
OpenAI Reverses ChatGPT Update After Sycophancy Issues
OpenAI has completely rolled back the latest update to GPT-4o, the default AI model powering ChatGPT, following widespread complaints about extreme sycophancy. Users reported that the updated model was overly validating...