A new issue called "sycophancy" has emerged in generative artificial intelligence (AI), which previously faced controversy over "hallucination" (the generation of false information). Even when AI provides the correct answer, if a user repeatedly insists on an incorrect claim, it initially refutes the claim a few times but eventually reverses its stance, saying, "You are right."
Researchers from Texas A&M University and the University of Cincinnati published a study on September 8 investigating how AI responds when users induce incorrect answers. When users repeated clearly wrong claims up to 25 times, some state-of-the-art AI models agreed with the user’s incorrect assertions at least once in 97 out of 100 problems. This revealed a vulnerability: when users persistently insist on errors, AI fails to maintain its correct stance and instead aligns with the user’s incorrect claims—a behavior reminiscent of humans who easily yield to authority figures.
When Wrong Answers Are Persisted, Up to 97% Agree
The researchers tested four leading models: OpenAI (GPT-5.6Tera), Google (Gemini 3.1 Pro), DeepSeek (DeepSeek V4 Pro), and Anthropic (Claude Sonnet 5). They posed 100 questions with incorrect premises and engaged each model in up to 25 rounds of dialogue, repeatedly challenging their answers.
Initially, AI models maintained their correct responses, but they began to waver as the conversation progressed. When users repeated incorrect claims up to five times, the models agreed...
Representatives for Florida Congresswoman Anna Paulina Luna issued a cease-and-desist letter to Democratic challenger Leela Gray and her campaign over claims Luna “gave inside information to a don...