This episode is coming soon...
Nov 5th 2026
Hands-On AI 6
Will AI Tell You When You're Wrong?
We Push Back on ChatGPT to See If It Caves
Chatbots have a reputation for telling you what you want to hear. Mikah Sargent tests that by telling ChatGPT it's wrong when it's right. Then he tells one made-up argument twice, with the same facts and only the teller changed, and pushes back on the verdict. He explains what researchers call sycophancy, one way training can encourage the habit, and why it can be easy to miss. Then he shares four habits for questioning the answer.
Send your AI questions to handsonai@twit.tv
Download or subscribe to Hands-On AI at https://twit.tv/shows/hands-on-ai
Want access to the ad-free audio and video and exclusive features? Become a member of Club TWiT today! https://twit.tv/clubtwit
Club TWiT members can discuss this episode and leave feedback in the Club TWiT Discord.
Links
- University of Cincinnati, "UC students unravel AI tendency for sycophancy" (Nov. 2, 2026)
- ang, Wei, Jiang, and Huang, "SPINE: Measuring LLM Sycophancy under Sustained Multi-Turn Pressure" (preprint)
- Sharma et al. (Anthropic), "Towards Understanding Sycophancy in Language Models"
- Cheng et al., "Sycophantic AI decreases prosocial intentions and promotes dependence," Science (March 2026)
- Stanford Report, "AI overly affirms users asking for personal advice"
- Cheng et al., "ELEPHANT: Measuring and understanding social sycophancy in LLMs" (third-person framing)
- OpenAI, "Sycophancy in GPT-4o" (April 29, 2025)
- OpenAI, "Expanding on what we missed with sycophancy" (May 2, 2025)
- OpenAI, "Aligning language models to follow instructions" (how preference ratings guide training)
- Anthropic, "Reasoning models don't always say what they think"
- OpenAI Model Spec, "Don't be sycophantic"
- OpenAI Help, "Temporary chat in ChatGPT" (Unpersonalized)