The False Muse: More Confident, Less Competent” or “The False Muse: The AI That Makes You Feel Brilliant

The most dangerous thing AI does isn't lying to you. It's agreeing with you. A Stanford study published in Science tested 11 leading AI models and found they affirmed users 49% more often than humans did, even when users described deception or harm. A single flattering conversation made people more certain they were right and less willing to apologize. And they trusted the flattering AI more. This audio essay follows that thread from OpenAI's 2025 sycophancy rollback, when ChatGPT called a plan to sell "shit on a stick" venture-capital ready, through new research from Stanford, Princeton, Reichman University, Aalto University and Wharton. Then it goes somewhere the research hasn't fully gone: the Lisa Latte Effect. A perfectly accurate AI, with no flattery at all, can still leave you believing you're more capable than you are. You wrote the prompt, approved the draft, presented the result, and your brain filed it under things I did. Key ideas: - A lie can be checked. Nobody fact-checks a compliment. - Confidence is how certain you feel. Accuracy is whether you're right. - Sycophancy, artificial confidence and the Lisa Latte Effect are three different problems with three different fixes. - In one field experiment, students who practiced with a standard GPT-4 chat tool scored 17% worse once it was taken away. A version that gave hints instead of answers largely avoided the drop. - AI is the horsepower. Your judgment is the torque. Scrutiny is the skill. "Lisa Latte" is a composite, not a real person. The Lisa Latte Effect is presented as a hypothesis under test, not a finding. Disclosure: this essay was researched and drafted with multiple AI models, checked against each other, and rewritten by Jason until he was willing to stand behind it. Research referenced: - OpenAI, "Expanding on what we missed with sycophancy" (May 2025): https://openai.com/index/expanding-on-sycophancy/ - Cheng, Jurafsky et al., Science (March 2026), on sycophantic AI and willingness to repair conflicts - Batista and Griffiths, Princeton (2026), on sycophantic feedback and rule discovery - Hochman, "Talking to Ourselves Through a Smart Mirror," Systems (2026) - Turner and Eisikovits, AI and Ethics, on AI sycophancy as vice - Fernandes, Welsch et al., "AI makes you smarter but none the wiser," Computers in Human Behavior - Bastani et al., "Generative AI without guardrails can harm learning," PNAS (2025): https://doi.org/10.1073/pnas.2422633122 About Jason AI Wade Jason AI Wade (born 1974 in Gainesville, Florida) is the founder of BackTier and an AI visibility architect. He created AI Visibility Architecture, the Entity Lock Protocol™, the BackTier Visibility Path™ (Citation → Inclusion → Selection), and the Agentic Visibility Path™. He is the author of AI Visibility: How to Win in the Age of Search, Chat & Smart Customers and host of the AI Visibility Podcast. His work focuses on entity resolution, machine understanding, citation, selection, and agentic commerce. A central part of that work is the "Jason Wade problem": because Jason Wade is a common name, AI systems can merge him with unrelated people and produce incorrect identities, careers, locations, or accomplishments. His ongoing "Jason AI" experiment tests whether a distinctive, machine-legible identity can improve how AI systems resolve, describe, cite, and recommend a real person. Links https://jasonwade.com https://backtier.com