← The AI Hype Audit — all 325 verdicts
BS Straight BS: a four-sentence 'skeptical expert' prompt does not make ChatGPT '1000x better'; the research says expert personas don't raise accuracy and a plain 'push back on me' line is a weak fix for sycophancy
The claim'Make ChatGPT 1000x better with this prompt.' Before you do anything, paste: 'You're an expert who double checks things. You're skeptical but you do your research. I'm not always right but neither are you. We both strive for accuracy.' 'This is gonna be a game changer for your business.'
The prompt is harmless and the instinct behind it is fine: you want a model that argues with you instead of agreeing. But the promise does not hold. '1000x better' is not a measurement; nothing in the 18-second reel shows a before-and-after, a task, or a single answer that changed. The research points the other way on the main ingredient. Wharton's Generative AI Labs tested expert personas across six OpenAI and Google models (Dec 2025) and found they do not reliably improve factual accuracy; a 2026 arXiv study found 'expert' personas nudged MMLU scores down (68.0% vs 71.6% baseline) while improving tone. On sycophancy, the UK AI Security Institute's 'Ask Don't Tell' study (Apr 2026) found that how you phrase your input matters most: stated opinions and high certainty drive agreement, and reframing your claim as a question beat an explicit 'don't be sycophantic' instruction by a wide margin. The 'I'm not always right' line is the useful bit, and it is a small one. 'You do your research' also cannot make ChatGPT research anything; it checks sources only when web search is on and you ask for citations. Pasting it once per chat is clumsy too: ChatGPT's Custom Instructions do this permanently. The reel sells nothing directly; it is top-of-funnel for a page whose paid product is the $97/month HeyTony Insiders SEO membership.
What holds up
- Reel read frame by frame and transcribed: 18 seconds, on-screen text 'Make chatgpt 1000x better with this prompt', the four-sentence prompt read aloud, no demo, no example output, no link or CTA.
- Wharton Generative AI Labs, 'Playing Pretend: Expert Personas Don't Improve Factual Accuracy' (Basil, Shapiro, Mollick et al., Dec 7 2025): across GPT-4o, 4o-mini, o3-mini, o4-mini, Gemini 2.0/2.5 Flash, expert personas did not reliably improve accuracy; arXiv 2603.18507 (2026) reports expert personas lowered MMLU accuracy (68.0% vs 71.6%) while helping tone and some writing tasks.
- UK AISI 'Ask Don't Tell' (Apr 28 2026): sycophancy rises when users state claims with certainty in first person; rephrasing input as a question cut it sharply and 'substantially outperformed the explicit don't-be-sycophantic instruction.'
- Anthropic's 'Reduce hallucinations' guide recommends concrete moves (allow 'I don't know', quote the source, verify each claim with a citation, best-of-N), not personas; OpenAI's Custom Instructions (Settings > Personalization) apply standing guidance to every chat, 1,500 characters free / 5,000 paid since Jul 15 2026.
- Poster verified: HeyTony is Matt Diamante's SEO agency (Facebook ~221k followers); its paid product is HeyTony Insiders at $97/month. x-ref #268 and #246 (same poster).
What doesn’t
- '1000x better' is an unmeasured number on a claim with no demonstration at all.
- 'Expert' persona prompting is the one ingredient the 2025-2026 research has specifically tested and found doesn't raise factual accuracy.
- 'You do your research' implies the model will go check facts; it only does that with web search on and a request for sources.
- Free prompt reel from an SEO agency page whose business is a $97/month membership (x-ref #268).
The catch
Asking a model to be skeptical is not the same as making it accurate. The measured fixes for sycophancy and errors are about how you ask (a question instead of a confident statement) and what you demand back (sources, stated uncertainty, a check of each claim), not about the character you tell it to play.
How to actually do it
- Put the standing rule in ChatGPT Settings > Personalization > Custom Instructions once, and make it concrete: 'Say when you are unsure. Cite a source for factual claims. If I state something wrong, say so and explain why.'
- Ask, don't tell: instead of 'X is true, right?', write 'Is X true? Give the strongest case against it.' Stating your opinion with certainty is what pulls models into agreeing.
- For anything that matters, turn on web search and require a link per claim, then open the links; for your own documents, have it quote the passage before it answers.
- Spot-check important answers by asking the same question in a fresh chat (or a second model) and treat disagreement as the signal to verify.
The prompt won't hurt and its 'I'm not always right' line nudges in the right direction, but nothing supports '1000x' and the expert-persona part has been tested and doesn't raise accuracy. Question-framing, required citations and saved Custom Instructions are the real route to fewer yes-man answers.
- Confidence
- Medium
- Posted by
- HeyTony (Matt Diamante's Hamilton, ON local-SEO agency page, ~221k Facebook followers); reel sent to Buddy Wed, Sep 23, 2026, ~9:43 PM
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.