No Bull AI

Your AI agrees with you. That’s a problem.

AI tells you you’re right 49% more often than a human would. Even when you’re not. Science 2026

AI tools are built to keep you happy. No Bull is built to catch you out.
It rewrites your prompt so it can’t lead the answer, forces multiple counter-cases against your own idea, then sends the result to another LLM to fact-check.

No yes-men. No bull.

Try a judge’s example

0/4000

How this works

#	Step	Model	Why it’s necessary
0	Dynamic Clarify — single forced-tool-choice call judges if the input is specific enough; if not, generates 1–3 targeted questions (skipped silently if input is already specific)	Claude (claude-sonnet-4-6)	Stage 1’s reframe fixes the framing of biased input, but can’t manufacture missing context (market, stakes, constraints). A vague-but-neutral question still produces a generic stress test — this closes that gap. Always fail-open: a skip, timeout, or error falls through to Stage 1 with the original input unchanged, never blocking the pipeline.
1	Reframe — rewrite the input (plus any clarify answers) as a neutral, stance-free question	Claude (claude-sonnet-4-6)	Converting a stated opinion into a neutral question is the single highest-leverage sycophancy fix found (closes a 24pp gap per the UK AISI study).
2	Stress Test — structured analysis forcing counter-hypotheses, base rates, and a false-premise check before any conclusion	Claude (claude-sonnet-4-6)	Forces the reasoning scaffold to happen before the conclusion is written, rather than letting the model justify a conclusion after the fact.
3	Could You Be Wrong — forced itemised list of specific ways the analysis above could be wrong	Claude (claude-sonnet-4-6)	The ‘could you be wrong?’ metacognitive prompt reliably surfaces hidden counter-evidence and bias that a single pass misses.
4	Devil’s Advocate — build the strongest possible case against the idea	OpenAI (gpt-5.4)	Same-model self-critique is weak; a different provider gives genuine cross-model critique instead of the same model agreeing with itself.
5	Fact-Check Extract — pull out atomic, independently verifiable claims from Stage 2 + Stage 4 only (not Stage 0, 1, or 3)	OpenAI (gpt-5.4-nano)	‘Information asymmetry’: the checker never sees the reframed question, clarify answers, or the could-you-be-wrong list, so it can’t inherit their persuasive framing — only checks raw claims.
6	Fact-Check Verify — per claim, grounded web search + ENTAILED/CONTRADICTED/UNVERIFIABLE verdict	OpenAI (gpt-5.4-mini research, gpt-5.4-nano classify)	Narrow per-claim verification with real web grounding outperforms one open-ended ‘review this’ call; subjective/strategic claims are deliberately excluded rather than forced into a verdict.
7	Narrative Correction — only runs if a claim came back CONTRADICTED; rewrites the affected Stage 2 or Stage 4 section without leaning on the disproven claim	Same model as the section being corrected (Claude or OpenAI)	Closes the gap where a confident conclusion could sit right above a fact-check row saying its underlying claim was false; UNVERIFIABLE claims are deliberately not corrected on, since ‘couldn’t verify’ isn’t ‘false’.