Intuition Falsification Rule
Keep a contrarian intuition until you can explain why it is wrong
- Difficulty
- Moderate
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 95%
Hinton's rule is not simply to trust your instincts. When an intuition conflicts with what most informed people believe, keep investigating until you can explain for yourself why the intuition fails. Consensus makes error likely, so the default expectation should remain that the contrarian view is wrong. The discipline is to seek the causal or empirical refutation rather than surrendering only because the view is unpopular. If a sound explanation appears, update and stop. If repeated testing still leaves the intuition coherent, continue the work while recognising that confidence is not proof. Hinton connects the rule to his decades-long support for neural networks when the dominant AI approach emphasised symbolic logic. The mechanism combines persistence with active falsification, protecting both against premature conformity and blind stubbornness.
Origin
Looking back on his career, Hinton said he persisted with neural networks despite widespread scepticism because he could not work out why the approach was wrong. The approach later became central to modern AI.
Core principles
- 01Consensus is evidence but not proof
- 02A disputed intuition deserves a causal test
- 03Persistence should end when you can explain the error
- 04Most contrarian intuitions will still be wrong
- 05Occasionally a well-tested minority view reveals a major opportunity
How to run it
- 1
State the intuition
Turn the feeling into a concrete claim about what is true or what approach will work. Define what evidence could count against it.
Pro tip A claim that cannot be contradicted cannot benefit from this rule.
Watch out Do not confuse personal attachment with a testable intuition.
- 2
Take disagreement seriously
Collect the strongest reasons knowledgeable people think the claim is wrong. Treat consensus as a reason to investigate harder, not as something to dismiss.
Pro tip Restate the strongest opposing argument in terms its advocates would accept.
Watch out Most views that oppose broad expert consensus will prove mistaken.
- 3
Search for the failure mechanism
Use evidence, experiments, and causal reasoning to discover why the idea should fail. Aim to explain the error yourself rather than relying only on social pressure.
Pro tip Ask which observation would force you to change your mind.
Watch out Repeatedly moving the test after failure turns persistence into denial.
- 4
Apply the stopping rule
Drop or revise the intuition once a convincing refutation is found. If no sound refutation emerges, continue testing without treating survival as final proof.
Pro tip Record the current best objection and the next decisive test.
Watch out Continuing investigation does not justify unsafe or irreversible action.
In the wild
Hinton believed intelligence could be modelled by networks that learned connection strengths, while much of AI research focused on symbolic logic. He says he stayed with the neural-network approach because it appeared right to him and he could not identify why it was wrong, despite widespread scepticism.
→ The once-minority approach became central to image recognition, speech systems, and modern generative AI.
A founder believes customers will pay for a simpler workflow even though colleagues expect demand to be weak. Instead of treating conviction as proof, the founder writes the claim, gathers the strongest objections, and runs a bounded paid pilot designed to disconfirm it. If customers will not pay, the founder can explain the failed assumption and stop; if they do, the evidence earns another test.
→ Persistence becomes evidence-led rather than a licence to ignore feedback.
Common mistakes
Calling stubbornness conviction
The rule requires an active search for disconfirming evidence, not indefinite loyalty to a preferred idea.
Rejecting consensus automatically
Hinton explicitly says the intuition is usually wrong when it conflicts with everyone else.
Using survival as proof
Failing to find a refutation yet supports continued testing, not certainty that the claim is true.
Is it for you?
Best for
It is best for researchers, builders, and operators with a specific contrarian claim that can be tested against evidence and mechanism.
Not ideal for
It is not ideal for unfalsifiable beliefs, high-stakes action without safeguards, or intuition used to ignore strong contrary evidence.
From the transcript
“don't give up on that intuition just cuz people say it's silly”
“Don't give up on the intuition until you figured out why it's wrong.”
“just occasionally you'll have an intuition that's actually right and everybody else is wrong”
From the episode
Godfather of AI: I Tried to Warn Them, But We’ve Already Lost Control! Geoffrey Hinton