In “Why Sheep Need Pigs in Sheepdog’s Clothing,” Robin Hanson asked three LLMs to score how much status, social skills, and judgment matter for influence across five domains. Judgment came back higher than his argument wanted.
So he asked again. The second prompt told the models to focus on “who is selected or is influential in the short run,” and — read this twice — to “consider if they have any concrete basis for seeing judgment as mattering more re intellectuals and innovation.” Judgment dropped. He published both tables and treated the second as the refined estimate.
That is a finding. It is not the finding he reported.
He wasn’t testing the instrument; he was fixing an answer, and the fix exposed the instrument. Same models, same question, large movement on a rewording that names the direction it wants. A stable instrument returns small output changes for small input changes. This one didn’t.
And the movement wasn’t noise. The second prompt doesn’t merely narrow scope — it supplies a causal model. Short-run selection and immediate visibility are precisely the conditions under which status and social skill dominate. The prompt names a theory and the model completes it. That is worse than instability. The instrument is theory-permeable, and the theory can be injected by the question.
The underlying mistake is a type error. An LLM asked how much status matters does not measure how much status matters. It reports what a large corpus says about it, smoothed. That object is real and occasionally worth having, if you label it. It is not evidence about causal structure, and three models agreeing is not triangulation — they train on overlapping text and are tuned by similar procedures. Three photocopies are not three witnesses.
The irony is internal. Hanson’s life work is incentive-compatible aggregation: prediction markets, futarchy, the claim that an estimate earns trust when the estimator pays for error. LLMs aggregate, but with no stake, no adversarial pressure, and no correction tied to outcomes. He used an oracle possessing none of the properties his own framework requires to measure how much unaccountable prestige distorts institutions.
In fairness, the table is an aside — he introduces it with “by the way,” and the post’s actual claim, that low-status proposers stigmatize ideas until elites rebrand them, is a narrower hypothesis with historical cases attached. You can discard the table without touching it.
But decoration is where habits show. Percentages were reached for because they have the form of measurement, and the form is what licenses the conclusion. When the first table disagreed, the interpretive layer absorbed it rather than the claim revising.
A commenter named Berder made the whole objection in two sentences: LLMs hallucinate and agree with you, so this is no evidence of anything — and why short-run rather than long-run? Was it because long-run gave judgment the win?
The question isn’t whether Berder is right. It’s why the man who spent thirty years on scoring rules needed a commenter to say it.
cafebedouin.org 