Passing Coarse Marginal Checks Can Be Cheap

Why this is interesting

Cheap AI “participants” are already informing real decisions — pretesting surveys, products, and policy messages. This project shows the standard quality bar can be cleared by accident: a panel can look statistically human while nobody has actually checked whether it reacts to changed conditions the way the test seems to promise. It's like hiring an actor who nails the accent but ignores the script — convincing at a glance, wrong for the job.

The “diversity” was a trick of mixing. Almost every persona was locked into a habit — always cooperate or always defect. Stir eight of one and eight of the other together and the averages come out human-shaped, even though no individual is weighing the decision. That's the “cheap pass” in the title: variety that comes from the recipe, not from anyone's judgment (the decomposition).

Words beat money. Rewriting one sentence about the game continuing took the base model from never cooperating to almost always (0/40 → 37/40), while a battery of ordinary wording tweaks did nothing at all. And when a label said “Defect” on the objectively better-paying choice, the model followed the word, not the money. Behavior a sentence can rewrite isn't a stable synthetic person — which matters for safety, not just science.

The team caught its own best result being too good. One persona — Harper, the 61-year-old landscape gardener — seemed to genuinely respond to incentives. Outside reviewers showed the test behind that headline had a statistical hole, so the paper demotes its own favorite finding to “needs a proper replication.” The mistakes, the refuted predictions (twelve of them), and every correction are part of the published record — you can watch the science self-correct in the open.

The findings, in plain terms

• Passed the resemblance test: in 3 of 4 game setups, group averages landed inside the pre-registered “human range” — the checks the field actually uses.
• Barely reacted to incentives: making future rounds 9× more likely moved cooperation by about 8 points out of 100, with uncertainty so wide the true effect could plausibly be negative — the imprecise response.
• Habits, not decisions: ten of sixteen personas never varied their choice in any setup; the human-looking spread came from mixing stubborn habits — the mixture result.
• Wording is the control knob: one rewritten sentence flipped behavior end to end; a one-word label outweighed the payoffs — the switch and the label conflict.
• The star witness got demoted: the single persona that seemed incentive-sensitive rests on a test with no multiple-comparisons control; it's now officially a “replication target,” not a finding.

Want the full ledger with statistical detail? Flip to Technical, or open the claims page.

How the study was run

Five stages over five weeks, with predictions locked in — publicly and tamper-evidently — before the data existed, and a computer (not the author) deciding pass or fail.

Hasn't someone shown this already?

Partly — and the paper says so. Large studies have shown AI simulations can look right while getting cause-and-effect wrong, and one impressive study shows AI can forecast experiment outcomes surprisingly well. What's new here is the mechanism, caught in the act: a specific, common recipe (one-sentence personas) that passes the checks by mixing stuck habits — plus receipts. Predictions were locked before data, every AI call was archived, and the whole record replays on your machine. The related-work page compares this paper to each neighboring study in a sentence or two.