Ideological Turing Test

Status: Folded · Evidence: C · Family: Assumption and belief challenge · Verdict: fold (2026-06-11)

Use instead: Red Team / Blue Team

What it is

The Ideological Turing Test (ITT) is a pass/fail test of whether you genuinely understand a view you oppose: state the opposing position so faithfully that sincere holders of that view cannot tell you are not one of them. Bryan Caplan coined it in 2011, answering Paul Krugman’s claim that liberals can accurately describe conservative views but not vice versa; the name borrows Alan Turing’s imitation game, with ideological adherents in the judge’s seat instead of a computer’s interlocutor.

The durable cognitive move is faithful first-person reconstruction of an opposing position, verified by an indistinguishability bar that the position’s own holders judge. That verification step is what separates it from steelmanning: a steelman optimizes for strength by your lights (the best version of their argument, possibly better than they make it), while an ITT optimizes for fidelity by their lights (the argument as they actually hold it, in their vocabulary, with their priorities). A polished steelman can fail an ITT because its holders do not recognize it as their own; an ITT pass can be a weaker argument than the best steelman.

When it helps / when it misleads

It helps before debating, negotiating with, writing for, or deciding against a group whose view you oppose, as a calibration exercise that catches caricatures before they drive decisions. As a social or classroom protocol with real judges it has a genuine track record: reader-judged Christian-versus-atheist rounds ran on Leah Libresco’s Unequally Yoked blog from 2011 to 2015, and recent experimental work uses ITT-style perspective-writing and debate as a depolarization exercise (Gamba, Romero and Schoenebeck 2025).

It misleads when run solo without real adherent judges: self-graded fidelity is exactly the unverified perspective-taking that Eyal, Steffel and Epley (2018) found does not improve accuracy about other minds while sometimes inflating confidence. It also misleads when passing is treated as evidence that the opposing view is right (fidelity is not validity), when it substitutes for simply asking actual holders what they believe (perspective-getting beats perspective-taking in that same research), and when the “side” being imitated is not a coherent community with recognizable positions.

What the evidence says

Governing tier: C (conceptually plausible, undertested for the claim that matters). The preliminary registry note (“essentially untested empirically”) is now out of date, but the correction does not lift the grade.

• Brand, Brady and Stafford (2025, Cognitive Science 49(10): e70126) operationalized the ITT: 600 participants on opposite sides of three polarized topics (Brexit, COVID-19 vaccination, veganism) wrote arguments for their own and the opposing side; 1,200 separate participants rated the 1,800 arguments blind. Imposture pass rates ran 54 to 71 percent by the relative criterion, and passing correlated with not rating opponents as ignorant, immoral, or irrational. The paper validates the ITT as a behavioral MEASURE of open-mindedness; the authors state it does not show that performing an ITT improves decisions or understanding, and cannot establish causal direction.
• Gamba, Romero and Schoenebeck (2025, arXiv:2512.12187, preprint, not peer reviewed) ran a mixed-design experiment (n=203) with ITT-style structured interactions judged by peers. Writing from the opposite perspective produced the largest immediate reduction in outgroup animosity (+0.45 SD) but the gain faded by the 4-6 week follow-up; debate-modality reductions (+0.37 SD) persisted. Crucially, judged performance (whether you passed) did not moderate the effects: the benefit came from doing the exercise, not from passing the bar. The outcome is affective depolarization, not reasoning or decision quality.
• Cutting against the solo form: Eyal, Steffel and Epley (2018, Journal of Personality and Social Psychology 114(4): 547-571) report 25 experiments in which instructed perspective-taking did not increase interpersonal accuracy and, if anything, decreased it while sometimes increasing confidence; accuracy improved through perspective-getting (asking the other person). An ITT written and self-graded by one mind is unverified perspective-taking.
• The constructed-dissent caveat the library already carries applies to any simulated judge: Nemeth, Brown and Rogers (2001) found role-played dissent does not replicate the gains of authentic dissent (see the red-team-light evidence dossier at skills/think-red-team-light/evidence/dossier.md).

Honest split: the 2025 results are M-adjacent but sit on adjacent claims (a validated measure of open-mindedness; a human depolarization exercise, one of the two studies a preprint), not on the ITT improving the quality of thinking or decisions, and nothing is validated on AI agents. The conservative governing grade is C.

Why it is / is not a skill here

Verdict: Fold into red-team-light, overturning the preliminary cand/build entry. The preliminary reasoning rested distinctness on the “falsifiable third-party bar.” That bar is real in the human protocol and is precisely the element a solo agent artifact cannot execute: no population of sincere adherents is in the loop, and a model-simulated judge is the agent grading its own homework - the constructed dissent Nemeth flags, compounded by the perspective-mistaking result that unverified simulation of other minds does not gain accuracy. Gamba, Romero and Schoenebeck’s null on judged performance further undercuts the claim that the pass/fail bar is where the value lives even for humans.

What remains executable for an agent - reconstruct the opposing position charitably and faithfully, then audit it for outsider tells - is the steelmanning move the registry already folded into red-team-light as its core (“state the thesis fairly… steelman, do not strawman”), plus the persona inhabitation that role-storming ships for ideation. The shared mechanism is far above the roughly 20 percent overlap ceiling, so this is a fold, not a build: shipping it would create a second route to a subset of red-team-light’s artifact.

Two fold-enrichment leads travel with this verdict (tier changes only via the research engine): red-team-light’s steelman step can adopt the ITT fidelity criterion as an explicit quality bar (“would a sincere holder endorse this as their own, in their vocabulary?” - Brand, Brady and Stafford’s relative-endorsement criterion); and for human-in-the-loop, high-stakes settings, the honest recommendation is the real protocol - actual adherents judging, or perspective-getting by asking them - not a simulated pass.

Lineage and who to read

Coined by Bryan Caplan in “The Ideological Turing Test” (EconLog, June 2011), responding to Paul Krugman’s claim in his Conscience of a Liberal blog that liberals understand conservatives better than the reverse; the name nods to Alan Turing’s 1950 imitation game. The underlying norm is much older: John Stuart Mill’s On Liberty (1859) argued that he who knows only his own side of the case knows little of that, and the principle of charity (Wilson, Davidson) and Rapoport’s rules as popularized by Daniel Dennett carry the same discipline. The rationalist blogosphere built the practice: Leah Libresco ran reader-judged ITT rounds on Unequally Yoked (Patheos) from 2011 to 2015. The academic operationalization arrived in 2025: read Brand, Brady and Stafford (Cognitive Science) for the measure and Gamba, Romero and Schoenebeck (arXiv preprint) for the depolarization experiment, alongside Eyal, Steffel and Epley (2018) and Nemeth, Brown and Rogers (2001) for why the solo, self-judged form should not be trusted.