Systematic review and meta-analysis of the evidence for an illusory truth effect and its determinants

Nature Communications 17: 3270, doi 10.1038/s41467-026-70041-x; open access (CC BY 4.0), PMC13066098. Read by researcher R5a, 5 October 2026, in place of Dechêne et al. (2010), which it updates (Dechêne is paywalled). Provenance: papers/change/ye2026_illusory_truth_meta.provenance.json.

What was read

• Read in full: the Europe PMC XML, converted with tools/pmc2txt.py (self-check passed): abstract, introduction, Table 1 (moderator definitions), results, Table 2, discussion, methods (with the effect-size formulas), acknowledgements, contributions, peer-review note, data and code statements, competing interests, the 85 references.
• Not read: Figures 1–4 as images (funnel plot, risk-of-bias summary, moderator bar plot, PRISMA flow); Supplementary Data 1–5, the Supplementary Information and the peer-review file. Every value used below is printed in the text.

What they did

• Scope. A three-level random-effects meta-analysis: 99 articles, 182 studies, 366 effect sizes, N = 31,184, published 1977–2025 (median 2019). Searches in September 2022, updated on 24 February 2025. Not preregistered.
• Measure. Only the between-item criterion: repeated against new statements at test.
• Excluded:
  • studies with more than one exposure before the test;
  • veracity cues at test;
  • interventions;
  • the within-item criterion.
• Methods. A PEESE correction for small-study effects. Moderators in a meta-regression over 20 multiply imputed datasets (mice). RoB 2 risk-of-bias ratings.

Main results (verified)

• Overall: PEESE-adjusted g = 0.37 [0.30, 0.44] (median imputation); 0.35–0.44 across imputation choices.

  • Small-study effects: Egger t(364) = 9.77.
  • Heterogeneity: τ² within studies .054–.101, between studies .099–.148; Q(364) = 4446.

• Risk of bias. 43 of 182 studies low, 139 "some concerns", mostly for lack of preregistration. The rating did not moderate the effect (p = .104).

• Moderators (subgroup g; tests are meta-regression contrasts):

  Moderator	Result
  Item type	Standard statements .41; social-media headlines .17 (smaller, p = .038); claims (opinions, marketing) .47
  Task at first exposure	Truth judgement .10 [−.03, .23]; irrelevant task .42; passive reading .47 (judgement smaller than both, p < .001)
  Warnings that some items are false	No moderation, at exposure (.34 against .44) or at test (.38 against .37)
  Veracity cue at exposure	False cue −0.18; none .43; true cue .70 (false against none: b = −0.66)
  Type of false cue	Epistemic qualifier −0.57; source reliability −0.73; immediate feedback −0.91; delayed feedback −1.12; a "disputed" label .04 (no effect)
  Delay to test	None .46; within the day .33; up to a week .36; over a week .25. Only within-day against within-week differed (p = .023)
  Other	Verbatim .36 against gist .63; true .46, false .33, unverifiable .31; online .40 against face-to-face .31: none significant. Exposure over 5 s .48 against self-paced .34: not robust to imputation

• The moderators explain 37% of between-study variance (pseudo-R²).

• Cited, not read in the original:

  • Dechêne 2010: d .39 (fixed) to .50 (random); within-item .39 against between-item .53.
  • 96% of illusory-truth studies report significant results (Henderson et al. 2022).
  • The effect survives prior knowledge (Fazio 2015), accuracy incentives (Speckmann and Unkelbach 2022) and differences in cognitive ability (De keersmaecker 2020).

Limits

• One repetition only. Multiple exposures are excluded by design, so the dose–response is not estimated.
• Stimuli and samples. Trivia statements dominate; the samples are WEIRD.
• PEESE is unreliable under high heterogeneity (the authors).
• The delay moderation compares studies, not conditions.
• Inconsistencies found:
  • The abstract dates the previous meta-analysis to 2006. Dechêne et al. was published in 2010 and covers studies up to 2006.
  • CV moderator: "b = 0.53, 95% CI [0.39, 0.58], t(330) = −4.30" is described as a negative relation. The sign of b contradicts t and the text.
  • Delay: "b = −0.20, 95% CI [−0.37, 0.03], p = .023" has a CI that includes zero. "b = 0.12, 95% CI [−0.26, 0.01]" has b outside its own CI.
  • Testing environment: the counts are swapped between Results (online 196, face-to-face 169) and Methods (online 169, face-to-face 196).
  • The prior systematic map: 93 articles in Results, 89 in Methods.

What it means for Kurisutina (inference)

• One repetition of a statement, with no argument, raises its rated truth by g ≈ 0.37, for true and false statements alike.
  • A claim that interlocutors keep repeating is partly a repetition.
  • So some of what M1's "repeated pressure" carries over in people may be familiarity, not social pressure.
• Evaluating a claim at first encounter almost removes the effect (g .10), and a clear falsity tag reverses it. For B3, an entry stored as "X claimed this; I judged it unsupported" should gain little or nothing from later repetition, while an unevaluated claim should gain a little.
• Battery item. Repeat ambiguous statements across sessions and rate their truth. Human reference: g ≈ 0.37, with this moderator pattern as structure to reproduce.