Kurisutina
Library/Carrying on

Wagner et al.

The role of cognitive accessibility

Published in Journal of Experimental Psychology: General 153(6): 1605–1627 (2024), DOI 10.1037/xge0001580, closed access; the published version was not read. Read as the PsyArXiv preprint v1 (n94aj, 26 November 2021; no licence stated). The published paper may differ, for instance in added studies or revised analyses. University of Münster and Ruhr University Bochum. Provenance: papers/carry_on/wagner2024_shared_reality_retrieval.provenance.json.

What was read

All 2,847 lines of pdftotext -layout output (56 pages): abstract, introduction, three studies, general discussion, references, Tables 1–12 and Figure 1's caption and axis labels (the plotted values are in the tables). No supplement. The preregistrations (AsPredicted, Studies 2–3) were not read.

Question

The saying-is-believing (SIB) effect: people tune a description of a person to their audience's attitude and later recall the original information in that direction. Is this because audience-congruent information becomes more accessible in memory (the ROAR model), rather than because people filter at recall or recall their own message?

Method

  • Three online studies (German students and others): N = 165, 289 and 347.
  • Procedure. Participants read an ambiguous text about "Michael" and learn that their online partner "Jan" (a fiction) likes or dislikes him. They write Jan a description, then take an accessibility test, then freely recall the original text. Questionnaires follow: shared reality about the target (SR-T), epistemic trust, relational motivation and IOS.
  • Accessibility.
    • Reaction time to judge whether trait items (e.g. independent or obstinate) and behaviour items apply to Michael.
    • Study 1 also used a scrambled-sentences choice task.
  • Study 2. Commonality manipulations: a joint picture task with "93%" agreement; feedback that Jan identified Michael; both. Half the participants skipped the reaction-time task.
  • Study 3. Pictures, Feedback or Standard, and half communicated about Joe Biden instead of Michael.
  • Coding. Two blind raters scored message and recall valence from −5 to +5 (r = .81–.96).

Results

  • Messages are strongly tuned.

    • Study 1: 0.66 against −1.04 (t(163) = 5.46).
    • Study 2: 0.92 against −0.84 (ηp² = .228).
    • Study 3: 0.83 against −0.98 (ηp² = .223).

  • Memory shifts only a little, and not always.

    • Study 1: no recall effect, 0.20 against 0.09 (p = .47).
    • Study 2: 0.23 against −0.02 (ηp² = .016); significant only without the preceding reaction-time task (0.38 against 0.01).
    • Study 3: 0.25 against −0.31 for target-related communication (ηp² = .068); n.s. when talking about Biden.

  • Accessibility follows the audience at the trait level. People judged audience-congruent trait items faster:

    • Study 1: difference −91 against +84 ms (F = 9.39);
    • Study 2: −203 against +44;
    • Study 3: −53 against +82 (target-related).

    Behaviour items mostly showed no effect (except in Study 3).

  • Links. The trait accessibility effect correlates with the recall bias only weakly (r = .18, .09 n.s., and .18). It correlates with shared reality about the target (r = .24–.26) and epistemic trust (.18–.20), not consistently with relational motivation or closeness.

  • Commonality manipulations did little: none in Study 3; mixed in Study 2, where the combined manipulation numerically reversed the recall effect.

  • Authors' conclusion. Shared reality with a trusted audience makes congruent trait impressions more accessible. People can then produce biased memories "quickly and confidently".

Limits

  • Online studies with a fictitious partner. There is no no-audience control; the bias is the difference between the two audience groups. The recall effects are small, and the reaction-time task itself dampened them.
  • Correlational links between accessibility, recall and questionnaires are weak (r < .2 for recall).
  • Inconsistencies found:
    • Study 2's valence main effect is given as "F(1, 138) = 15.31, p = .023, ηp2 = .037". F = 15.31 on (1, 138) gives p < .001 and ηp² ≈ .10 (my computation).
    • Study 3 reports the Biden recall effect as "F(1, 167) = 2.87, p = .092, ηp2 = .017", identical to the Biden reaction-time statistics; Table 8 gives p = .29 for recall.
    • "All ps < .20" and "p< .28" where ">" is meant.
    • Study 3's target-related subset is tested with df (1, 335), the full-sample df.
    • Table 12's "Behavior – negative Items" and "Behavior – Diff." rows are swapped (raw times appear under Diff.).
    • Figure 1 labels Study 2 N = 289, but only about 144 did the reaction-time task (Table 6).
    • Checked: F = 9.39 on (1, 161) matches the stated ηp² = .055.

What it means for Kurisutina

  • Question 2: humans tune what they say a lot, but what they remember only a little.
    • Audience tuning of messages is large (ηp² about .22). The later shift in memory of the original is small (ηp² .02–.07; about 0.1–0.6 points on a −5 to +5 scale) and fragile (absent in Study 1, damped by a filler task).
    • For a replica, this separates two things to measure:
      • expressed accommodation to an interlocutor, which people also show strongly;
      • carried-over change in its later, private answers and memories, which in people is small.
    • A replica whose private answers after a conversation shift as much as its tuned statements is drifting more than a person would (inferred).
    • The test: ask the same questions privately before and after interlocutor exposure. Also ask in the interlocutor's presence. Compare both shifts with the person's own.
  • The effect scales with epistemic trust in the audience, and it enters through trait impressions, not stored behaviours. A replica's memory of other people held as trait summaries (a common agent-memory design) is where interlocutor influence would land. Episode-level records resist it more (inferred from the trait/behaviour asymmetry).
  • A calibration for Yin & Liu (2025). Their recall shifts of 3–4 points on the same scale are far outside these careful studies' 0.1–0.6. That is one more reason to treat that paper's effect sizes as unreliable (see its summary).

Cross-references

  • summaries/carry_on/yin2025_saying_is_believing.md: the open replication with implausible effect sizes.
  • summaries/carry_on/mooney2025_behavioral_coherence.md, summaries/carry_on/venkit2026_companion_drift.md, summaries/carry_on/abdulhai2025_persona_consistency.md: interlocutor drift in LLM agents.
  • summaries/carry_on/stjacques2013_reactivation.md, summaries/carry_on/hirst2009_911_memory.md: memory updating and social correction.
  • summaries/carry_on/park2023_generative_agents.md: agent memory with reflections (trait-like summaries).

This summary is our record of the paper, written after reading the full text and published as written; links into our own repository have been removed.