The influence of communication modality on the "saying-is-believing" effect

Behavioral Sciences 15(5): 639 (May 2025; MDPI), DOI 10.3390/bs15050639, PMC12109010, open access (CC BY 4.0). A correction was published later (Behav Sci 15(11): 1446, PMC12649374), which was not read. Tianjin Normal University. Provenance: papers/carry_on/yin2025_saying_is_believing.provenance.json.

Why it is here. The core saying-is-believing papers could not be obtained openly: Higgins & Rholes 1978; Echterhoff, Higgins & Groll 2005; Echterhoff et al. 2008; Wagner et al. 2024. This is an open replication of their paradigm, used for the human baseline of drift toward an interlocutor. It is a weak paper (see Limits); rely on the direction of the basic effect, not on its modality claims.

What was read

All 591 lines of the text derived from the PMC XML: abstract, all sections, appendices A–D (materials, IOS, SR-G and SR-T items), Tables 1–6 and references. The converter checks that the text keeps every non-whitespace character of the XML. Figures 1–6 are images. The correction notice's content was not fetched.

Question

Does the saying-is-believing (SIB) effect occur? People who describe a person to an audience tune the description to the audience's attitude, and then remember the original information in the tuned direction. Does its size differ between typed online and face-to-face communication?

Method

• Participants. 80 Chinese students; 10 excluded for suspicion, leaving 70. The audience attitude groups are given as positive 34 and negative 36 in the text, but negative 35 in Table 1.
• Classic paradigm (materials adapted from Echterhoff et al. 2005/2008).
  • Participants read seven ambiguous behaviour descriptions of a target person (e.g. persistent or stubborn; frugal or miserly).
  • They were told "casually" that the listener, a confederate, liked or disliked the target.
  • They described the target so the listener could identify it among 15 people.
  • After a 3-minute filler task they recalled the original text as accurately as possible.
• Design. Audience attitude between subjects × modality (typed online; face-to-face at 50 cm) within subjects. Face-to-face listeners could nod or frown in line with their attitude, and participants expected to meet and rate each other afterwards. Online participants were told there would be no further contact.
• Coding. Two blind coders rated each description and recall unit from −5 to +5 (r = .986 and .976). Participants also completed shared reality (SR-G, SR-T) and IOS closeness questionnaires.

Results

• Audience tuning of descriptions. F(1,68) = 127.95 for audience attitude. Descriptions for a liking audience were positive and for a disliking one negative: online 1.24 against −1.40; face-to-face 2.15 against −0.60.
• Memory follows the tuning. F(1,68) = 526.58 for audience attitude on the valence of the recalled original text: online 1.70 against −2.10; face-to-face 3.17 against −0.59, on a −5 to +5 scale. People remembered the same ambiguous facts as better or worse depending on whom they had described them to.
• Modality.
  • Both descriptions and recall were more negative online.
  • The interaction on absolute deviation: face-to-face amplifies a positive tuning, online amplifies a negative one. The authors attribute this to "deindividuation" through visual anonymity.
• Shared reality. Shared reality and closeness did not mediate the modality effect.

Limits

• Modality is confounded.
  • The procedure puts the online description in Phase 1 and face-to-face in Phase 2. Materials were counterbalanced; modality order was not (my reading of the procedure).
  • Face-to-face also added live non-verbal feedback matching the attitude, and an expected later meeting.
  • The modality effect cannot be attributed to anonymity.
• No no-audience or neutral-audience control, so "bias" is the difference between the two audience groups.
• Small single-site student sample; the data are not shared.
• Inconsistencies found:
  • The negative group is n = 36 in the text and n = 35 in Table 1 (34 + 35 = 69, not 70).
  • The IOS mediation estimate (0.058) lies outside its own 95% CI [−0.007, 0.025].
  • The recall-deviation "SDs" (0.12–0.17) are about ten times smaller than the valence SDs (0.73–1.10) and look like standard errors.
  • "The Pearson correlation … reached 90% (r = 0.986)".
  • SIB is once expanded as "social identity bias".
  • SR-T items 3 and 4 are identical.
  • The consent statement mentions "the patient(s)".
  • Arithmetic checked: the pooled means in the text match Tables 3–4 when the negative group has 36 (e.g. online description (34 × 1.242 + 36 × −1.400) / 70 = −0.117 against the stated −0.12).

What it means for Kurisutina

• Question 2: some drift toward the interlocutor is human.
  • People who describe something to an audience with a known view shift both what they say and what they later remember, in the audience's direction.
  • So a replica that drifts toward whoever it talks to is not by itself un-person-like. The test is how much, compared with how much this person would shift.
  • This refines the drift criterion from Venkit, Mooney and Abdulhai. The null is not "zero movement toward the interlocutor" but the person's own susceptibility, which could be measured with this paradigm: ambiguous targets, audience attitude, delayed recall.
• Question 3: who the replica has talked to becomes part of its memory. For a person, retelling an ambiguous episode to a partial listener re-colours the episode itself. A replica's memories of events it has discussed should be logged together with its conversational partners (the brief's rule to log cues). A replica that never re-colours may be as un-person-like as one that re-colours too much (inferred).
• Effect size caution. Wagner, Axmacher & Echterhoff (2021/2024; summaries/carry_on/wagner2024_shared_reality_retrieval.md) found recall shifts of only 0.1–0.6 points on the same −5 to +5 scale in three larger online studies, and none in one of them; the 3–4 point shifts here are implausible. Use this paper for the direction of the effect only.
• Next reading, if obtainable. Echterhoff, Higgins & Groll (2005) and Wagner et al. (2024) are the primary sources. They show that the memory effect depends on creating shared reality, i.e. on trusting the audience, which may give the person-level moderator a replica needs.

Cross-references

• summaries/carry_on/mooney2025_behavioral_coherence.md, summaries/carry_on/abdulhai2025_persona_consistency.md, summaries/carry_on/venkit2026_companion_drift.md: drift toward the interlocutor in LLM agents.
• summaries/carry_on/stjacques2013_reactivation.md: post-event information changing memory.
• summaries/carry_on/hirst2009_911_memory.md: social retelling and correction.
• summaries/carry_on/acerbi2023_llm_transmission.md: ambiguity resolution defaults in LLMs.