Found on 2026-10-10 by a search the owner asked for: transmission-chain studies that measure a stance at each generation, the literature CLAIM-234 found absent from the record.
The case
Moussaïd, Brighton and Gaissmaier (PNAS 2015). Fifteen chains, of 7 to 12 people, analysed to the tenth position. The first person read six media articles on the antibacterial agent triclosan, then discussed it face to face with the next person, “in open, unstructured discussion”, who then discussed it with the next. Participants waited in one room and were told not to interact. It is, the authors say, the paradigm “pioneered by Frederic Bartlett” (LIT-768).
The transcripts were coded twice, on two schemes that the paper calls independent of each other.
- Content. 61 units of information, each tracked from position to position and marked “distorted” when it changed.
- Signal. Every sentence of the informed speaker labelled Positive, Negative or Neutral by the assessment of triclosan it expresses: “Sentences that highlight the suspected dangers of Triclosan or express a negative judgement about it received a ‘Negative’ label”. Two coders with the same instructions; “The results were then averaged.” No agreement coefficient is reported.
The results, in the authors’ words (arXiv preprint): messages “tend to become shorter, gradually inaccurate, and increasingly dissimilar between chains. In contrast, the perception of risk is propagated with higher fidelity” (abstract). Positive statements decayed faster than negative ones, so the negative share rose down the chain. Speakers differed: the direction in which a speaker changed the signal correlated with that speaker’s risk perception (c = 0.25, p = 0.019), so “the direction of the ‘mutation’ of the message corresponds to the opinion of the speaker” (p. 7). A receiver’s change in risk perception tracked the signal received (c = 0.55). The discussion: “changes in risk perception occur less as a result of the message content, and more as a result of its overarching, subjective signal” (p. 10).
Jagiello and Hills (Risk Analysis 2018). Fourteen written chains of eight. The first person read four documents balanced by construction, 21 positive and 21 negative statements, on nuclear energy (high dread) or food additives (low dread). Each later person read only the previous message and wrote their own. At position 6 the chain forked: one branch saw the original documents again, one did not. Each participant’s knowledge and perceived risk were measured by questionnaire before reading. All 154 messages were coded for negative and positive statements twice, by the experimenter and by five confederates blind to the hypothesis: “Cronbach’s alpha indicates high interrater agreement for both negative (α = 0.961) and positive statements (α = 0.974)” (§2.2).
- “the more a message was transmitted the more negative statements it contained. This was highest for the high dread topic” (abstract). Chain position had a main effect on the proportion of negative statements, F(10, 154) = 5.506; the low-dread chains “were not significantly affected by chain position” (§3.1).
- Re-exposure “did not restore negativity to its level in position 1” in the high-dread chains (§3.1.2).
- Across both conditions, a person’s output negativity was predicted by the negativity received (β = 0.565), by their initial risk perception (β = 0.188, p = 0.005) and, negatively, by their knowledge (Table I). Within either condition alone, initial risk was not significant.
What it can show
For CLAIM-068: drift along a stance dimension, measured. The claim quotes A50’s “communicative drift along a stance dimension, not merely lexical degradation”. Here the evaluative signal and the content were coded separately, at every link, by different retellers, and they behaved differently: the content decayed and diverged while the evaluation was kept and pushed one way. Jagiello and Hills add a start fixed by construction (balanced), a drift away from it, and a failure of the source to pull the chain back. The drift depends on the topic, as C1 Appendix D’s qualification expects: the endpoint is “a property of the specified transmission mechanism”.
For CLAIM-090: filtering by measured biases. The claim’s defeat condition asks for a chain whose output tracks the ancestor “rather than any measured prior of the transmitters”. Here the transmitters’ risk perception was measured on them, not assumed, and it predicts the direction of the change, in both studies. Kalish et al. (LIT-820) took their prior from earlier studies.
For CLAIM-234. One item of the absent literature is now documented. The objection’s list of what the record lacks narrows; see the note there.
What it cannot show
Footing, or any interpersonal stance. The stance coded is evaluation of a hazard, the evaluation vertex of Du Bois’s triangle (LIT-779). Its object is a chemical or a power source, not the addressee. Nothing in either study codes who stands behind a statement, whether it teases or reproaches, or alignment between speakers. The closest work the search found on footing, Altay, Claidière and Mercier’s study of source attribution in rumours (Evolutionary Human Sciences 2020, DOI 10.1017/ehs.2020.53), is a single transmission step, with the chain simulated. A52’s extension of iterated learning “to speaker footing” still has no study under it.
Convergence from different starts. CLAIM-068’s defeat condition asks whether chains “started from different stances” converge to a few conventional ones. Every Jagiello and Hills chain started balanced, and the Moussaïd chains all started from one set of articles. The chains are 8 and 10 long and are not stationary. Moussaïd’s simulations end “strongly positive or strongly negative” depending on the group’s initial perception, which reads as several attractors, as in Kalish et al.; but those are simulations, not data.
A clean prior. CLAIM-090 is met only partly. In Moussaïd the correlation with the direction of change uses the risk perception the subjects reported “in questionnaires conducted after the experiment”, and, since nobody knew triclosan, it was measured indirectly, on “chemical use in food safety and cosmetics” generally. A post-experiment measure is contaminated by the message received. Jagiello and Hills measured risk before reading, which is cleaner, but the received message predicts output three times as strongly, and the prior is significant only when the two topics are pooled.
Reliability. Moussaïd reports no agreement statistic for the signal coding. Jagiello and Hills report Cronbach’s α between two passes on counts, which is lenient for coding.
The defender’s reply. A defender of the objection, that no reading measures stance under transmission, would say this is content selection under another name. A “Negative” sentence is one that “highlight[s] the suspected dangers”, and its count is a count of which propositions about harm survived. The finding is then a negativity bias in recall, the kind of content bias Mesoudi and Stubbersfield study, not a stance kept apart from content. The reply has force: the two coding schemes are independent as schemes, but the signal is computed over statements whose content is evaluative. It does not reach the whole result. The signal was kept while the information units carrying it changed and diverged between chains, and Moussaïd’s receivers changed their own risk perception with the signal rather than with the content. But the case cannot separate an evaluative stance from the valence of retained content.
What would close the gap. A chain study of footing. Seed chains with one utterance under different attributed speakers, such as CASE-018’s sentence as a friend’s tease and as a judge’s condemnation, and with different starting stances, so that convergence can be tested. Have each generation’s retelling coded by blind raters for footing and alignment, on a definition and response format validated first, which is Stage 0 of the discriminating study, CASE-055, and the question QUESTION-076 asks. The intralingual retelling arm of CASE-055 is the natural place to run it, and these risk-message designs give its procedure.
Sources
Every URL below was retrieved on 2026-10-10. None of these works is held in this record or, on a check of the anthology clone, in the anthology. This record is their likely home; none is filed here.
- M. Moussaïd, H. Brighton and W. Gaissmaier, “The amplification of risk in experimental diffusion chains”, PNAS 112(18) (2015), 5631–5636, DOI 10.1073/pnas.1421883112; preprint arXiv:1504.05331. The preprint was read directly and in full; page numbers are the preprint’s. The PNAS text was not seen, so the published wording is unverified.
- R. D. Jagiello and T. T. Hills, “Bad news has wings: Dread risk mediates social amplification in risk communication”, Risk Analysis 38(10) (2018), 2193–2207, DOI 10.1111/risa.13117. Read directly and in full in the author copy, https://warwick.ac.uk/fac/sci/psych/people/thills/thills/2018jagiellohills.pdf.
- Y. Altay, N. Claidière and H. Mercier, “It happened to a friend of a friend: inaccurate source reporting in rumour diffusion”, Evolutionary Human Sciences 2 (2020), e49, DOI 10.1017/ehs.2020.53. Read only through the research report’s summary. Unverified here.
- A. Mesoudi, A. Whiten and R. Dunbar (2006), British Journal of Psychology 97(3), 405–423, DOI 10.1348/000712605X85871, and J. M. Stubbersfield, J. J. Tehrani and E. G. Flynn (2015), British Journal of Psychology 106(2), 288–307, DOI 10.1111/bjop.12073: content-bias chain studies named in the defender’s reply. Not read; known only through the research report, which found that they measure recall, not stance.
- J. W. Du Bois, “The stance triangle” (2007), held as LIT-779.
- F. C. Bartlett, Remembering (1932), held as LIT-768 (Deferred, unread).