Source
Weatherall, O’Connor and Bruner (2020), BJPS 71(4), LIT-1341, read in full in NOTE-1038 from the arXiv manuscript of August 2018: §3 and its Figs 2–4.
What was actually shown
K scientists on a cycle or complete network choose between action A (success .5, known) and B (.5 + ε, believed .5 ± ε), play B when their credence in it exceeds .5, run n trials a round and update by Bayes on their own and their neighbours’ results. Policy makers, initially sceptical (credence uniform on (0, .5)), listen to k scientists and gather nothing. The propagandist forwards to every policy maker each round’s results that favour A. Runs end at a scientific consensus; those ending in the wrong one are discarded.
The paper’s finding: “while the community of scientists converges on true beliefs about the world, the policy makers reach near certainty in the false belief”, over many parameter settings, with the effect weaker for larger k (Fig. 2), larger ε (Fig. 3) and larger n (Fig. 4). Each comparison could have come out the other way: an audience that hears every real result might have tracked the community, and larger studies might have yielded more ammunition rather than less.
Re-run here, from the paper’s description, 100 runs a setting unless stated (policy makers’ mean final credence in the true theory, B):
| setting | no propagandist | propagandist |
|---|---|---|
| cycle, K = 20, n = 10, p_B = .55, k = 1 | 0.873 | 0.000 |
| same, k = 5 | 1.000 | 0.000 |
| same, k = 10 | 1.000 | 0.039 |
| same, k = 20 | 1.000 | 0.938 |
| complete, K = 10, k = 9, p_B = .55, n = 5 (60 runs) | 0.980 | 0.247 |
| same, n = 20 | 0.988 | 0.755 |
| same, n = 100 | 0.995 | 0.989 |
| cycle, K = 20, n = 10, k = 10, p_B = .51 / .6 / .7 (40 runs, with propagandist) | 0.000 / 0.981 / 1.000 |
The re-implementation’s own choices, not fixed by the paper: each policy maker hears k scientists drawn at random once per run; scientists update on each neighbour’s result in turn within a round; there are ten policy makers; a run is capped at 20,000 rounds. Every trend has the paper’s direction. The values are this re-implementation’s, not the paper’s, whose plotted values were not seen.
What this does not say
- Not that real audiences were misled this way. The authors: “modeling alone cannot establish that this did happen in the case of tobacco industry propaganda efforts.”
- Not that a Bayesian who knows of the selection is misled. The policy makers treat forwarded results as unselected draws. A receiver who models the sender is THEORY-404’s case, a different result.
- Not that the effect holds whatever the parameters. It fades when the problem is easy, studies are large or the audience hears most scientists.
- Not about influence on the scientists. The propagandist never touches them; that is Holman and Bruner’s setting (LIT-744, LIT-751).
- Not that more scientists always help or always hurt. With k fixed, more scientists supply more spurious results, “often” but “not always” worsening the audience’s belief.
Connections
- CLAIM-404. Its evidence route, with genuine results selected and no fabrication, in a model; distinct from THEORY-404’s committed design and THEORY-146’s selective funding.
- THEORY-845. The same model’s biased-production treatment, and the sufficient statistic that explains both.
- THEORY-143. The Zollman effect concerns the community’s learning; with an audience and a propagandist, slow consensus can cost the audience (NOTE-1038).
- QUESTION-047: answers in part. The audience settles on a belief none of the inquirers reasoned to, through selection of their honest outputs.