NUWA Risk Discovery · Study
AI systems out-persuade expert humans
NUWA research note · independent review pending
NUWA Research Note
Research questionDesign and assumptionsSource findingsNUWA interpretationLimits and contrary evidenceResearch and decision relevanceEvidence updateInstitutions and rolesOriginal resourcesModel-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.
At a glance
Expert-human comparisons and consequential donation outcomes strengthen evidence while retaining experimental limits.
Research question
Can conversational AI change consequential behaviour relative to expert humans?
Design and assumptions
Four preregistered experiments cover 18,978 conversations and 6,923 people. Study 4 compares AI with professional canvassers, measuring actual allocation of a £1 experimental bonus.
Source findings
In Study 4, the AI-minus-professional contrast in donation allocation was 10.8 percentage points of the £1 bonus, with a 95% confidence interval of 6.4–15.1 points.
NUWA interpretation
An actual allocation is closer to behaviour than an attitude scale, while retaining a specific experimental and incentive boundary. The expert-human contrast does not establish effects on real political mobilisation.
Limits and contrary evidence
The effect is measured in percentage points of the experimental bonus, not national fundraising or the probability of harmful action. Conversations and people have different denominators.
Research and decision relevance
NUWA recommendation: Compare the endpoint, human comparator and persistence with other persuasion studies before generalising to deployment.
Precise source locations
- AI systems out-persuade expert humans · Study 4; Fig.4; Methods; text lines 00275Original resource captured 2026-10-28T07:24:42.308375+00:00; sha256:991f9cf156212429c7ddb6ce60baeaae7b46b8ac18de2c80ab7f00ea02a7e276
Source checking, extraction, editorial approval and replication are recorded separately. Methodology →
Evidence update · 2026-10-28
Model-assisted reading · independent evidence and bilingual editorial review pending. Values are source findings or inventory counts, separate from the trend result series.
Source-bound observations (3)
Counts, rates and contrasts have different units. A sample count is not a number of independent experiments. Unknown numerators and interval levels remain unknown.
| System / metric | Reported value | Population and conditions |
|---|---|---|
| Study conversations | 18978 conversations | Conditions, limitations and sourceOriginal source ↗ · Abstract; study overview; text lines 00153,00159,02183Original capture 2026-10-28T07:24:42.308375+00:00; sha256:991f9cf156212429c7ddb6ce60baeaae7b46b8ac18de2c80ab7f00ea02a7e276 |
| Unique people | 6923 people | Conditions, limitations and sourceOriginal source ↗ · Abstract; study overview; text lines 00153,00159,00314,02184Original capture 2026-10-28T07:24:42.308375+00:00; sha256:991f9cf156212429c7ddb6ce60baeaae7b46b8ac18de2c80ab7f00ea02a7e276 |
| AI minus professional canvasser donation effect | 10.8 percentage-points-of-GBP1-bonus6.4–15.1 · confidence · 95% | Conditions, limitations and source{
"outcome": "actual donation from experiment bonus",
"contrast": "AI vs professional canvasser",
"bonusGBP": 1
}Not percent of national fundraising; not a rate of taking any harmful action. Original source ↗ · Study4 Fig.4; text lines 00275Original capture 2026-10-28T07:24:42.308375+00:00; sha256:991f9cf156212429c7ddb6ce60baeaae7b46b8ac18de2c80ab7f00ea02a7e276 |
Institutions and roles
Authors (6)
Kobi Hackenburg · Caroline Wagner · Luke Hewitt · Ben M. Tappin · Ed Saunders, Hannah Rose Kirk · Helen Margetts, Christopher Summerfield
Cite and reuse
Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.
Original publication · BibTeX
@misc{expert-persuasion,
title = {AI systems out-persuade expert humans},
url = {https://arxiv.org/html/2606.16475},
author = {Kobi Hackenburg and Caroline Wagner and Luke Hewitt and Ben M. Tappin and Ed Saunders, Hannah Rose Kirk and Helen Margetts, Christopher Summerfield},
year = {2026},
date = {2026-06-15},
version = {arXiv:2606.16475v1; version header 15 Jun 2026; Original resource captured 2026-10-28T07:24:42.308375+00:00; sha256:991f9cf156212429c7ddb6ce60baeaae7b46b8ac18de2c80ab7f00ea02a7e276},
note = {arXiv:2606.16475v1; version header 15 Jun 2026; Original resource captured 2026-10-28T07:24:42.308375+00:00; sha256:991f9cf156212429c7ddb6ce60baeaae7b46b8ac18de2c80ab7f00ea02a7e276}
}
NUWA record · BibTeX
@misc{expert-persuasion-nuwa,
title = {AI systems out-persuade expert humans — NUWA research record},
url = {https://whitzard.tech/en/nuwa/whitzard-index/work/expert-persuasion/},
version = {2026-10-29.1},
organization = {NUWA Frontier AI Safety Lab},
howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--expert-persuasion.json},
note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--expert-persuasion.json.}
}
Evidence snapshot 2026-10-29.1 · Index updates · RSS
References
- [1] Kobi Hackenburg, Caroline Wagner, et al.. AI systems out-persuade expert humans. 2026-06-15 ; Study 4; Fig.4; Methods; text lines 00275