CoChat

How persuasive is AI-generated propaganda?

Open access

The DOI matches its Crossref record; 28 of the 28 references it cites were found in a registry or an index; 1 of the 3 sources its claims cite were checked for support.

Abstract

Can large language models, a form of artificial intelligence (AI), generate persuasive propaganda? We conducted a preregistered survey experiment of US respondents to investigate the persuasiveness of news articles written by foreign propagandists compared to content generated by GPT-3 davinci (a large language model). We found that GPT-3 can create highly persuasive text as measured by participants’ agreement with propaganda theses. We further investigated whether a person fluent in English could improve propaganda persuasiveness. Editing the prompt fed to GPT-3 and/or curating GPT-3’s output made GPT-3 even more persuasive, and, under certain conditions, as persuasive as the original propaganda. Our findings suggest that propagandists could use AI to create convincing content with limited effort.

Overview

CoChat Summary

TL;DR

A preregistered survey experiment with 8,221 US respondents found that GPT-3 davinci could generate propaganda articles that were highly persuasive, though slightly less persuasive than real-world foreign covert propaganda on average. Human editing of prompts and curation of outputs made GPT-3-generated propaganda as persuasive as, or more persuasive than, the original propaganda.

Key findings

  • Original propaganda was highly persuasive: agreement with thesis statements rose from 24.4% in the control to 47.4% among those who read original propaganda, a 23 percentage point increase.
  • GPT-3-generated propaganda was also highly persuasive: agreement was 43.5% versus 24.4% in the control, a 19.1 percentage point increase, but 3.9 percentage points lower than original propaganda on average.
  • Human curation helped: after removing two of 18 GPT-3 outputs that did not advance the intended thesis, agreement rose to 45.6%, and the difference from original propaganda was no longer statistically significant.
  • Editing prompts and selecting the best of three outputs produced even stronger effects: agreement was 52.7% compared with 47.4% for original propaganda; edited-prompt-only articles reached 46.4%, not statistically different from 47.4%.

Methods

  • Preregistered survey experiment in December 2021 with US adults recruited via Lucid using quota sampling; final sample was 8,221 after attention-check failures and completion times under 3 minutes were excluded.
  • Six English-language foreign covert propaganda articles (151–308 words) from real-world campaigns originating from Iran or Russia were compared with GPT-3 davinci outputs on the same topics; GPT-3 generated three articles per topic, with outputs under 686 or over 1,936 characters discarded.
  • Respondents rated agreement with thesis statements for four randomly selected topics without reading an article (control), then read articles on the remaining two topics; analyses regressed percent agreement and scaled agreement on issue and article indicators.

Limitations

  • The paper states its estimates may represent a lower bound because large language models are rapidly improving and because the experiment tested exposure to a single article, whereas propagandists could expose citizens to many articles.
  • The paper states future research could test AI-generated propaganda across a wider range of issues and with information from multiple sources.
  • Evident from the text: the study uses US respondents, six topics, and GPT-3 davinci in December 2021, so findings may not generalize to other populations, topics, models, or later AI systems.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/pnasnexus/pgae034

Cites

28 of 28 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1. Original propaganda and GPT-3-generated propaganda were highly persuasive. The top panel shows the percentage of respondents who agreed or strongly agreed with the thesis statement. The bottom panel shows the average level of agreement on a 5-point scale, coded 0 if strongly disagree, 25 if disagree, 50 if neither agree nor disagree, 75 if agree, and 100 if strongly agree. Estimates are pooled across topics and outputs. SEs clustered by respondent and 95% CIs are shown.
Fig. 1. Original propaganda and GPT-3-generated propaganda were highly persuasive. The top panel shows the percentage of respondents who agreed or strongly agreed with the thesis statement. The bottom panel shows the average level of agreement on a 5-point scale, coded 0 if strongly disagree, 25 if disagree, 50 if neither agree nor disagree, 75 if agree, and 100 if strongly agree. Estimates are pooled across topics and outputs. SEs clustered by respondent and 95% CIs are shown.CC BY · PubMed Central
Fig. 2. Most GPT-3-generated output were as persuasive as the original propaganda, but a few articles performed worse. Average agreement with the thesis statement for each treatment group for each topic. SEs clustered by respondent and 95% CIs are shown.
Fig. 2. Most GPT-3-generated output were as persuasive as the original propaganda, but a few articles performed worse. Average agreement with the thesis statement for each treatment group for each topic. SEs clustered by respondent and 95% CIs are shown.CC BY · PubMed Central
Fig. 3. Human curation made GPT-3 as persuasive as the original propaganda. Average agreement with the thesis statement for each treatment group, averaged over topics. “All outputs (no curation)” includes all GPT-3 propaganda articles. “Outputs that make thesis” excludes GPT-3 outputs that did not make the claim of the thesis in the title or body of the article. “Best performing outputs” is the average agreement with the thesis statement for the best performing GPT-3 output for each of the six topics. SEs and 95% CIs are shown.
Fig. 3. Human curation made GPT-3 as persuasive as the original propaganda. Average agreement with the thesis statement for each treatment group, averaged over topics. “All outputs (no curation)” includes all GPT-3 propaganda articles. “Outputs that make thesis” excludes GPT-3 outputs that did not make the claim of the thesis in the title or body of the article. “Best performing outputs” is the average agreement with the thesis statement for the best performing GPT-3 output for each of the six topics. SEs and 95% CIs are shown.CC BY · PubMed Central

Verification

Claims the paper makes

1 Valid · 1 Not valid · 2 Not checked · 29 Own results and interpretations

11 more claims

Create a free account to see all 33 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 11
  • Probably not a substantial claim. 12

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 28 references

Cited by

Items 101–126 of 126

Metadata

Authors
Goldstein, J. A., Chao, J., Grossman, S., Stamos, A., & Tomz, M.
Affiliations
  • Georgetown University (US)
  • Stanford University (US)
  • Stanford Institute for Economic Policy Research (US)
Corresponding author
Grossman, Shelby
Venue
PNAS Nexus
Published
2024-02-01
Volume
3
Issue
2
Pages
pgae034-pgae034
ISSN
2752-6542 (ISSN-L)
OpenAlex
W4391985880
Cited by
133
References
18
Citations by year
2026: 45 · 2025: 64 · 2024: 22
FWCI
93.55
Type
journal article
Language
en
Open access
Locations
  • PNAS Nexus (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
Stanford University (#216301)
Topics
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Social Sciences › Sociology and Political Science › Media Influence and Politics
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
Keywords
large language models, GPT-3, persuasion, automated content generation, disinformation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.