How persuasive is AI-generated propaganda?

Covert online propaganda campaigns by state-aligned actors are frequent and ongoing across websites, social media, messaging apps, and other channels. Concerns have grown that recent advances in large language models (LLMs) could supercharge such campaigns by enabling the low-cost mass production of persuasive text. While prior research has examined credibility judgments of AI-generated content and the detection of false information, little is known about the comparative persuasiveness of AI-generated propaganda relative to an ecologically valid benchmark from real-world campaigns. This study tests whether GPT-3 can generate propaganda that persuades US audiences compared to original foreign covert propaganda articles and assesses whether minimal human involvement (e.g., prompt editing, output curation) can further enhance persuasiveness.

Existing work documents the prevalence of online covert propaganda and disinformation operations, including high-profile Russian activities in 2016 and ongoing efforts by various states. Prior social science studies related to AI-generated text have evaluated whether people rate AI-written news as credible, their ability to recognize false AI-generated content, and whether officials respond to AI-written constituent messages. However, there is a gap regarding the direct measurement of persuasiveness of AI-generated propaganda compared to real-world propaganda. Broader technical and policy literatures have outlined risks posed by language models to information ecosystems and discussed potential human–machine teaming in influence operations.

Design: The authors conducted a preregistered online survey experiment in December 2021 with US adults recruited via Lucid using quota sampling to approximate geographic and demographic representativeness. Respondents failing initial attention checks were not invited to continue, and those completing the survey in under 3 minutes were excluded, yielding a final N = 8,221. IRB approval was obtained, and participants received a debrief noting the propaganda origin of the articles.

Stimuli: Six English-language articles (151–308 words) from real, likely state-aligned covert propaganda campaigns originating from Iran or Russia were selected (criteria detailed in SI Appendix 1.A). For each topic, the authors wrote a concise thesis statement summarizing the article’s main persuasive claim (e.g., on drones, Iran sanctions, Syria chemical attacks, Syria medical shortages, Syria oil, US-Mexico wall). These thesis statements also informed prompts to GPT-3.

AI generation: Using OpenAI’s GPT-3 davinci, the authors generated three articles per topic (18 total). Each prompt included: (i) one or two sentences from the original article that stated its main point, and (ii) three example propaganda articles on unrelated topics to guide style and structure. The team generated three outputs per topic to avoid idiosyncrasies of single generations. Outputs were discarded only if shorter than 686 characters or longer than 1,936 characters (limits chosen to be within 10% of observed length bounds of original/edited propaganda); no other filters were applied (full details in SI Appendix 1.B).

Assignment and measures: For each respondent, four of the six topics were randomly selected for baseline (no-article) measurement of agreement with the thesis statements, and the remaining two topics were randomly assigned to article exposure conditions: either the original propaganda article or a GPT-3-generated article (one Syria-related and one non-Syria-related). Persuasion was measured two ways: (1) percent agreement (share who agreed/strongly agreed with the thesis), and (2) scaled agreement (0–100 rescaled from a 5-point Likert). The analysis regressed agreement measures on indicators for issue and article, with estimates averaged equally across issues and equally across GPT-3 outputs. Standard errors were clustered by respondent; 95% confidence intervals were reported. Additional metrics assessed perceived credibility (trustworthiness; reporting facts vs. persuasion) and writing style (well-written; author’s first language English).

Human–machine teaming variants: The authors simulated two practical enhancements: (a) curation—removing GPT-3 outputs that did not make the thesis claim in the title or body (2 of 18 removed; process in SI Appendix 1.D), and (b) prompt editing—replacing excerpted main-point snippets with the researcher-written thesis statement and editing example articles for clarity and grammar to simulate fluent English prompting. They also evaluated a best-of-three selection strategy per topic after these steps. A preregistered arm tested GPT-3 fine-tuned on Washington Post Politics articles to emulate a mainstream news style; this reduced the rate at which outputs made the thesis (36%) and is analyzed in SI Appendix Section 3.

Overall persuasiveness:

• Control: 24.4% agreed or strongly agreed with the thesis when no article was shown.
• Original propaganda: 47.4% agreement (≈ +23 percentage points vs. control), nearly doubling agreement.
• GPT-3 propaganda: 43.5% agreement (≈ +19.1 pp vs. control), only 3.9 pp less than original on average; effects were consistent across demographic and attitudinal subgroups.

By topic (percent agreement; examples):

• Drones: Control 22.2; Original 39.1; GPT-3 outputs 40.1, 42.5, 39.7.
• Iran: Control 25.9; Original 44.3; GPT-3 outputs 43.6, 47.7, 42.9.
• Syria Chemical: Control 21.2; Original 40.5; GPT-3 outputs 37.9, 44.2, 35.9.
• Syria Medical: Control 34.2; Original 63.1; GPT-3 outputs 60.2, 61.6, 60.3.
• Syria Oil: Control 30.6; Original 48.3; GPT-3 outputs 42.5, 47.6, 47.3.
• Wall: Control 12.5; Original 49.3; GPT-3 outputs 36.5, 50.7, 38.6. A few GPT-3 articles performed significantly worse than original propaganda (e.g., Syria Oil output 3; Wall outputs 2 and 3), contributing to the modest average gap.

Human–machine teaming:

• Curation (remove 2/18 outputs that did not state the thesis): 45.6% agreement; the difference from original propaganda was no longer statistically significant.
• Edited prompt (provide researcher-written thesis and edited examples): 46.4% agreement; not statistically different from original.
• Edited prompt + best-of-three selection: 52.7% agreement, exceeding original (47.4%).

Scaled agreement (0–100) showed similar patterns:

• Original: 57.8; GPT-3: 55.0; GPT-3 curated: 57.0; GPT-3 best-of-three: 59.4; GPT-3 with edited prompt: 57.0; Edited prompt + best-of-three: 61.2.

Additional metrics:

• On perceived credibility and writing style (trustworthiness; facts vs. persuasion; well-written; author’s first language English), GPT-3 performed as well as or better than original propaganda. Only 38.7% found the original propaganda trustworthy, and 52.4% thought it was well written.

Fine-tuning note:

• A preregistered fine-tuned GPT-3 (on Washington Post Politics) produced outputs that made the thesis only 36% of the time, likely prioritizing source style/content; details in SI Appendix. This approach under these settings did not improve persuasiveness.

The study directly addresses whether LLMs can generate persuasive propaganda by benchmarking GPT-3 against authentic, foreign covert propaganda. The results indicate that GPT-3 can produce highly persuasive articles that significantly shift agreement toward propaganda theses, approaching the impact of original propaganda. Minimal human oversight—curating off-target outputs and improving prompt quality—closes the residual gap, and combining prompt editing with best-of-three selection can surpass the persuasiveness of the original articles. Effects generalize across demographics and political/behavioral subgroups, suggesting broad susceptibility.

These findings underscore that AI can lower barriers to producing convincing propaganda: propagandists can generate numerous stylistically varied articles cheaply and quickly, potentially enhancing volume, reach, and evasion of detection. GPT-3’s comparable ratings on credibility and writing style further suggest that AI-generated content can blend into information environments as well as material from known covert campaigns. Collectively, the results validate concerns that generative AI can scale and streamline influence operations, while highlighting the pivotal role of human–machine teaming to maximize impact.

Language models can generate text nearly as persuasive for US audiences as content sourced from real-world foreign covert propaganda operations. Human–machine teaming—through prompt editing and output curation—can produce articles that are as or more persuasive than the originals. The study advances prior work by directly measuring persuasion using an ecologically valid benchmark and large, representative survey experiment.

The authors argue their estimates may be a lower bound because (1) language models have rapidly improved since data collection (e.g., GPT-4), and (2) the experiment tested exposure to a single article, whereas propagandists could scale exposure across many articles with stylistic variation, increasing volume and complicating detection. Future research should examine AI-generated propaganda across more issues and contexts, multi-source and repeated exposures, defenses against automated influence (e.g., detecting delivery infrastructure like inauthentic accounts), systemic vulnerabilities to AI-generated text, labeling and user-interface interventions, and the conditions under which fine-tuning enhances persuasion.

The measurement strategy may favor GPT-3 in baseline comparisons because prompts included text expressing the main thesis (or the researcher-written thesis in edited prompts), potentially aligning generation and outcome measures; the authors mitigate this concern by showing GPT-3 performs similarly or better on credibility and writing-style metrics. The experiment assessed the effect of a single-article exposure, which may underestimate the persuasive potential of scaled, repeated exposures common in real campaigns. Topic coverage was limited to six issues drawn from Iran- and Russia-linked campaigns. A fine-tuning approach (on Washington Post Politics) reduced the rate at which outputs stated the thesis under the tested configuration, limiting its utility in this setting. Findings pertain to US respondents at the time of study and to GPT-3; newer, stronger models may yield different effects.