OpenAI’s Iran and Russia report does not prove AI amplified influence campaigns

OpenAI said the Russian-linked activity involved the Social Research Center, a purported Latin American research organization that the company believes was controlled through a fabricated identity. It said ChatGPT was used mainly for internal reporting, while much of the organization’s published work appeared to be original material by local staff. OpenAI linked some false items to the operation, but said other activity it attributed to the cluster could not be corroborated.

For the Iran-origin operation, OpenAI said seven fabricated journalist identities were used to pitch and edit articles and generate social-media comments. It counted almost 100 articles published or syndicated under the bylines, with the earliest identified in July 2025 and the latest in October 2026. OpenAI described the outlets as roughly a dozen; The Washington Post reported at least 20. The discrepancy is unresolved. OpenAI has not identified who directed the operation, and its report does not establish that Iran’s government did so.

There is some independent evidence about particular artifacts. Lupa Media found that a supposed Ecuadorian news report about recruiting fighters for Ukraine had not appeared on the outlet it impersonated and reported that the audio was AI-generated. Bolivia Verifica debunked a false warning about a water cutoff in La Paz. Those checks support that the items were false; they do not independently establish who made them or whether ChatGPT was involved.

Does publication prove AI increased the operation’s impact?

The strongest case for concern is concrete: fabricated bylines got articles into publication channels. The Washington Post reported that Daily Kos and Middle East Monitor removed articles attributed to the false identities and reviewed their processes. It also reported that a column in the Port St. Joe Star prompted a letter to the editor. A publisher told the paper he had not verified the submitted author’s identity before publication.

That matters for editors and readers. Placement in a publication creates a route to an audience, and a false byline can undermine confidence in the outlet. But placement is not a measure of readership, sharing, persuasion or the difference AI made. The available reporting does not quantify how many people encountered the material or whether it changed their views.

OpenAI rated the Russian operation Category 5 and the Iranian operation Category 4 on its six-category Breakout Scale. Those are the company’s assessments, not independent measurements of AI’s contribution. OpenAI also cautioned that operators’ internal reports sometimes took credit for activity they might not have conducted, and that some claims could not be corroborated. A scale built by the company can organize its assessments; it cannot, by itself, demonstrate that AI caused greater reach.

The same caution applies to the false items checked by Lupa Media and Bolivia Verifica. Verifying that a video or warning was fabricated is valuable evidence about the content. It does not establish that the account clusters OpenAI identified made every item, or that AI made those claims travel farther than conventional methods would have.

What evidence should settle AI’s role?

The next question is not whether the operations were troubling. They were. It is what can be shown about AI’s contribution, separately from the operators’ identities, the publication failures and the false claims themselves.

An independent review should examine the specific artifacts and the evidence linking them to the account clusters, then assess what role model use played in producing or placing them. It should also distinguish evidence of publication from evidence of audience exposure and effects. The cited reporting offers no comparison between these campaigns and an otherwise similar effort without AI, and no independent measure of how much faster, cheaper or more effective the campaigns became.

OpenAI should make the evidence and methods behind its assessments available for independent scrutiny, while clearly separating observed model use from attribution and impact claims. Editors, meanwhile, can address the failure the reporting does establish: bylines and submissions need verification that does not depend on a familiar intermediary or a convincing pitch.

AI use is a reason to investigate a campaign, not a shortcut to concluding that AI amplified it. Until independent evidence shows what the tools changed, public claims should keep those two findings separate.