False identities, with AI support

OpenAI says it banned ChatGPT accounts associated with two influence campaigns, which it describes as originating in Russia and Iran. Its October 8 report examines how deceptive identities helped messages reach established publications.

The company describes seven journalist personas used to pitch articles, and a separate research front that appears to have employed people unaware of its operators' identity. It has not identified the particular actor behind the Iran-origin campaign. The report should not be read as establishing state sponsorship.

Writing was not the only task

According to OpenAI, the models helped refine pitches, comments and internal reports. In much of the Russia-origin activity, the company observed report-writing rather than creation of the campaign's public content. This is not evidence of an entirely automated operation.

The company also questions the operators' own success metrics. One calculation counted views on posts they replied to, rather than on their replies. That confuses the audience of an existing discussion with the attention earned by a new contribution.

A useful scale, with a limited question

The report uses the Breakout Scale, a framework published by Ben Nimmo at Brookings in 2020. It sorts influence operations by how far they spread across platforms and communities, how media or prominent people amplify them, and whether they prompt concrete action.

Those categories provide a way to compare observable distribution. They are not a percentage of people convinced by a message. An article reaching an established outlet and a reader changing their mind are different events.

The framework also explains why editors and other high-profile intermediaries matter: they can bring an operation's messages to audiences beyond the channels it controls.

The practical issue is who stands behind a contribution

For a newsroom, the useful response is a verification question: can the contributor's identity and claimed affiliations be checked separately from the polished draft? Fluent writing is not an identity check.

For readers, the remaining questions concern attribution, how much activity was observed and what effect it had. Model Current has not independently verified the operators' identities or measured audience response. This company investigation offers evidence about a workflow, not a verdict on AI's overall influence on public opinion.

Sources

  1. OpenAI: Disrupting AI-enabled false-front operations, October 8Primary company investigation supports attributed bans, campaign origins, seven bylines, the research front, writing/reporting use and questionable view-counting. The particular Iran-origin actor is unidentified. It does not establish state sponsorship or measured persuasion.
  2. Brookings: Ben Nimmo's Breakout Scale, September 2020Original methodological executive summary defines six observable distribution/amplification categories. Used as older context, not a new study or direct measurement of individual persuasion. No numerical campaign score or independently confirmed effectiveness asserted.