OpenAI Uncovers Russian and Iranian Influence Operations Using AI

OpenAI has exposed the use of its AI tools by Russian and Iranian operatives to conduct influence campaigns. The company, in a recent blog post, announced the banning of accounts from these countries after identifying their role in spreading false narratives.

Internal reports revealed the operatives’ success in getting fake narratives picked up by news outlets. One instance even saw a response from a European Parliament member. These operations, primarily through ChatGPT, targeted Latin American countries and aimed to undermine support for Ukraine while affecting local political dynamics.

OpenAI’s investigation showed that embedding false narratives within established media reached wider audiences compared to using fake social media accounts. In a significant campaign, Russian operatives impersonated a Peruvian educational authority. They tricked schools into organizing activities featuring Stepan Bandera, a controversial historical figure.

“The Russian fake thus both fed on, and fed into, historical tensions between Ukrainians and Poles,” OpenAI wrote.

Extra, a Peruvian outlet, reported such an event happening in a school, leading to a political reaction in Poland. Russians utilized ChatGPT to create internal memos claiming further influence operations. They managed a fake think tank in Latin America, enlisting unwitting researchers to produce reports.

The operations included narrative laundering tactics, as noted by Darren Linvill, a co-director at Clemson University’s Media Forensics Hub. This included creating fake stories and personas to align with Russia’s political goals. A similar effort in 2020 involved posing as a news outlet linked to Yevgeny Prigozhin’s Internet Research Agency.

Fake credit was taken by Russians for events beyond their doing, and AI-generated material formed only a small part of their workload. Linvill highlighted how AI use intended to lessen work ironically exposed the operatives.

Iranian operatives, contrastingly, used AI differently. They focused on disseminating content regarding the U.S.-Iran conflict and American politics, appearing like a commercial actor’s influence campaign. OpenAI was unable to pinpoint the responsible party.

Through seven fabricated authors, the Iranian operation published long articles, nearly 100 in total, across small media worldwide. Targets included Daily Kos in the U.S. and the Middle East Monitor in the U.K. The latter published 21 articles by four fake authors, but with no response to requests for comment had removed the profiles.

Daily Kos clarified that its community section, used by a fake persona, is distinct from staff reports. This section is open for public posting and not pre-reviewed by Daily Kos staff, though rules prohibit fake identities. The persona’s posts remain accessible.

Linvill emphasized AI’s role beyond simplifying tasks. It enabled creating content that appeared authentic enough to be accepted by legitimate outlets, a challenge Iranian operators would have faced greater difficulty with in the past.