Listen again · In Conversation
How AI Is Used in Influence Campaigns October 9, 2026 · 11:18 · With Maya, Alex, and Leo
Leo covers technology and digital life, and Alex is a station journalist. Thanks for listening, wherever you are.
Transcript Tap a sentence to hear it from there.
Maya Welcome to Slow English FM. This is our round table on technology and world events. I’m Maya. Leo covers technology and digital life, and Alex is a station journalist. Thanks for listening, wherever you are. We’re looking at reports that Russian and Iranian networks used ChatGPT to support deceptive online campaigns, and what those cases can—and cannot—tell us about AI and information security. Maya Alex, let’s begin with the latest report. What does it say happened? Alex NPR reports that OpenAI identified Russian and Iranian operatives using ChatGPT in influence campaigns. France 24 reports that OpenAI said it had shut down operations linked to those countries. The reported uses included creating fake journalist identities and supporting a supposed think tank that spread anti-Ukraine propaganda. These are descriptions of activity identified by OpenAI, not proof that every post reached a large audience or changed anyone’s views. Maya So the concern is not simply that someone used a chatbot to write a post. The concern is that the tool may have helped make a wider campaign look like ordinary reporting or discussion. Leo, what can AI add to that kind of operation? Leo It can help with the work around the message. In a 2024 report, OpenAI said generative AI could make it more efficient to write, translate, and post propaganda. In plain terms, a group can use a text-generating system to produce or adapt material with less effort. That does not make the material true, and it does not prove that people will believe or share it. It changes a possible part of the production process. Alex The earlier cases also show why it is useful to look beyond the text itself. The Guardian reported that OpenAI’s May 2024 review described five covert influence operations linked to state and private actors. The operations were connected to Russia, China, Israel, and Iran. OpenAI said they used AI alongside other methods, including text written by people and memes copied from the internet. So the reports do not describe campaigns run by AI alone. Maya Leo, Alex says the AI was one tool among several. Does that distinction matter? Leo Yes. It helps us avoid treating AI as either the whole campaign or a magic solution. A campaign still needs a purpose and a way to put messages in front of people. The reports describe human operators using AI for tasks such as writing, translation, or technical work. OpenAI also said one Russian operation used an AI model to help debug code and create a bot that posted on Telegram. That is a different use from writing a fake news article, even if both involve AI. Alex And the content varied by operation. According to The Guardian’s account of OpenAI’s May report, Russian operations created material criticizing the United States, Ukraine, and several Baltic nations. The same account said an operation linked to China generated text in English, Chinese, Japanese, and Korean, then posted it on Twitter and Medium. OpenAI’s report therefore described more than one language and more than one kind of online activity. Maya That gives us a sense of the range. But a message appearing online is not the same as a message having an effect. Alex, what do the reports say about audience response? Alex The Guardian reported that none of the campaigns in OpenAI’s May 2024 report gained traction or reached large audiences. It also reported that, in August 2024, OpenAI said an Iranian operation called Storm-2035 did not appear to achieve meaningful audience engagement. OpenAI said most of the social-media posts it identified received few or no likes, shares, or comments. Those are separate reports about particular operations. They do not establish how every influence campaign performs. Leo That distinction is important. A system may make it easier to create many versions of a message, but more output does not automatically mean more attention. The Storm-2035 case shows that the effort to publish content and the effort to reach people are not the same thing. The evidence we have describes low engagement for that operation. It does not tell us that every future operation will have the same result. Maya The operation’s presentation matters too. Leo, what does it mean to pose as a news outlet? Leo It means presenting a site as if it were a source of news, when its real purpose may be to promote a particular message. A website can look like a publication without being an independent newsroom. The Guardian, citing a Microsoft threat-intelligence report, described Storm-2035 as four websites posing as news outlets and engaging groups of US voters with polarizing messages. That finding concerns the websites’ presentation and activity. It does not mean every reader was deceived. Alex There was also a mix of formats. The Guardian reported that Storm-2035 used ChatGPT to generate long articles and shorter comments for social media. Its subjects included US election candidates, Gaza, and Israel’s participation in the Olympic Games. The same report said Iranian actors had used generated articles attacking the United States and Israel, then translated them into English and French. Translation can help a message cross language boundaries, but the reported facts do not show that every translated item found an audience. Maya So, on one side, AI may help produce and translate content. On the other, the reports give us examples where the audience response was limited. Alex, how should we understand the newer report about Russian and Iranian networks alongside these earlier cases? Alex As evidence that companies are continuing to report attempts to use their tools in influence activity, and that the reported tactics can include false identities or apparently independent outlets. But the public details in the current reports, as summarized by NPR and France 24, are limited. We should not assume that the newly reported networks used every method seen in Storm-2035 or in the 2024 cases. They are different operations, reported at different times. Leo That is also why the word “influence” needs care. It describes an attempt to shape what people see or think; it does not, by itself, show that the attempt succeeded. In these reports, there is evidence of content creation and online activity, and some evidence about engagement in specific earlier cases. Those are useful but different questions. A campaign can exist without reaching a large audience. Maya Does the evidence point to a new kind of information threat, or an older kind of deception with another tool? Leo My reading is that it points to both continuity and change. The continuity is deception: presenting a message through a false identity or a site that looks like news. The reported change is that generative AI can assist with writing and translation, and in one earlier case with code and a posting bot. That may lower the effort needed for some tasks, as OpenAI said. But the reports also describe human-written material and copied memes. It is not a switch from human campaigns to fully automated ones. Alex And we should keep the claims attributed. OpenAI is reporting what it identified on its own service, and its statements describe its own findings and actions. The Guardian’s account of the Microsoft report adds another view of the Storm-2035 websites. These reports help explain the tools and methods described. They do not show, on their own, that an operation changed an election or persuaded a particular group. Maya That is a useful limit. The story is about reported attempts and methods, not a proven effect on voters. Leo, what should listeners take away about AI’s role? Leo AI can support several parts of a deceptive campaign: drafting text, translating it, or helping with technical tasks. It can make some of that work more efficient. But a generated article is only one part of a larger effort, and the reported cases show that producing content does not guarantee that people will engage with it. The tool matters, but so do the people using it, the channels they choose, and the response of the audience. Maya Alex, your summary? Alex The reports describe Russian and Iranian networks using ChatGPT in influence activity, with earlier cases showing a wider mix of countries, languages, and tactics. They also show why we should separate a campaign’s activity from its impact. Some earlier operations received little engagement, according to OpenAI’s reports as covered by The Guardian. The newer reports add to the picture of how such tools may be used, but they do not establish that every effort succeeds. Maya Leo has explained how AI can assist with producing and adapting messages. Alex has set out what the reports say about the operations—and where their evidence about impact is limited. That is our round table on AI and influence campaigns. Thanks for listening to Slow English FM.
