AI firms face scrutiny over state-linked propaganda ops

Oversight questions are mounting over how artificial intelligence providers and platforms will police state-linked influence operations, after Anthropic disclosed that its models were deployed in covert campaigns run by or on behalf of foreign governments and political organizations, including the United Arab Emirates and Iran’s exiled opposition movement, the People’s Mojahedin Organization of Iran (MEK).

The Anthropic report, released last month, describes how four individuals working for the MEK built an AI agent called “Viktor” that impersonated a real Iranian activist on Telegram. The agent had been trained on 8,400 of the activist’s genuine posts, allowing it to mimic his style so convincingly that activists inside Iran who exchanged messages with it about politics and news had no idea they were not speaking to a person. Beyond the impersonation, Viktor surveilled individuals inside Iran and flooded social media with pro-MEK content presented under “independent” branding.

The scale is considerable. Viktor infiltrated more than 500 Telegram groups and, according to Anthropic, grouped its targets by city, age, occupation, political alignment, and arrest history, likely to help tailor messages to specific audiences. One MEK-linked account managed by the agent, @Tehranchekhabar19, has more than 170,000 Instagram followers and presents itself as an independent news account while posting content supporting the MEK’s views, including videos promoting the U.S.-Israeli war against Iran and attacking Reza Pahlavi, the former Crown Prince of Iran who leads a rival opposition faction. Those videos sometimes draw hundreds of thousands of views, and other Viktor-run Instagram accounts with similarly large followings appear to have since been shut down.

The MEK’s political footprint in Washington gives the disclosure particular weight for accountability watchers. Last year, 230 members of the House co-sponsored a resolution urging U.S. support for the MEK’s plan for Iran. Former National Security Adviser John Bolton, former Vice President Mike Pence, and former New York City Mayor Rudy Giuliani are among the high-profile Republican politicians who have spoken at MEK rallies, often for a hefty speaking fee. The group, which the U.S. designated a Foreign Terrorist Organization until 2012, is widely regarded as a cult.

Craig Holman, Government Affairs Lobbyist for Public Citizen, told Responsible Statecraft by email that the tactic represents a new form of influence. “These accounts of foreign governments using AI to manipulate the opinions of legislative bodies and the public are based on the root source of AI deception: falsely depicting a real person or entity doing or saying something that they never did,” he said.

Meanwhile, the UAE’s campaign, detailed in the same report, centers on an AI agent nicknamed “Deadshot” that Emirati officials created on Claude and charged with whitewashing the UAE’s role in the Sudanese civil war. A UAE-backed Sudanese paramilitary group, the Rapid Support Forces, has been credibly accused of ethnic cleansing and sexual violence, and a U.N. fact-finding mission said the RSF’s actions against non-Arab communities in Darfur bear the “hallmarks of genocide.”

Deadshot’s work targeted the U.N. directly. According to the report, the agent hijacked the identity of a real Sudanese human rights organization and ghost-wrote U.N. testimony in hopes the world body would be swayed by what appeared to be independent local witnesses rather than state messaging. Anthropic’s account is inconsistent on whether the effort succeeded: at one point it says it could not confirm the testimony reached its intended audience, while a summary elsewhere describes it as “delivered in a live UN Human Rights Council session.” Anthropic did not respond to a request for comment. The agent also compiled “counter-accountability dossiers” on U.N. Special Rapporteurs who had criticized the UAE’s role in Sudan.

Alongside this, the UAE ran a broader information campaign on X, where Deadshot created and managed roughly 300 inauthentic accounts posting content that pins blame for the war on the Muslim Brotherhood, which backs the Sudanese Armed Forces against the RSF. The posts often take the form of humanitarian appeals, deflecting blame away from the RSF. One post from an account with 27,000 followers warns that the longer the war drags on, the deeper the famine crisis becomes, citing analysts who point to hardline Islamist elements favoring continued confrontation. Another, from a user with 119,000 followers, says the Muslim Brotherhood’s role “makes ending the war far harder.” The operator instructed Claude to carry out a “coordinated transatlantic and regional operation to dismantle the Muslim Brotherhood globally” hundreds of times, according to the report.

Kholood Khair, Founding Director of Khartoum-based think tank Confluence Advisory, told Responsible Statecraft that the Sudanese conflict is fought on three fronts: the battlefield, the mediation table, and the information space, with the last becoming more crucial. “These posts are carefully deployed to attempt to co-opt the pro-civilian rule, pro-democracy movement and camouflage what the UAE is actually trying to achieve in Sudan,” she said. While many posts are not factually incorrect, she noted, they are selective in the narrative they tell. “AI is the next battleground in many of these wars, and Sudan’s war has clearly become a testing ground, with the UAE getting a head start,” she added.

The disclosures sit within a wider pattern of accountability gaps. An OpenAI report published last week found that Iranian operators placed almost 100 articles in international affairs-focused publications under phony bylines using ChatGPT, typically promoting an anti-Iran war point of view, with headlines such as “Is America’s Iran Policy Being Privatized? The Jared Kushner Question” and “Trump’s Iran Gamble Hardened the Regime It Meant to Break.” The same report found that Russia created and controlled a fake think tank called the “Social Research Center,” with real staff in Latin America who had no idea they were working for Russia; operators used ChatGPT to draft content and internal reports while the unknowing staff conducted real interviews and published research, mostly on India-Latin America relations. The organization’s X account has been suspended, but its website remained active as of publication, carrying a statement rejecting OpenAI’s claims as “factually incorrect and as constituting a deliberate attempt to discredit an independent research project.”

Taken together, the reports suggest foreign actors are taking concrete steps to add AI to their information-war arsenals, raising unresolved questions about how platforms, AI providers, and governments will account for operations that blur the line between ordinary model use and coordinated deception.