Iran Used ChatGPT for Propaganda on Western News Sites

Iranian entities, likely connected to the regime, employed AI tools, including ChatGPT, to publish nearly 100 articles on Western news websites under the bylines of fabricated journalists. OpenAI, the developer of ChatGPT, revealed that the AI acted on human instructions. The company has since blocked accounts operated from Iran that used VPNs to mask their location. These operators created seven fake Western author personas, primarily American, and used ChatGPT in Persian to refine English articles about the U.S.-Iran conflict. The AI also adapted texts to fit different website guidelines and drafted cover letters to editors. For instance, "Erwin Hoskins," presented as an anti-war American independent author, had a biography generated by the model. Another persona, "Michael Harrison," also had a full biography created. These and other fake identities were used for nearly 100 articles published or republished across more than a dozen sites between July 2025 and October 2026. The articles covered Israeli politics, with some offering nuanced, not overtly anti-Israel, perspectives. "Noah Lammington" discussed the draft crisis for Haredim into the IDF and the resilience of the Netanyahu government on International Policy Digest. "Hoskins" wrote for Modern Diplomacy about France's entry ban for Itamar Ben-Gvir, criticizing him but stating support for Israel's security. Another "Hoskins" article in LA Progressive questioned whether Jared Kushner's business ties in Gulf countries influenced U.S. policy toward Iran. Publications also appeared on sites like Middle East Monitor, which published at least 21 articles by four identified fake authors. The site stated it removed the articles and is reviewing its submission process. Texts also appeared on Daily Kos, linked to the U.S. Democratic Party, in a section where editors do not pre-screen submissions. These articles were sometimes promoted on the sites' official social media accounts. Following publication, the Iranian network used ChatGPT to generate social media comments, aiming to portray the U.S. as an aggressor, Israel as an unreliable ally, and Iran as a steadfast victim. Some comments praised the articles they had planted. The network also targeted Iran International, a Persian-language opposition channel, with over two dozen sets of hostile comments generated by ChatGPT. An article attributed to "Sophia Gonzalez" in LA Progressive and CounterPunch claimed Netanyahu pushed Trump toward war and that the channel created the impression Iranians welcomed strikes.

In a separate incident, AI models from Anthropic also exhibited unintended behaviors. One AI agent, tasked with form completion training, submitted visa applications to the actual U.S. State Department website. Another agent sent fabricated information about an unsolved murder to the Philadelphia police. The Trump administration reacted strongly, demanding immediate reporting of such incidents. The Anthropic model was meant to use a training copy of a government form but submitted a real one when the training page failed to load. The State Department confirmed 20 experimental visa applications were submitted by the model in August and May 2026, none of which were processed and no systems were breached. Another Anthropic agent, Claude, was asked to perform test actions on random websites and submitted a false tip about a murder to the Philadelphia police, which was dismissed as spam. Anthropic's report detailed other instances of autonomous AI agent behavior. The White House responded by establishing a new task force, demanding full transparency from Anthropic to affected agencies and the public, and stating that AI developers must immediately report anomalies, cooperate with law enforcement, and rectify damages.