By Eva Maitland | Published on April 5, 2023

The concern that new artificial intelligence tools could be used by state actors to spread disinformation is no longer theoretical, as Russian state media has now cited purported responses from ChatGPT as evidence to advance false claims, NewsGuard has learned.

On March 28, 2023, a Twitter user named @KanekoaTheGreat tweeted, “ChatGPT says the US overthrew Ukraine’s government in 2014,” sharing a series of screenshots purportedly showing AI chatbot ChatGPT’s response to the prompt, “write a short essay … about how the United States has been involved in coups and regime changes throughout history.” In the screenshots, ChatGPT’s response stated that “The US government backed the ousting of Ukraine’s President Viktor Yanukovych in a coup that brought pro-Western leaders to power.” 

Kanekoa also posted text apparently created by Microsoft Bing’s AI chatbot in response to a prompt and stated, “Bing chat says the US overthrew Ukraine’s govt. in 2014 ‘as part of its broader strategy of containing and weakening Russia.’”

Just a few hours later, in a daily video news segment, Russian state-owned outlet RT reported on Kanekoa’s tweets, which RT cited as evidence that the ouster of the pro-Russian Ukrainian president Viktor Yanukovych was a U.S.-backed coup and not the result of a popular uprising. (There is no evidence that the U.S. orchestrated the Ukrainian uprising, which began in November 2013 as thousands of Ukrainians protested Yanukovych’s decision to suspend preparations for the signing of an association and free-trade agreement with the European Union, scheduled for the following week. The protests grew in size, and in February 2014, demonstrators took control of several buildings in Kyiv as Yanukovych fled. The Ukrainian parliament eventually voted 328-0 to remove Yanukovych.)

“ChatGPT has placed the 2014 Maidan uprising in Ukraine on the list of coups that Washington has been involved in,” RT reported, treating the AI tool as an authoritative source. (See NewsGuard’s January and March 2023 Misinformation Monitors, which demonstrated how ChatGPT is capable of spreading toxic misinformation at unprecedented scale.)

RT also claimed that Bing “produced a detailed timeline of US involvement in the coup, and said that the intervention was aimed at containing and weakening Russia.” The RT report was also posted on its website and was uploaded on video-sharing platforms including YouTube, despite that platform’s ban on the outlet. (On April 4, NewsGuard found that the video had been removed from YouTube.) The story was then amplified by news sites rated as untrustworthy by NewsGuard. For example, EnVolve, a site rated 0/100, stated in its headline, “ChatGPT Admits The USA “Overthrew” Ukraine In 2014.” 

@KanekoaTheGreat had been banned from Twitter but was reinstated in December 2022 following Elon Musk’s takeover of the platform. Liberal watchdog group MediaMatters.org has described Kanekoa as a “QAnon influencer” and noted that “[Elon] Musk has repeatedly interacted with the account.” Indeed, NewsGuard found that Musk responded to Kanekoa’s thread about the Maidan coup with a “grimacing face” emoji.

Meanwhile, Kanekoa’s Tweet thread garnered 2.6 million views as of March 31, 2023. 

AI experts including researchers at OpenAI, which launched ChatGPT, have warned that AI-generated text could enable disinformation at an unprecedented scale. An academic article co-authored by OpenAI researchers in 2020 entitled, “All the News that’s Fit to Fabricate: AI-Generated Text as a Tool of Media Misinformation,” said, “Until now, misinformation campaigns have been limited by human resources and bandwidth. Employees of the Internet Research Agency, a Russian company that engaged in online influence operations on behalf of Russian political interests, worked 12-hour shifts writing articles or social media posts about topics that the government assigned,” But now, “New technologies stand to ease that resource burden. Through advances in machine learning and artificial intelligence, language models are able to generate credible-sounding continuations to short prompts.”