OpenAI Bans Accounts Reviving Russia's 'Stop News' Influence Operation

OpenAI has banned a set of accounts that tried to use its models to resurrect "Stop News," a covert Russian influence operation the company first disrupted and disclosed in October 2024.

The case study, originally published as part of OpenAI's October 2025 report on disrupting malicious uses of AI, identifies the activity as originating in Russia and consistent with an operation run by a marketing company. The stakes for the industry are clear: this is a documented case of an influence actor returning after disruption by OpenAI and its industry peers, adapting its toolchain rather than abandoning the campaign. It also shows how operators chain together multiple AI providers — OpenAI for text, unidentified services for video — to build persuasive content at scale.

What the operation did

The revived campaign mainly used OpenAI's models to generate content posted on social media and a set of websites that impersonated the same fake news outlets in Africa and the UK that OpenAI described last year. The actor also attempted to use other companies' AI tools to generate videos later posted on YouTube and TikTok. OpenAI states it cannot independently confirm which models were ultimately used for those videos.

OpenAI's investigators flagged two notable shifts from the original 2024 activity. First, image generation dropped sharply. The original Stop News operation was, as OpenAI noted at the time, unusually prolific in its use of imagery; the new activity focused on text generation, with images relatively rare. OpenAI calls this consistent with the assessment by the French digital agency VIGINUM that "since the publication of the Meta and OpenAI reports, the use of AI-generated images has fallen drastically."

Second, the operation used ChatGPT to produce scripts and descriptions for news-style short videos. The workflow was methodical: operators would input a lengthy Russian-language text and ask for a video script, then request a translation into French, then ask the model to generate an SEO-optimized description and hashtags.

An AI newsreader with a changing face

Some of these scripts appeared in videos on a YouTube channel linked to the operation's Africa-focused website, newstop[.]africa. Others ran on a YouTube channel and a TikTok channel that did not carry the "Stop News" branding. Both featured an AI-generated person reading the news. In most cases, brief clips of the newsreader were interspersed with real footage of events in, or relating to, Africa. OpenAI says the newsreader clips do not appear to have been generated with its models, and it cannot confirm which service produced them.

The newsreader's appearance evolved through 2025 — screenshots from the operation's TikTok account show the avatar praising Russia in January 2025, early June 2025, and late June 2025.

The most elaborate workflow OpenAI documented shows cross-model orchestration. Operators first generated a Russian-language video script with ChatGPT, broke it into two-sentence snippets, and asked ChatGPT to write a video prompt for each snippet, apparently for use with other AI models. They then had the model translate the entire script into French. OpenAI used this method to reconstruct a TikTok video praising Russia's "Africa Corps." A telling forensic detail: the video's caption contained a grammatical error — "ces" instead of "ses" — that was absent from ChatGPT's French translation, suggesting the caption was transcribed from audio and never proofread by anyone with French skills.

Three categories of output

The operation's completions fell into three groups. It generated French-language content criticizing the roles of France and the United States in Africa while praising Russia's role there. It generated English-language content criticizing Ukraine and its international supporters, occasionally with accompanying images. On rare occasions, it asked ChatGPT for promotional articles and Google ad texts for its main brand, "Newstop Africa" — some of which bracketed the Newstop brand together with legitimate news outlets, likely to borrow credibility by association.

The third category was different in kind: promotional materials for what appeared to be a range of commercial companies in Russia, including ads for unlicensed online gambling. OpenAI notes that this mix of covert influence work and conventional advertising is consistent with a commercial company running influence operations for hire, and consistent with what it reported last year.

Limited impact, downgraded assessment

Despite the pivot to short video and more complex multi-tool workflows, the content barely registered. The core "Newstop Africa" X account had 172 followers as of August 2025, and X's "Top tweets" function showed a maximum of four retweets on any post.

The YouTube and TikTok channels featuring the AI newsreader each had roughly 1,900 followers. The TikTok channel recorded 5,855 likes across 56 videos — an average of 105 likes per video — with the most-viewed video at 63,300 views and the least-viewed at 87. The YouTube channel logged some 255,000 views across 50 videos, averaging 5,100 views per video, but OpenAI found no evidence of the videos being re-shared, cited in media, or achieving wider resonance.

OpenAI has also revised its overall assessment downward. In its October 2024 report, it rated the operation Category 3 on the Breakout Scale, largely on the basis of apparent "information partnerships" with UK-based websites. Subsequent research by VIGINUM and open-source researchers showed those partnerships were likely fictional and had "exploited technical flaws on these external sites" to insert content without administrators' knowledge. OpenAI now assesses the operation at Category 2: activity across multiple internet platforms, but without significant breakout into authentic communities.

The episode offers two lessons for platform integrity teams. Disruption forces adaptation — the actor shifted from images to text and video, and toward chaining multiple AI providers — but adaptation does not guarantee reach. Whether that pattern holds as video generation tools improve is the question OpenAI's threat researchers will keep watching.