Dougan Leak: Did a Russian Propagandist Use Claude and Qwen for 43,000 Articles a Month? What Is Claimed and What Is Verified

Leaked files from John Mark Dougan, a US-born propagandist living in Russia, show that he used Anthropic's Claude and Alibaba's Qwen while running influence work, and that he told Claude he was producing 43,000 articles a month. That is what NewsGuard reports from the files. The claim that AI made all 43,000 articles, or that chatbots now repeat his stories because of it, goes beyond what the sources show.

The story spread fast on November 12 with some summaries stating the volume as fact. This post separates what the two primary reports say from what remains unproven.

TL;DR: claimed vs verified

What was leaked and how

NewsGuard says that on November 10, 2026, documents from Dougan's laptop were posted to a website that mimics his own. They include emails, Telegram and Signal messages and AI chatbot prompts. The link was shared from his apparently hacked X account. Dougan confirmed the leak was authentic to NewsGuard over Signal on November 11.

Also on November 10, The Insider published a long investigation based on a hack of his messages. It describes him as a former US police officer who fled to Russia in 2016 and became a Kremlin-linked disinformation asset, and it argues his career then declined. The two reports use overlapping material, but they focus on different things. The Insider is about the man and his handlers. NewsGuard is about the AI use.

What the AI chats show

NewsGuard says the files hold dozens of interactions from April to August 2026 with Claude and Qwen. Dougan also ran a Qwen model locally through Open WebUI. The examples it lists are:

• Claude, June 8, 2026: helped him set up a "free speech site for Europe," including name ideas and advice on French phrasing.
• Qwen: wrote a Telegram post about the midterms claiming Democrats staged protests.
• Qwen: helped draft anti-Western arguments built on a false claim about US bioweapons labs.
• Qwen, July 2026: gave tips on making text evade AI detection.
• Qwen: wrote a "machiavellian" dating profile.
• Unspecified tools: he asked how to edit his Wikipedia page to add comments from Tulsi Gabbard.

NewsGuard notes the documents only confirmed publicly disclosed US funding for storing Soviet-era pathogens, which is not the same as the bioweapons claim. It also says the chatbots complied in cases that supported his efforts. The Insider's article, by contrast, says that in March 2024 Dougan asked Claude what an API was, and it uses that to question his technical skill. It does not mention Qwen.

The 43,000 number

The figure that travels is 43,000 articles a month. Here is what is behind it.

According to NewsGuard, in a prompt to Claude Dougan wrote that he was producing 43,000 articles a month. NewsGuard adds that his overall operation created tens of thousands a month and that costs were about 2 million rubles, roughly $23,000, a month.

Those are statements by a man talking to a chatbot, and he has reasons to inflate. They are not a measured output. The Insider's article gives no output volume beyond saying the Storm-1516 network, which Clemson researchers describe, includes hundreds of outlets that claim to be in small US cities or European capitals and are populated using AI. Some social posts and aggregator pages turned "he said he produced 43,000" into "he used Claude and Qwen to produce 43,000." The sources do not say that. Nor do they say how much of any output came from Claude or Qwen as against other tools or human work.

What The Insider adds

The Insider's investigation is the stronger source on who Dougan worked for. It reports that:

• Yuri Khoroshenky, an officer of GRU Unit 29155, contacted Dougan on May 19, 2022 and asked him to promote links about US-backed biolabs.
• The GRU used Dougan as a financial cutout to hide the source of payments.
• He dealt with two FSB Economic Security Service handlers, Vladislav Yevsyukov and Yevgeny Firsov.
• Clemson, NewsGuard and Insikt Group traced the Storm-1516 network back to Dougan, and he denied involvement.
• In June 2026 Tulsi Gabbard endorsed his biolabs claims, and she resigned on June 30.

These are the outlet's reading of messages it obtained. We have not seen the underlying files and cannot verify them. Dougan denied involvement in the Storm-1516 tracing, and the leak is a hack, so readers should weigh provenance.

What is not shown

The most important gap is the link between this leak and chatbot behavior. NewsGuard says the documents do not show whether Dougan succeeded in fine-tuning a model or infecting existing ones. A file titled "Request for LLM Data" describes a goal of tuning an uncensored model on pro-Russian sources, but it is unclear to whom the request was made.

That means the leak does not prove that Claude, Qwen or any other chatbot now repeats his stories. Earlier NewsGuard audits of a different Russian network, Pravda, found that leading chatbots repeated its falsehoods about a third of the time in a 2025 test. That is a separate result, and it should not be merged with this one. Our coverage of the Nous Research propaganda-reduction claim shows how hard it is to measure this problem even when the test is public.

Why the timing and the dates matter

The chat logs run from April to August 2026, and the leak surfaced in November. That gap matters. The Qwen request for AI-detection evasion tips is dated July 2026, and the Claude exchange about the European site is dated June 8, 2026. Any safeguards either lab shipped after those dates would not show up in these files. The midterm Telegram post that Qwen wrote also predates the November vote. So the files describe past behavior of the models and of a user, not how Claude or Qwen would respond today. A fair test would rerun similar requests against current versions, and nobody has published one yet. Until then, the safe reading is narrow: a known propagandist used mainstream chatbots for drafting and advice, and at least some requests were answered.

Where the AI companies stand

NewsGuard emailed both companies on November 11. Neither had replied when it published. Anthropic's usage policies prohibit using Claude to create or spread deceptive information and to impersonate people or entities. Alibaba's policy bars using Qwen to spread misinformation.

Using a chatbot to rephrase text, pick a site name or draft French is hard to tell apart from ordinary use. That is the hard part for any lab. A refusal at the prompt level works only when the request itself shows the intent. Anthropic has said it runs election safeguards, which we covered in Claude's midterms safeguards, and it has banned accounts for other abuse, as in the Russian passport persona bans. Neither covers this case, because no company has commented on it.

How it fits the pattern

This is the latest in a run of influence stories. OpenAI disrupted Russia and Iran front operations in October. Our check of the Storm-1516 Talarico deepfake found a confirmed link to the same network name. A Bloomberg report on Matryoshka AI celebrity videos showed the same playbook with video. The pattern is cheap volume plus a few polished pieces.

For readers who build or use AI products, the takeaways are practical:

1. Do not treat a volume claim as a measurement. A propagandist's brag in a chat is a claim, even when it appears in a real leak.
2. Separate use from effect. Proof that a tool was used does not show the output reached or fooled anyone.
3. Watch the training-data angle. The "LLM grooming" idea, publishing text so that models absorb it, is a real concern in public research, but this leak does not demonstrate success.
4. Check provenance. The files came from a hack, confirmed by the owner to one outlet. Look for independent reporting before you quote numbers.

What to watch next

• Replies from Anthropic and Alibaba to NewsGuard.
• Independent review of the files by other researchers, and whether the 43,000 figure holds up against the actual site output.
• Any audit that tests whether chatbots repeat Storm-1516 claims after this period.
• Whether the "Request for LLM Data" file leads to a named recipient.

Related reading

• Storm-1516 and the Talarico deepfake: what is confirmed
• OpenAI disrupts Russia and Iran influence operations
• Matryoshka AI celebrity videos and Democratic Senate candidates
• Anthropic's US midterms safeguards for Claude
• Claude Russian passport persona bans: claimed vs verified
• Russian army AI fake battlefield videos

This post reflects sources available on November 12, 2026. We will update it if Anthropic, Alibaba or other researchers respond.