# Bot or not? How China, Iran and Russia use AI to fake consensus in social media comments

> Comments written with ChatGPT to praise an article or tear down an opponent, made-up journalists, fake think tanks. OpenAI's reports show how China, Iran and Russia use AI to simulate a consensus that doesn't exist. So far with little success. The catch is that we only know because someone reads their chats.

- URL: https://mautoblog.com/en/posts/ai-bots-social-media-comments-fake-consensus-china-iran-russia-2026/
- Language: en
- Published: 2026-11-20
- Author: Matteo
- Categories: Culture and Society, Artificial Intelligence
- Tags: Social Media, OpenAI, Politics, Digital Society, Local AI
- Translation (it): https://mautoblog.com/posts/bot-ia-commenti-social-falso-consenso-cina-iran-russia-2026/

---

In the autumn of 2023, an X account that had spent weeks attacking Alexei Navalny and his associates posted an odd reply. Instead of the attack, it contained ChatGPT's refusal: the request "goes against OpenAI's use case policy by promoting hate speech or targeted harassment".

It was one of at least 64 accounts that researchers at the Institute for Strategic Dialogue had seen working together ([ISD](https://www.isdglobal.org/digital-dispatch/hey-fellow-humans-what-can-a-chatgpt-campaign-targeting-pro-ukraine-americans-tell-us-about-the-future-of-generative-ai-and-disinformation/)). The targets were Navalny, Maria Pevchikh and their Anti-Corruption Foundation, presented to pro-Ukraine Americans as "controlled opposition" run by the Kremlin. From September 2023 all those accounts wrote with ChatGPT. One forgot to check what it was pasting.

Three years later the method is the same, just more polished. In 2026 OpenAI published three reports on campaigns from China, Russia and Iran that used its models to fill social media with fake comments. I read them, and they tell a more interesting story than "AI is rigging elections".

---

## The dirty work happens in the comments

When people talk about disinformation they picture the viral post, the hoax that goes around the world. These campaigns almost always work lower down: **in the replies under other people's posts**.

The mechanism is simple. Under an article, thirty enthusiastic comments appear, or thirty outraged ones. Readers don't check each commenter's followers; they get a sense of "what people think". It's a thermometer we all use, often without noticing.

Faking it used to be expensive. Bots were easy to spot because they repeated the same sentence with the same mistakes. Writing a hundred different, believable comments in the right tone took a hundred people. Today a prompt is enough: "write me thirty replies to this post, in different voices, as Americans of various ages". That's what AI has changed. It didn't invent propaganda; it **lowered the cost of variety**.

## China: data centers and tariffs, in Italian too

In June OpenAI shut down two clusters of accounts that very likely came from China ([June report, PDF](https://cdn.openai.com/pdf/96b559fa-c165-4575-805d-e636909e2f78/June-2026-Threat-Report.pdf)). According to OpenAI they were the social media team of a private tech company working for provincial governments. We know because they uploaded one of their internal reports to ChatGPT, with their goals and their techniques for creating accounts that evade platform detection.

The first cluster, which OpenAI called "Data Center Bandwagon", posed as ordinary Americans angry about their power bills. Comments and cartoons on X claimed that AI data centers push up electricity prices and that families end up paying. The clever detail: next to the comments they posted links to real news about electricity capacity auctions. They didn't invent a concern; they latched onto one that already exists.

The second, "Tech and Tariffs", attacked US tariffs as a way to dominate technology. The prompts included a precise instruction: **don't mention Xi Jinping, only Trump**. And among the languages requested, besides English, Japanese and Traditional Chinese for Taiwan, there was **Italian**. Even my corner of the world got some attention.

Around them was a network of X profiles claiming their ChatGPT data had been stolen and their lives ruined. All false. Those profiles had been created at the end of 2025, had very few followers and reposted each other. In one case, accounts from both networks quoted the same tweet by a stranger within a few hours. X had already suspended them on its own.

OpenAI points out the irony: they used American AI to attack American AI.

## Iran: seven journalists who don't exist

The most organized case is the Iranian one, disclosed in October ([CNN](https://www.cnn.com/2026/10/08/politics/iranian-operatives-used-ai-to-plant-fake-stories-in-us-media), [Iran International](https://www.iranintl.com/en/202610097033)). OpenAI called it "Bogus Bylines".

Seven made-up journalists, mostly presented as Americans, pitched articles to small and mid-sized online outlets, mainly about the war between the United States and Iran. The operators wrote their prompts in Persian. ChatGPT polished the English, adapted each piece to the outlet's guidelines and wrote the pitch emails. The result: **nearly 100 articles** published or syndicated by about a dozen outlets between July 2025 and October 2026. OpenAI rates it Category 4 on a six-point scale: the articles really did get out beyond the operators' own circle.

Then there were the comments, and this is where the praise-or-demolish trick shows. On one side, more than two dozen batches of hostile Persian replies under posts by Iran International, the London-based opposition outlet. On the other, comments **praising articles they had written themselves**, to make them look popular. OpenAI believes they were probably a for-hire company, whose reports to clients inflated the results with misleading calculations. Propaganda has its smoke-selling consultants too.

## Russia: the think tank that isn't there

In August it was the turn of a Russian group ([OpenAI](https://openai.com/index/disrupting-malicious-uses-of-ai-influence-campaign-russia/), [CNBC](https://www.cnbc.com/2026/08/25/openai-russia-chatgpt-influence-campaign.html)). They promoted the "International Burke Institute", a research institute supposedly based in Israel that doesn't exist. The institute published a "sovereignty index" that rewarded Russia and marked down Western countries. Comments on X, Facebook, Telegram, LinkedIn and Substack pushed its "analyses".

The detail that struck me: they prompted in Russian and asked ChatGPT for English comments, **stripping out the linguistic clues that could give away where they came from**. AI isn't only for writing more; it's for hiding better. One of the articles on the site was also copied from a Cambridge University Press publication and attributed to a professor at a different university.

The October report also included another Russian operation, aimed at Latin America. It revolved around a made-up persona, "Mia Clark", and a fake research center, with forged documents against Ukraine ([Business Today](https://www.businesstoday.in/technology/news/story/openai-bans-russian-and-iranian-accounts-over-ai-powered-fake-news-campaigns-targeting-global-audiences-560608-2026-10-09)).

## So far it barely works

Before panicking, there's one point the reports repeat in every case: **almost none of these campaigns broke through**.

OpenAI measures reach on a scale from 1 to 6, the Breakout Scale. Apart from the Iranian articles, the Chinese campaign on data centers and tariffs is Category 1: one platform, almost no engagement, no spread beyond its own network. The Russian fake-institute campaign sits at the low end of Category 3. The Iranian comments got likes in the single or double digits and were a minority among genuine comments. Researchers at the Carnegie Endowment, in an October analysis, also doubt these operations really shift opinions ([Carnegie](https://carnegieendowment.org/russia-eurasia/politika/2026/10/russia-AI-political-influence)).

So no, there's no AI army deciding elections. The damage is slower and subtler: if some of the comments are fake, comments stop being a reliable thermometer. Back in February 2024, after Navalny's death, the analytics firm Cyabra estimated that 29% of the profiles interacting with Biden's post and 25% of those under Trudeau's were inauthentic ([GZERO](https://gzeromedia.com/amp/tracking-anti-navalny-bot-armies-2667343614)). These are estimates, but they make the point: in some discussions, one commenter in four may not be who they seem.

## We only know because someone is reading

There's one aspect of this story that bothers me more than the bots themselves.

Everything I've described, we know because **OpenAI reads how its models are used**. It saw the Persian prompts, the order not to mention Xi, the internal report uploaded by mistake. It's a form of surveillance, and in this case it protected everyone. But it only works as long as the propagandists use an American cloud service.

OpenAI says so itself: China's strategy for "cyber special operations" already favors **open-weight models running locally**. A model on your own server reports to no one. I've always defended local models, and I still think they're the real alternative to the power of the big companies: I wrote about it in the post on [DeepSeek's ban in the United States](/en/posts/deepseek-us-ban-backdoor-weights-chinese-models-2026/). But it has to be said plainly that the same freedom applies to everyone, including people who write fake comments for a living. The cases we see are those of operators lazy enough to use ChatGPT.

It's the same theme as my [post on vibe coding](/en/posts/vibe-coding-risks-security-skills-ai-costs-2026/), seen from the other side. There, Moltbook was a social network where AI agents openly talked to each other. Here the agents talk among us, and do everything they can not to be noticed.

## What I think

I don't think the answer is to stop reading comments, or to see a bot in everyone who disagrees with us. That would be propaganda's real victory: convincing us that anyone who disagrees is fake.

A few habits do help, though:

- **Comment counts are not a poll.** Thirty similar replies under a post tell you how many replies there are, not how many people think that way.
- **Look at the profile, not just the sentence.** Recently created, few followers, only replies and no posts of its own, always on the same topics: it's a clue, not proof.
- **Be wary of waves.** Many different accounts saying the same thing in different words at the same moment is the typical sign of this work: AI now makes the words different, but not the timing.
- **Check who's behind the "institute".** A think tank nobody has heard of, with an index that always rewards the same country, deserves a search before a share.

Platforms and AI companies have to do their part, and partly they do. But we're the ones reading the comment thermometer, and it's worth remembering that someone is trying to get their hands on it.

