Manufacturing Credibility: Russia Used ChatGPT to Build a Think Tank That Never Existed

We have been told to worry that AI will flood the internet with propaganda — an ocean of cheap, fake posts. This week OpenAI revealed something more sophisticated and more unsettling. A Russia-based operation used ChatGPT not to generate noise, but to manufacture legitimacy: a fake think tank, invented experts, borrowed academic authority, and a bogus index rating the world in Moscow's favour. The danger was never going to be the volume. It was always going to be the credibility.

What OpenAI Found

On Tuesday, 25 August, OpenAI announced it had banned a cluster of ChatGPT accounts originating in Russia that were running a covert influence campaign. The operators worked in Russian, routed their access through VPNs to evade OpenAI's ban on use from Russia, and used the model to draft social media content — mostly in English — for Substack, Telegram, X, Facebook, and LinkedIn. Tellingly, they instructed ChatGPT to strip out any linguistic markers that might reveal the authors were Russian speakers.

At the centre sat a fabricated institution: the International Burke Institute, a self-described expert community claiming a street address in Israel, its website registered in early 2025. The site listed world-class scholars among its experts, including names as prominent as Francis Fukuyama and Noam Chomsky. It was almost entirely hollow. Of a sample of 36 articles attributed to its experts, OpenAI found that 34 had simply been copied from elsewhere on the internet, some misattributed, some apparently machine-translated from a Slavic language. And it published a "sovereignty index" — a metric engineered to cast Russia favourably while denigrating Western countries, with reports singling out France, Germany, the United States, and the European Union.

The Real Innovation: Credibility, Not Volume

Here is the line from OpenAI's own analysis that matters most, and it reframes the entire threat. The company said the significance of the operation lay not in the audience it reached — which was relatively small — but in the infrastructure it had built. Read that again, because it inverts the fear we have been carrying. The worry was that AI would let propagandists produce more. The reality is that AI lets them produce credibility.

Think about what used to be expensive in a disinformation operation. Not writing posts — that was always cheap. What was costly was the scaffolding of legitimacy: a plausible institution, a roster of credentialed-sounding experts, a body of published research, a proprietary-looking analytical framework that lends numbers an air of objectivity. Building that by hand took time, money, and skill. AI collapses that cost to almost nothing. In an afternoon, a handful of operators can conjure an institute, populate it with stolen scholarship, borrow the names of real thinkers, and generate an official-looking index to launder a political agenda into the appearance of research. The machine did not make the lie louder. It made the lie look respectable.

Tool and Guard, in the Same Company

There is an uncomfortable structural fact running underneath this good-news story. The same company whose model was used to build the operation is the company that detected and dismantled it. OpenAI deserves genuine credit — this is careful, valuable work, and it is not the first such campaign it has disrupted, having taken down Russian, Chinese, Iranian, and North Korean operations before. But notice the dependency. Our defence against AI-enabled influence campaigns currently rests, to a remarkable degree, on the voluntary vigilance of the private firms that build the tools being abused. They are both the armoury and the guard. When the same entity supplies the weapon and polices its use, the public has no independent way to know how much is being caught and how much is not.

Claim and Counter-Claim

The reassuring reading is legitimate. This operation was caught, its reach was small, and its output bore the tell-tale seams of machine generation — clumsy, non-idiomatic English that a careful reader would clock. OpenAI's threat-intelligence work is real and effective, and the episode shows the guardrails and detection systems functioning as intended. This is a system working, not failing.

The alarming reading is harder to shake, and it is the honest one. Detection this time does not guarantee detection next time; the operators will learn, the output will smooth, the seams will vanish. The point of this campaign was never its current audience — it was the template. A cheap, replicable method for manufacturing institutional credibility has now been demonstrated, and templates spread. The immediate impact was small; the precedent is not. And the deepest problem is not solved by banning a few accounts: once it is trivially cheap to fake the form of expertise — the institute, the index, the credentialed voice — the corrosion is not just more lies, but a general collapse of trust in the real institutions the fakes imitate. The honest synthesis: the incident was contained, and it is a rehearsal for something that will be much harder to contain.