Claude misused (3/7) : Cheap AI propaganda mostly reached nobody
The short version
A firm in Istanbul built a platform to target voters in all 222 Malaysian constituencies, and Anthropic grades its real reach at two out of six. The operations that worked already owned a broadcaster.
The fear with AI and elections was always volume, a flood of fake people saying fake things at a price too low to stop. Anthropic’s September threat report has nine influence operations in it, and the awkward finding is that the flood mostly arrived and mostly nobody read it.
This is part three of seven, and part two was surveillance. The report grades each operation on the Breakout Scale, a six-step measure researchers use for how far fake content travels, where One means it never left its own cluster of accounts. Most of the nine sit at Two or Three. In Anthropic’s words, “Most of the content we discovered drew little or no authentic engagement.”
Writing the lie got cheap. Getting anyone to hear it still costs a radio station.
The first is next door. I’m in Singapore, so the Malaysia case is the one I can’t shrug at. Anthropic removed an account running a commercial election-manipulation platform aimed at Malaysian voters, and linked it to BBS Bilisim Teknolojileri, a technology company in Istanbul that sold it as a service. The operator’s own documentation called it a “military-grade, AI-driven, real-time political operations ecosystem.” It used real census and electoral data, including millions of voter records, to profile all 222 parliamentary constituencies along the lines of race, religion and royalty. It ran about a thousand fake X accounts, a fake news site called Malaysia Pulse, and invented dossiers against an opposition politician. Anthropic saw one request for a million artificial views on the sitting Prime Minister’s account. The report doesn’t say who the customer was, and it found no evidence that the operator’s pitch to Malaysia’s communications regulator went anywhere. Its reach grade is Two, meaning the content never broke out of its own accounts and sites.
The reaction in Malaysia travelled a lot further than the campaign did. MCA president Wee Ka Siong called it “not basic political manipulation, but serious manipulation using AI,” and asked how millions of voter records got out in the first place. The regulator, MCMC, said it was taking the report seriously and that it wasn’t involved in anything the report described. UMNO information chief Azalina Othman Said had the line I’d keep: “Technology can produce a thousand accounts and a thousand comments. But technology cannot determine the truth or what people actually feel on the ground.”
The second actually got on air. Four accounts used Claude as a sub-editor for Russian state media. A former editor-in-chief of Sputnik Moldova turned Romanian and Moldovan news into Russian-language articles that were published and then echoed between outlets until the same claim looked independently confirmed, including fabricated claims about President Maia Sandu before the September 2025 election. An employee of a Russian state-owned outlet used it for the tickers, captions and voiceover scripts of live broadcasts, and in at least one confirmed case the material went to air. This one worked for a boring reason, which is that the audience already existed. Claude didn’t find anyone new to lie to, it just let one staffer do the work of an editing desk for a channel that was already broadcasting.
The third never needed an audience. An operation Anthropic links to the MEK, an Iranian opposition group in exile, had an AI agent read roughly 8,400 Telegram posts written by a real activist, told it in Persian that it was now that person, and used it to hold live political conversations with his contacts. As far as Anthropic knows, they didn’t realise. The same network scraped more than 500 channels to build profiles of people inside Iran that included their arrest history, in a country where that history can get someone jailed or executed. Its reach grade is also Two, but reach is the wrong measure when the whole trick is one trusted person talking to people who think they know him.
So I’d split the nine into two piles. The fake-crowd operations, like the thousand accounts in Malaysia or the seventy fake news sites another case traced to a single ad agency in France, mostly talked to each other and didn’t get far, and in several cases Anthropic caught them before they’d built an audience. The ones that mattered either borrowed an existing megaphone or skipped the crowd entirely and went one to one. It’s the second pile that isn’t going to show up in anyone’s engagement numbers.
If you only ever scroll, here’s what changes. A thousand accounts agreeing with each other isn’t evidence of anything now, because agreement is the cheap part. Don’t share a “local news” story before you’ve checked how old the site is. Malaysia Pulse’s domain was registered on 10 May 2026, and it was republishing Xinhua and Sputnik copy with the state attribution stripped out. And if someone you know starts messaging about politics in a way that’s slightly off, don’t argue with the text, ring them or ask something only they would know.
The influence section is here, all nine cases of it. Part four is weapons, and it doesn’t get lighter.