Latest / Minds, Bodies, and Terawatts / Russia Ran a Latin American Think Tank Through an AI Persona Named Mia Clark. The Researchers Did Honest Work and Never Knew Who Signed Their Paychecks. Iran Got a Hundred Articles Past Real Editors Under Seven Bylines That Don't Exist
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, where the fake news turned out to have a real payroll. OpenAI said on October 8 that it shut down a Russian propaganda operation that ran an actual think tank in Latin America, with real employees, through a boss who did not exist.
- 0:17[Guest] Her name was Mia Clark. The operators' own notes show them deciding pay, hiring and firing, and she was the face they used to deal with the staff. The Russian team invented her.
- 0:29[Host] Hold on, the researchers were real people doing real work?
- 0:33[Guest] That's OpenAI's read of the evidence. The staff interviewed experts across the region and wrote papers on things like Brazil's economy, and the site has more than sixty articles that are mostly original. OpenAI says they had no idea they were working for a Russian group.
- 0:50[Host] So where does the AI come in? I assumed this would be a flood of machine-written fakes.
- 0:56[Guest] Mostly it didn't write the fakes. The operators used ChatGPT above all to draft status reports to a superior up the chain, which is why OpenAI could read the whole playbook. Creating propaganda was a small slice of the workload.
- 1:11[Host] Status reports. So the robot's main job in a propaganda ring was paperwork.
- 1:15[Guest] Padded paperwork. The reports took credit for things the team had nothing to do with, like Argentina's decades-old position on the Falklands. A researcher at Clemson, Darren Linvill, told NPR the underlings were using AI to do less work, and that is what got them caught.
- 1:32[Host] What did they actually pull off, then, if they were inflating it?
- 1:37[Guest] Enough that OpenAI rates it a five on the six-step scale researchers use for how far propaganda travels, the first five it has seen in thirty operations. The one that stands out happened in Peru in the spring. A fake email dressed up as the Lima education office told schools to hold Ukraine-themed events and to feature Stepan Bandera, a nationalist from World War Two whom Poland and Russia associate with wartime atrocities.
- 2:03[Host] And schools did it?
- 2:05[Guest] By the operators' own account, some wrote back to the fake address confirming they had, with photos. Then stories about Ukraine exporting extremism ran in Peru, in Poland, even in Hungary, and a Polish member of the European Parliament suggested barring anyone with anti-Polish views. A fake email in Lima moved a politician in Poland.
- 2:25[Host] Okay, that's Russia. And the second operation?
- 2:27[Guest] Iran, and this one targets readers like you. Seven journalists who do not exist pitched articles about the US war with Iran to small and mid-sized outlets, and got almost a hundred published across about a dozen of them. OpenAI's models polished the drafts to each outlet's submission rules, wrote the pitch emails, and for one persona, Michael Harrison, wrote the bio.
- 2:50[Host] Which outlets, anything I'd actually read?
- 2:52[Guest] CNN and NPR name the Middle East Monitor in London, which has nearly two million Facebook followers, plus the LA Progressive and the open community section of Daily Kos, where anyone can post. The same crew also pumped out batches of comments on social media, and OpenAI found none of them got any real traction. The reach came through real editors, not through volume, and OpenAI's best guess is a contractor for hire, client unknown.
- 3:19[Host] I'm still stuck on how a persona runs people. Would the staff never ask to meet the boss?
- 3:26[Guest] How many people now have a boss they've only ever seen in a chat window? Remote hiring made a manager who exists only in messages ordinary, and that's the gap this operation walked through. The researchers got paid, got assignments, got reviewed, and never had a face to put on the name on the door.
- 3:46[Host] Fine, but couldn't the outlets have run the articles through an AI detector?
- 3:51[Guest] It would have found nothing in Latin America, because the humans wrote that research in good faith. And the Iranian pieces were human drafts that a model tidied up. What was fake wasn't the prose, it was the name on the door, and no text scanner checks that.
- 4:08[Host] Then what does? Because editors read a pitch, like the writing, and say yes.
- 4:13[Guest] That's Turing's old imitation game, a machine trying to pass as a person in text, and look at what it actually measures. When people are asked to tell a machine from a human, studies keep showing they go on warmth and style, so a persona wins by being charming and on spec. The editor was the judge, the byline was the contestant, and the contest was never about truth.
- 4:39[Host] So the honest people are the ones who got used, the researchers, the editors, the schools. Could any of them have caught it?
- 4:47[Guest] Not from where they stood, because every one of them made a decision in the dark. The researchers took a job without seeing who paid, and the editors ran a piece without seeing who wrote it. Each call looked fine up close, and the lie lived one layer up.
- 5:05[Host] So what's the fix, if detectors don't help? Ban anonymous bylines?
- 5:09[Guest] Not a ban, a rule that whoever pays and whoever assigns the work can be seen. Patrick Deglon in Unscarcity calls it Truth Must Be Seen, no decision in the dark, and Mia Clark was built to break exactly that. For more on false fronts and the imitation game, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.