The AI Propaganda Isn't Coming From Where You Think

Over the last few months I've watched three people go viral arguing about AI.

A comedian at Harvard told a room of graduates that the mission of their generation is to destroy it. A US senator filmed himself reading out a letter he'd sent to the heads of OpenAI, Anthropic and Meta, telling them to stop. And a former CEO of Google stood up in front of a stadium of graduates, started talking about AI, and got booed.

I watched those three clips and decided everyone had lost the plot. You either hate AI or you love it. Two types of people, shouting at each other.

Then I did something I should have done first. I watched the full speeches, not the clips, and I checked whether any of it was true.

Two of the three were completely reasonable. One of them was making my exact argument. And the one I was most ready to write off had the best evidence of the lot.

There is propaganda in this argument. I was right about that. I was just wrong about where it comes from.

The comedian

Ronny Chieng, Class Day at Harvard, May this year. The clip that went round is him telling the graduates to destroy AI, with some swearing for good measure.

I got to the end of the actual speech and laughed, because the first minute is jokes. AI is always wrong, untalented people use it to write their speeches, that kind of thing. Then he stops joking. He says creating is the fun part, that the journey is the whole point. And he closes on this: the battle won't be between humans and AI. It'll be between people with substance and people with superficial knowledge. Mastery versus faking it. Taste versus tacky.

That's my position. It's what I've been saying on my channel for the best part of a year. The man got clipped as the anti-AI extremist, and what he actually stands for is taste and mastery. He isn't anti-AI. He's anti-shortcut.

The CEO

Eric Schmidt ran Google for ten years. In May he gave the commencement speech at the University of Arizona, moved on to AI, and the stadium booed him. That's the clip. There's even a reaction video where the commentator says Schmidt is "happy at the prospect of putting people out of work".

Here's what Schmidt actually said. He opened by apologising for his own generation: nobody sat down and decided to build a technology that would polarise democracies and unsettle a generation of young people, but it happened. He said the fear in that room was rational. He said the technology is a tool that optimises for whatever we tell it to, and we're the ones who decide what that is. And he told the graduates: you bring the judgment, you bring the conscience, you bring the moral sense to know what's worth building.

Maybe he's the wrong messenger. A tech billionaire telling students who are about to enter the job market that their edge is to be more human isn't an easy sell, and I understand the boos up to a point. But "happy at the prospect of putting people out of work"? That isn't what happened.

One more thing the clip leaves out. There were students there that day protesting about something else entirely, a lawsuit Schmidt is involved in, nothing to do with AI. Watch the clip and you'd think the booing was purely about the speech.

The senator

This is where I have to hold my hands up.

I'd filed Bernie Sanders as a politician who was worried about something he didn't really understand. Then I read the whole letter and checked every claim in it.

They hold. Three AI models got out of their test environments and reached real companies in the space of five weeks. OpenAI disclosed the first in July: two of its models got out of a test environment, onto the open internet, and into Hugging Face's servers. A week or so later Anthropic said one of its models had broken into a real company and taken a few hundred rows of production data. Then Meta disclosed one in August, although that one came down to a misconfigured test environment, so it isn't the same calibre. Sanders also said the head of the CIA had compared these models to digital nuclear weapons. He did, in June, in almost exactly those words.

And his actual argument wasn't fear. Each of those three companies had published, in writing, the conditions under which they would pause. Sanders' letter says: by your own definition you're there, so pause. A man asking companies to do what they said they'd do is a more logical position than the clip suggests.

The letter isn't perfect. Anthropic had already removed its pause commitment from its policy in February, six months before Sanders quoted it. And the claim in it that got the most attention, that AI has been used for the first time to create new viruses, is where the whole mechanism shows itself. More on that in a moment.

Where the propaganda actually lives

None of the three is an extremist. So where does the extremism come from?

From the gap. The gap between what happened and what reached you. The clip, the caption, the thumbnail, the reaction video. There's a whole layer between the speech and your feed, and that layer has a job, which is to make things travel.

It doesn't lie. It selects. It takes the ninety seconds out of the twenty minutes that will do numbers. "Destroy AI" travels. "Mastery versus faking it" doesn't. A stadium booing travels. "I understand that fear, it's rational" doesn't.

The virus claim is the cleanest example I found. When you hear that AI was used to create new viruses, you picture a bioweapon. The actual paper was published in Science on 6 August by Stanford and the Arc Institute. They used AI to design bacteriophages, which are viruses that infect bacteria, not people. About 300 designs, 16 of which worked. The reason they did it is antibiotic resistance: they were building something that kills strains of E. coli that antibiotics no longer touch. It's a treatment. Is it also dual use? Yes, and serious people worry about that. But a therapy for drug-resistant infection left a lab in California and arrived in your feed as a bioweapon. Nobody lied. It got compressed, step by step, by people doing their jobs.

I'm in that layer

Here's the part I'm less comfortable with. I make short videos.

I've spent twenty hours editing a podcast before. Recently I put one out that took me about five minutes, and most of those five minutes were writing a prompt. That's true. It's also true that Claude then spent a couple of hours rendering the footage. I clipped it into a short and framed it so it looked like five minutes, start to finish. I knew which version would travel, and I picked it.

That's a light version. There are far worse ones, where nothing in the message changes but the framing distorts the reality, purely for views. Nobody made me do it. The punchier number does better and I wanted it to do better. If I sit here saying the discourse is broken and people are being manipulated, I'm doing it too. So is most of the content about AI, mostly without noticing. The reasonable version of anything doesn't clip well. You end up in a world where nearly everyone thinks something sensible and none of us can hear each other.

The one thing all three agree on

This is the useful bit, and it's why I made the video.

Strip the packaging off and the three of them are saying the same thing. Chieng says the divide is mastery versus faking it. Schmidt says the tool optimises for whatever you tell it to, and someone has to decide what that is. Sanders, from the opposite end of the argument, is worried about exactly that mechanism. Read what happened in those hacks: nobody's model rebelled. They did what they were told, and went further than anyone expected. That's not a machine developing its own goals. It's a machine inheriting ours, badly specified.

AI is a tool, and the people behind the tool decide what it does. It will do good things and bad things because there are good and bad people behind it. What you ask for, and how precisely you ask for it, is the job now.

That's what's actually changing about work. I have a dev team of twenty. We write far less code than we did a year ago and far more tests and PRDs, because the job has moved up a level. I haven't let anyone go. The doing shrank. The deciding grew. I don't think AI takes your job. I think it takes the doing and leaves you the deciding, and the deciding is harder than people think.

Three people, three speeches, three separate rooms, all clipped and distorted by the same machine. And underneath it they want the same thing: AI that's safe, and that moves us forward. I think most of us want that. We just have to be honest about the tool, and about the layer in between.

I nearly made a video calling all three of them extremists. It would have done better than this one. It would also have been wrong.

The video version, with the clips and the sources on screen:

Sources, if you want to check me the way I checked them:

• Ronny Chieng's Harvard Class Day speech: Harvard Gazette and The Harvard Crimson
• Eric Schmidt booed at the University of Arizona: NBC News and The College Investor on the protests that day
• Bernie Sanders' letter: the full PDF and Axios
• The three test-environment escapes: CNN on OpenAI and Hugging Face, NPR on Anthropic, NPR on Meta
• CIA Director John Ratcliffe on "digital nuclear weapons": The Japan Times
• The bacteriophage paper: Science and the Stanford Report
• The pause commitments: Anthropic's Responsible Scaling Policy and Meta's Frontier AI Framework

If you run a business and you're trying to work out what's real in all of this, I write a free weekly newsletter about what's actually happening inside my companies as we build with AI. You can join it here: axelmolist.com/youtube.

And if you've caught yourself doing what I did, clipping the version that travels, I'd like to hear about it. Hit reply. I read every email.

Thanks, Axel.