1,200 AI Lab Workers Just Asked Washington to Slow Down AI — Here's Why

The short version

More than 1,200 employees from OpenAI, Anthropic, Google DeepMind, and Meta signed a letter asking the U.S. government to build a plan to slow AI that can write and improve its own code. It’s rare for people at rival labs to agree on anything, let alone jointly ask for brakes. The signatures are aimed squarely at self-improving systems, not chatbots that help you write emails.

Why did this letter get my attention?

I read a lot of open letters about AI. Most of them are easy to ignore because they come from people with no skin in the game, or from executives doing PR. This one is different, and I sat up when I saw the numbers.

When more than 1,200 employees from OpenAI, Anthropic, Google DeepMind, and Meta sign the same document, that’s not a press release. Those are four companies that compete brutally for talent, funding, and headlines. Getting a thousand-plus of their engineers and researchers to co-sign a request to slow their own field down is like watching four rival kitchens agree the stove is too hot.

The letter, described in the original report from Fortune and the Washington Post, formally backs a government-led plan to slow the pace of AI that can write its own code. Read that twice. The people building the thing are asking the referee to step in.

What exactly are they worried about?

This is the part that matters, and it’s easy to miss if you skim. The signatories aren’t scared of ChatGPT writing your cover letter. They’re worried about self-improving systems: AI that can write, test, and refine its own code with less and less human involvement.

That’s a specific technical fear, not a vibe. Once a model can meaningfully improve itself, the feedback loop gets fast. A system that gets 5% better at coding each cycle, then uses that ability to speed up the next cycle, doesn’t move in a straight line. It compounds. The people closest to the training runs understand that curve better than any of us on the outside, and that’s exactly why their signatures carry weight.

The letter reportedly asks Washington to build a plan to slow that particular capability down. Not ban AI. Not pause everything. Slow the narrow, high-stakes part where a machine improves the machine.

Why does it matter that the workers, not the CEOs, signed it?

Here’s the distinction I keep coming back to. CEOs of AI labs have said cautious-sounding things for years, often while raising billions and shipping faster. Talk is cheap when your equity depends on growth.

But rank-and-file engineers signing a public letter is a different kind of signal. These are people who could get quietly passed over for promotion, or worse, for putting their name on something that embarrasses their employer. When over a thousand of them decide the reputational risk is worth it, they’re telling you the internal conversations are more serious than the polished public messaging suggests.

I trust the person writing the training loop over the person pitching investors. That’s the whole reason this letter feels credible to me instead of like more noise.

Isn’t asking the government to regulate AI a little rich coming from the labs?

Yes, and I won’t pretend otherwise. There’s a real tension here. Some AI companies have publicly asked for regulation while privately lobbying to shape it in their favor, or to lock out smaller competitors. “Please regulate us” can be a moat dressed up as responsibility.

So I hold two thoughts at once. First, the specific ask here is narrow and technical, which makes it harder to read as pure regulatory capture. You don’t build a competitive moat by slowing down the exact capability that would make you the most money. Second, I’d still want to see the actual policy language before cheering, because the gap between “slow down self-improving AI” and a workable law is enormous.

The named tradeoff is this: move too slow and other countries race ahead with fewer guardrails; move too fast and you lose the ability to course-correct at all. There’s no clean answer. Anyone selling you one is lying.

What could a government “slowdown plan” even look like?

The letter asks for a plan, not a specific bill, so the details are still open. But based on how these things usually go, a slowdown plan for self-improving AI would probably touch a few areas.

Compute thresholds are the obvious lever. Training the biggest models takes staggering amounts of specialized chips, and that’s measurable and controllable in a way that software isn’t. A rule might require government notification or review before a training run crosses a certain size.

Evaluation requirements are another. Before deploying a system capable of automated self-improvement, a lab might have to prove it passed specific safety tests, similar to how a new drug clears trials before it reaches pharmacies.

And there’s monitoring of the recursive-improvement loop specifically, which is the technical heart of what the signatories flagged. The honest answer is nobody has fully figured out how to regulate this cleanly yet, which is part of why the workers want a plan built now rather than after something breaks.

Should you, a normal person, actually care?

If you use AI to write emails and make slide decks, this letter changes nothing about your Tuesday. That’s the truth, and I won’t inflate it.

But the longer arc matters for everyone. The tools I recommend, the ones you and I use daily, all sit downstream of what these frontier labs build. If the base models start improving themselves at a pace no human team can supervise, that reshapes the entire field within a couple of years, not decades. The people warning about it aren’t preppers. They’re the ones with badge access.

My honest take: I’m glad this letter exists, I’m skeptical Washington will move fast or smart enough, and I think the fact that 1,200 insiders felt the need to write it at all is the real headline. The message isn’t “AI is scary.” The message is “the people building it don’t feel in control of the pace,” and that’s worth a moment of your attention.

FAQ

Did the AI companies themselves sign this letter? No. The letter was signed by more than 1,200 individual employees across OpenAI, Anthropic, Google DeepMind, and Meta, not by the companies as corporate entities. That distinction is why it reads as a grassroots safety push rather than official corporate policy.

Are they asking to ban or pause AI? No. The letter asks the U.S. government to build a plan to slow the specific capability of AI systems that can write and improve their own code. It targets self-improving systems, not everyday AI tools or chatbots.

Is this the same as previous “pause AI” letters? Not quite. Earlier public letters were often signed by outside academics, executives, or public figures. This one is notable because it comes from a large number of working employees inside the top four labs, giving it more insider credibility.

Will this actually lead to new laws? Unclear. The letter requests a government-led plan, but there’s no bill yet, and the gap between a broad request and enforceable policy is wide. Watch for whether Washington responds with concrete compute or evaluation rules.

Does this affect the AI tools I use today? Not directly. Consumer AI tools are downstream of frontier labs, so nothing changes for daily users right now. The concern is about future systems that improve themselves faster than humans can oversee.