Your employer builds the thing you want to slow down

More than 1,200 researchers at OpenAI, Anthropic, Google, and Meta recently signed an open letter. They are not asking for better pay or fewer meetings. They are asking governments to coordinate tools that could slow down AI development. Tools their own employers have no interest in deploying.

That is the contradiction in one sentence. The people who best understand what these systems can do are employed by the companies most incentivised to build faster. And the only lever they have is a public letter addressed to politicians who, by and large, are still trying to understand what a large language model actually is.

The collective action problem no one wants to name

Here is a structural fact that tends to get buried in the ethics discourse: no single company will slow down voluntarily if it believes its competitors will not. This is not cynicism. It is the basic logic of competitive markets, and it applies with unusual sharpness to AI development, where a six-month lag in capability can mean losing a product cycle, a talent cohort, or a major enterprise contract.

The researchers signing the letter understand this. That is precisely why they are asking for coordination at the government level rather than appealing to their employers' better instincts. Collective action problems require collective mechanisms. A voluntary pause is not a pause. It is a unilateral slowdown that benefits whoever ignores it.

What the letter implicitly acknowledges is that internal safety teams, model cards, and responsible AI principles, however well-intentioned, are insufficient structural answers to a structural problem. They are good-faith gestures operating inside a competitive frame that neutralises them.

What governments are actually being asked to do

The letter does not ask for a ban. It asks for coordination tools: mechanisms that could, in theory, give regulators meaningful visibility into training runs, compute thresholds, and capability evaluations before deployment rather than after.

This is a narrower and more technically grounded ask than the 2023 open letter that called for a six-month moratorium and was signed by figures ranging from Elon Musk to academics with no direct lab affiliation. The current signatories are, by definition, insiders. They are not critics of the industry. They are people writing the code, running the evaluations, and shipping the products. That distinction matters.

But the gap between what is being asked and what governments can realistically deliver is wide. Compute governance, for instance, requires tracking the supply chain of high-end GPUs globally, something that involves export controls, chip manufacturers, and cloud providers across multiple jurisdictions. The EU AI Act addresses some of this. The US Executive Order on AI from October 2023 required companies to report on training runs above a certain compute threshold. These are not nothing. They are also not coordination at the scale the letter envisions.

The honest answer is that no government currently has the technical capacity, legal authority, or geopolitical alignment to implement what the researchers are describing. Which is not an argument against trying. It is an argument for starting earlier, not later.

Why this matters beyond the AI safety community

If you run a product team, an agency, or a startup that is actively integrating AI into your work, you might read a story like this and conclude it has nothing to do with you. The existential risk debates feel distant. The governance discussions feel abstract. You have a sprint to ship.

I think that framing is a mistake, and here is why.

The speed at which capabilities are advancing is already creating practical problems for teams building on top of these systems. A model behaviour you designed around in Q1 may not exist in Q3. An API you integrated may be deprecated, upgraded, or repriced with a few weeks' notice. The regulatory environment in your market may shift faster than your product roadmap can accommodate. These are not hypothetical risks. They are operational realities for anyone who has tried to build a durable product on a foundation that is itself moving.

A credible governance framework would not just reduce tail risks for humanity in the abstract. It would reduce uncertainty for every team building on top of these systems. Predictable capability curves and stable deployment windows are good for product development. Regulatory clarity is good for investment. Slowing the release cadence, even modestly, would give builders more time to understand what they are actually working with.

The researchers are not asking for slower AI because they are afraid of it. Most of them are asking because they have seen enough of the internals to know that the current pace of deployment is outrunning the capacity to evaluate, understand, and govern what is being shipped.

The agency angle

For studios and agencies, this contradiction shows up in a specific way. We are being asked by clients to build faster with AI, while the people who build the AI are asking someone to slow down. Both pressures are real. Both come from people acting rationally inside their own context.

The practical response is not to wait for governance to catch up. It is to build with assumptions that are more conservative than the optimistic scenario. Design for model substitutability rather than coupling tightly to one provider's API. Treat capability benchmarks as snapshots, not contracts. Keep a human in the loop not because regulations require it, but because the cost of a failure at scale is higher than the cost of the check.

This is not a counsel of caution for its own sake. It is basic architecture thinking applied to a genuinely unstable environment. The researchers signing that letter are not wrong that something structural needs to change. But structural change takes time. In the meantime, the teams building on top of these systems have to make design decisions that account for the instability, not assume it away.

The contradiction at the heart of the field is not going to resolve itself cleanly. No letter will do that. What it can do is shift the Overton window on what governments and regulators treat as reasonable to ask of frontier labs. That is slow, unglamorous work. It is also the kind of work that tends to matter more than it looks like it does at the time.

Where that leaves us

The open letter is a symptom, not a solution. What it reveals is that the people closest to the technology no longer trust that competitive dynamics alone will produce acceptable outcomes. That is a significant signal, regardless of what you think about the specific policy asks.

For anyone building products, strategies, or organisations around AI right now, the relevant question is not whether you agree with the letter. It is whether your planning assumptions account for a world where the current pace of development is, at some point, meaningfully constrained. Either by regulation, by geopolitical friction, by compute limits, or by a market event that changes the risk calculus for deployers.

None of those scenarios are remote. Some of them are already in motion. The researchers who signed the letter are not asking you to stop building. They are asking someone with actual authority to create the conditions in which building responsibly is not a competitive disadvantage.

That is a reasonable thing to want. The fact that it requires asking governments rather than employers says everything about where the field actually is.