Pacing the Frontier Won't Pace What's Already Here

What last weekend's rare agreement between three rival AI chief executives actually leaves unanswered

By Louize Clark

Something unusual happened in AI this weekend. Three people whose businesses depend on pushing the frontier forward publicly agreed that perhaps the frontier needs pacing.

The obvious question is whether they mean it.

I'm interested in a different one.

On Saturday, Anthropic's Dario Amodei published an essay called “We Must Pace the Frontier,” arguing that AI companies should deliberately slow how quickly they push forward the capabilities of their most advanced models not because progress should stop, but because capability has started outrunning the industry's ability to understand what that capability actually does. Within hours, Elon Musk said he agreed. Sam Altman backed the substance of it too, and said the question was already being discussed seriously inside OpenAI.

Markets reacted on Monday. President Trump pushed back, arguing that slowing down simply hands the advantage to China. China's own state media dismissed the warning as fear-mongering.

All of that is real, and all of it matters. But it isn't actually the question I want to spend this piece on.

The question everyone's asking

Most of the coverage this week is circling the same thing: will they actually do it? Will three companies simultaneously racing each other for market share, for talent, for the next funding round, genuinely hold themselves back? Amodei's own proposal is careful on this point, independent evaluators inside the labs first, broader industry cooperation second, international agreement a distant and much harder third, precisely because he knows unilateral restraint is close to unenforceable in a competitive market. Altman was careful to say pacing doesn't mean stopping. Nobody involved is promising to slow down in a way that actually costs them the race.

That's a fair question to sit with. I don't think it's the one that matters most to the organisations I actually work with, though. Because even in the world where this succeeds, where the frontier genuinely does pace itself, where independent evaluators get real access, where some version of international coordination eventually holds I don't think that solves the problem most businesses are already living inside.

Two different speeds

I think there are two speeds at work here, and they are not the same speed.

One is the speed at which frontier labs build new capability. That's the one everyone's talking about right now, and it's the one Amodei's essay is addressed to.

The other is the speed at which everyone else absorbs what's already been built. That one hasn't slowed down.

Imagine, for a moment, that the frontier genuinely does pace itself starting tomorrow. Model releases slow. Capability jumps get smaller and more carefully evaluated. Suppose it works exactly as intended.

Microsoft still has AI running through every layer of the products your organisation already uses. Your CRM still updates itself on its own schedule. The employee who found a tool that saves her an hour a day is still using it. The workflow that quietly reorganised itself around an AI recommendation eighteen months ago doesn't reorganise itself back. None of that pauses because three chief executives agreed, over a weekend, to build more carefully next.

Slowing the frontier doesn't slow what's already installed.

I'd go a step further, actually, because “installed” undersells what's already happened. I was at a parliamentary discussion on AI in Westminster last night, one of a series run through the All-Party Parliamentary Group on AI, where the UCL economist Dr Cecilia Rikap presented the argument from her new book, The Rulers. Her research, built on more than a hundred interviews, alongside patent, acquisition and investment data traces how a small number of cloud providers have become something closer to infrastructural landlords for the rest of the economy: the businesses building on top of them don't just rent their compute, they increasingly route knowledge, data and decision-making through infrastructure those few companies control.

If that's right, then what's already installed isn't only the habits and workflows sitting inside individual organisations. It's the underlying architecture those habits and workflows now depend on. Slowing the frontier doesn't unwind that either. Three companies agreeing to build their next models more carefully changes very little about who already holds the infrastructure everyone else is standing on.

Six years without a settled moment

I think this is where the pandemic comparison actually earns its place and I want to be careful how I use it, because it's easy to reach for and easy to overstate.

I'm not saying AI has been forced on us the way lockdown was. What I'm pointing at is narrower than that, and I think more useful: what happened to our adaptation time.

In March 2020, behaviour that would ordinarily have taken years to become normal video GP appointments, remote banking for people who'd never used it, entire workforces working from a kitchen table became normal in a matter of weeks. That was phenomenal, and it worked, largely because people had no real choice but to make it work.

What's said less often is that we never really got a settled period after that. Cloud migration accelerated. Hybrid work rewired how organisations operate. Then generative AI arrived, then copilots got folded into the tools people were already using, then agents started doing things without being asked step by step. Six years, and at no point has there been a genuine pause to check whether what changed was actually being absorbed properly, rather than simply being coped with.

I think we've quietly confused our ability to adapt quickly with our ability to absorb change safely. They are not the same skill and only one of them is currently being tested.

The race underneath the race

There's a version of this argument that stays at the level of labs and nations, and I don't think that's where it should stay, because the same dynamic is playing out one level down, inside almost every organisation I speak to.

A board hears that a competitor has launched an AI initiative. Nobody wants to be the business that's visibly behind. Something gets adopted, not because someone worked out exactly where it fits and what it changes, but because the alternative, being seen to have done nothing, felt like the bigger risk.

It's worth being precise about what “adopted” actually means here, because it usually isn't one thing. Sometimes a business deliberately chooses a tool and signs for it. Sometimes the capability arrives because a supplier updated a platform the business already trusted. Sometimes an employee brings something in on their own account because it made their week easier. Sometimes nobody in the organisation could tell you, if you asked them today, which of their systems now has AI running somewhere inside it. An organisation can be racing to adopt AI in the boardroom while simultaneously inheriting AI it never consciously raced towards at all, through the ordinary churn of the software it already runs on. Both are happening at once, in most businesses, right now.

That's the race underneath the race. Labs are racing each other to build it. Nations are racing not to be left without it. Businesses are racing to adopt it before a competitor does. Employees are quietly racing to use it before it becomes obvious they haven't. At every one of those layers, the question actually driving behaviour is essentially the same: what happens if everyone else moves and we don't?

There's a question worth sitting inside that race rather than simply naming it, though, and it came up more than once in that same room in Westminster: why does every organisation assume it needs the frontier at all? A hospital administration system doesn't obviously need the most capable reasoning model available. Neither does a routine logistics workflow, or a local authority processing correspondence, or most of what a typical SME actually does day to day. Capability and suitability have quietly become the same word in how most organisations talk about this, and they were never the same thing. The more useful questions are usually smaller: is this sufficiently capable for the task in front of it, what does it need to run, where does the data go, and what happens to us if the provider changes it, or we need to leave?

If what's actually driving adoption is fear of being left behind, rather than a considered view of where a tool belongs and how much of it a task actually requires, then pacing the frontier doesn't touch any of that. That fear operates independently of how fast the underlying technology happens to be moving, and independently of whether the organisation needed the frontier in the first place.

Losing visibility before losing control

I want to be careful with language here, because “losing control of AI” is doing a lot of work in the current conversation, and most of that work is happening at a scale, systems improving themselves, autonomous behaviour outrunning oversight that is genuinely Amodei's territory to worry about, not mine.

The version I actually see, inside real organisations, most weeks, is quieter and arrives much earlier. An organisation doesn't need an autonomous system slipping its constraints for something to go wrong. It only needs to stop knowing what it's already standing on: which tools now have AI embedded in them that didn't a year ago, which decisions are being shaped by a system nobody evaluated for that purpose, which process that used to require someone to understand why a thing was done a certain way has quietly become a workflow the organisation can still run without anyone left who could explain it.

That last point is worth sitting with. The risk isn't really that a platform “remembers” something a person forgot, however it can feel that way. It's that once a task becomes AI-mediated, an organisation can go on producing exactly the same output for years while the understanding that used to sit behind that output why it's done this way, what happens if it's wrong, what to check quietly stops living anywhere at all.

Colonel Sean Lamb MBE, Deputy Head Capability Development, UK Space Command put the wider version of this better than I've heard it put before: trust is not the same as accuracy. His point, as I understood it, was that trust in these systems doesn't stay with the person using them. It extends all the way back to the provider, and the burden of checking the machine currently sits with the wrong person in that chain. An employee can't meaningfully verify every output a system gives them. A procurement team can't continuously inspect every vendor's latest update. A board can't personally understand every model change a supplier ships this quarter. At some point, “there must always be a human in the loop” stops meaning very much, if that human has neither the time, the information, nor the standing to actually exercise judgement over what they're being asked to sign off.

That's not a control problem yet. It's a visibility problem, sitting on top of an assurance problem. But visibility is what control depends on, further down the line you cannot meaningfully govern what you've already lost sight of, and an organisation that has stopped being able to answer “what changed, and did we choose it?” doesn't get that answer back simply because the labs supplying its tools have agreed to build more carefully.

The ability to leave, not just the ability to build

This is also where the frontier conversation connects to a question Britain keeps having separately, about sovereign AI. Most of that conversation is still about compute: British data centres, British-trained models, chip supply, energy. All necessary. None of it is really the point I want to make here.

Tony Reeves, who leads AI work in defence at Deloitte and has long argued that organisations need a change of mindset rather than a change of toolset, made a related point on the same panel, about sovereignty, foreign interference and whether Britain can genuinely compete with the scale of investment behind programmes like the US's Stargate. The risk isn't only that Britain lacks frontier capability of its own. It's that years of digital transformation have already been built on someone else's. An organisation, or a country, can build sovereign infrastructure and still discover that its workflows, its skills and its accumulated knowledge were built around infrastructure it doesn't own and didn't choose.

Sovereignty, on this view, isn't only the ability to build. It's the ability to choose and, just as importantly, the practical ability to leave, if you needed to. An organisation can be fully compliant and still be completely dependent. It can have a governance policy and still lack visibility into what it's actually running on. And it can, in principle, have a choice of providers on paper while discovering that years of integrations, trained staff and rebuilt workflows have made changing provider something closer to impossible in practice. That's not a technology problem. It's an operational one, and it's the one most sovereignty conversations still skip past.

The dilemma nobody can solve alone

I don't think the honest answer to any of this is “everyone should just slow down,” because I don't think that's actually available to anyone, at any level.

Amodei's own proposal accepts this. A framework that could look like unilateral disarmament to a competitor, or to Beijing, isn't one Washington is going to embrace easily and the White House's response over the weekend more or less confirmed that. Even people who privately agree the race should slow can't easily be the ones who slow first.

The same dilemma sits inside individual organisations, just quieter. A leadership team might privately think they need six months to properly understand what they've already adopted before adding anything else. Then a competitor announces a major AI programme, and that six months starts to feel like a luxury nobody can afford. It's the same shape of problem, just without the geopolitics attached to it.

What would we do with the time?

Perhaps that's the question underneath all of this.

If the frontier does slow even slightly, what exactly are we hoping the extra time gives us?

More safety research, certainly. Better regulation and stronger international coordination, hopefully. But I think businesses and institutions need some of that time too.

Not to wait for AI to become safer before using it. That moment may never arrive neatly enough to recognise. But to understand what has already changed while we've been busy keeping up: the systems we've become dependent on, the capabilities we've inherited without deliberately choosing them, the knowledge that may have quietly moved out of the organisation, and the choices that are becoming harder to reverse.

Because perhaps the measure of whether we have kept pace with AI isn't how quickly we adopt the next model.

Perhaps it's whether we still understand what we're standing on and whether we still have enough freedom to choose to stand somewhere else.

Louize Clark is the founder of AI Policies UK, Publisher of The AI Law Report and creator of the Invisible AI Infrastructure™ . aipolicies.uk