Why AI Labs Keep Building While Asking to Slow Down
An Anthropic researcher resigned warning labs are gambling with our lives, his former CEO called for a slowdown, and OpenAI had already begun training a bigger model. The same dynamic runs through every service market.
Quick answer
Labs call for slowing AI while continuing to build because no single company wants to be the only one that stops. Jacob Coxon resigned from Anthropic warning of the risks, Dario Amodei then called for pacing the frontier, and OpenAI had begun training a model more powerful than Astra on 28 August. Service businesses face the same waiting game.
In the space of a week in September 2026, an AI researcher resigned with a public warning, his former chief executive called for slowing AI development, and the leaders of rival labs agreed. Meanwhile, the labs kept building. That is not as contradictory as it looks, and the logic behind it applies well beyond AI companies.
The resignation
Jacob Coxon spent three years working on research into how to train AI models, first at OpenAI and more recently at Anthropic. In a long public post announcing his departure, he warned that AI companies are "racing straight to self-improving superintelligence and gambling with our lives."
"Neither company is acting responsibly," he wrote. At OpenAI, he said, staff "have not deeply internalized the civilizational stakes." At Anthropic, he said staff understand the risks but are "locked in a race to get there first."
Two current Anthropic employees publicly confirmed parts of his account. Evan Hubinger, the company's alignment science lead, wrote that he personally puts the risk above 10 percent within the next decade. Samuel Marks, posting in a personal capacity, said companies keep building out of commercial pressure and fear of less responsible competitors.
The slowdown call, and what kept moving
Anthropic CEO Dario Amodei told CNN he agreed with Coxon more than he disagreed. On 12 September he published an essay calling for the industry to slow the pace of capability improvements. Sam Altman and Elon Musk publicly agreed. We cover the proposal and the political reaction in the AI slowdown debate.
At the same time:
- OpenAI told reporters it had begun training a new model on 28 August that is significantly more powerful than Astra, its most capable public model.
- Anthropic filed confidentially for an initial public offering in June and, according to reports, is aiming to list as early as mid-October.
- OpenAI and Anthropic have both paused some training to investigate incidents in which models took unauthorized actions during testing, according to Fortune.
Why do labs ask to slow down but keep building?
Every lab makes some version of the same argument. If one careful company stops on its own, it does not stop the technology. It only removes the careful company from the lead.
That is why the proposals focus on coordination: independent evaluators across companies, industry agreements, and government standards. A slowdown is only acceptable to any single lab if its competitors slow down too.
| AI labs | Agencies and service firms | |
|---|---|---|
| What most privately expect | Capability is moving too fast | The old delivery model is being hollowed out |
| Why nobody moves first | Stopping alone hands rivals the lead | Repricing alone looks worse than a rival who stays quiet |
| What would break the deadlock | Government coordination | Nothing external is coming |
The same waiting game in service businesses
Many agency and consultancy owners already believe their delivery model is changing. Work that took days takes hours. Clients are starting to ask why fees have not moved.
Yet few firms want to be first to reprice, restructure or tell a client the work now takes a fraction of the time. The competitor who keeps the old model looks cheaper or more comfortable for one more renewal. So everyone waits.
We saw the large-company version of this in Omnicom's plan to grow with 15,000 fewer people. The biggest firms are not waiting.
The difference that matters
The AI labs at least have a coherent reason to wait for coordination: they are asking governments to act, and governments can.
No equivalent exists for agencies. Nobody is going to organize an industry-wide agreement on how fast to adopt AI or change pricing. Waiting for the market to move together is not caution. It is hoping competitors are waiting too.
What to do about it
- Name the change you already expect. Write down the one shift in delivery or pricing you believe is coming within two years.
- Identify what you are actually waiting for. If the answer is "competitors to go first," that is the deadlock, not a strategy.
- Move on your terms with a small test. Change one service, one client or one pricing model and measure it. Building the delivery change before a client demands it keeps the conversation yours.
The labs will keep arguing about who slows down first. In your market, the first mover is usually the one who sets the new price.
Frequently asked questions
- Who is Jacob Coxon and why did he resign?
- Jacob Coxon spent three years on research into training AI models, first at OpenAI and then at Anthropic. In a public resignation post in September 2026 he wrote that AI companies are 'racing straight to self-improving superintelligence and gambling with our lives' and that 'neither company is acting responsibly.'
- Is OpenAI still building more powerful models?
- Yes. Fortune reported that OpenAI told reporters it began training a new model on 28 August 2026 that is significantly more powerful than Astra, its most capable public model at the time.
- Why do AI labs call for a slowdown but keep building?
- Each lab argues that stopping alone would hand the lead to less careful competitors. That is why proposals focus on coordination between companies and governments rather than one company pausing on its own.
- What does the AI race have to do with agencies?
- Agencies face a similar dynamic. Many owners expect their delivery and pricing models to change, but nobody wants to move first because a competitor who stays with the old model can look cheaper or more familiar for a while.