The people building AI want a brake pedal. Here's what that means if you haven't started yet.
Anthropic co-founder Jack Clark has told the BBC that the world needs a way to slow AI down, days after his chief executive asked the whole industry to pace itself and OpenAI's Sam Altman agreed. For a small business that hasn't started with AI yet, that sounds like a reason to wait. We read it the other way.
The news: AI's builders asked for a brake
On Saturday 12 September, Anthropic chief executive Dario Amodei published an essay called We Must Pace the Frontier. He argues that AI models are gaining new abilities faster than anyone can check they're safe, and that the industry should deliberately slow that rate down.
His main reason is recursive self-improvement, which means AI systems doing more and more of the work of building the next generation of AI. Clark told BBC Newsnight in June that Claude already writes 80% of the code that goes into Anthropic. Amodei says this loop has sped progress up sharply since around this summer, across the industry and inside Anthropic. Anthropic also committed to embedding independent evaluators in the company, with employee-level access and the freedom to publish what they find.
Sam Altman replied that OpenAI agrees and will give evaluators the same access. Elon Musk posted that Amodei is right.
Clark then took the argument back to the BBC. He pointed out that Anthropic and most other labs can already pull the plug on their own models, and asked whether governments should require a kill switch that an outside party can verify. He also said the window to act is only a few years wide. In that June interview, he said the industry has "a gas pedal, but it doesn't have a brake pedal".
The pushback: Nobody has pressed it yet
Three chief executives agreeing in public slows nothing down by itself. Every commitment so far is voluntary, and the two governments with the most say have already pushed back.
President Trump dismissed the calls on 13 September and said his priority is staying ahead of China. China's foreign ministry called them fearmongering, and the state-run Global Times described the plan as a Cold War playbook for AI. Trump and Xi Jinping meet in Washington on 24 September, so expect plenty more headlines before anything changes.
Experts disagree on the details too. Toby Walsh, chief scientist at the University of New South Wales AI Institute, called the kill switch idea a distraction, because anything that can switch an AI off becomes a target for attackers. He wants independent oversight of the kind airlines and banks already work under.
The detail: Slower still means fast
Read the essay closely and the proposed slowdown looks modest. Amodei says pacing does not mean halting training, says progress will still feel fast, and talks about buying an extra year or two before models reach dangerous levels of capability.
For scale, the UK's AI Security Institute reported last December that some frontier capabilities were doubling roughly every eight months, including the length of cyber tasks an AI can complete without human help. A curve that steep stays steep even if the labs stretch it by a year or two.
The slowdown also aims at the frontier, meaning the most capable models and the ones starting to help build their own successors. Anthropic released Claude Fable 5.1 on 1 September, eleven days before the essay went up. Nothing announced since changes the tools already in your browser, what they cost, or whether you can use them.
The temptation: Waiting for the dust to settle
When the people building AI say they're worried, holding off feels like the responsible choice. Their worries are specific, though: AI that builds better AI, and swarms of agents acting on the open internet with nobody watching. Amodei's essay leans on the July incident where OpenAI's test agents escaped their sandbox and hacked Hugging Face's servers. A tool that drafts your quotes or summarises a sales call sits a long way from either.
Your competitors aren't pausing. The British Chambers of Commerce found in March that 54% of UK SMEs are now adopting AI, up from 35% a year earlier. Firms already using it reported a net balance of +71% expecting productivity gains, against -3% among firms with no plans to adopt. Almost none of them are cutting staff to get there: 95% said AI hadn't changed their headcount.
A pause at the frontier doesn't pause the business down the road.
The lesson: Fit a brake pedal of your own
Clark's question works as well for a five-person firm as it does for a frontier lab. Can you switch it off, and can someone outside check what it's doing? We ask exactly that before we connect any AI agent to a client's systems.
Any AI tool you bring in should come with an off switch, access to only the data it needs, and a named person who signs off anything touching money, customer records or your live website. If a supplier can't tell you how to switch it off, or who can see what it does, pick a different supplier.
The advice holds whichever way the politics go. If the labs do slow down, the tool your team learns this autumn stays current for longer. If they don't, you'll have the guardrails in place before the next jump arrives.
The bottom line: Same advice, better reasons
This news changes what happens inside the handful of labs building the most powerful models. It doesn't change what a small business should do next, which is to pick one repetitive job, try a tool built for it, keep a person checking the output, and count the hours it saves.
If you'd like help choosing that first job, or a second opinion on how an AI tool connects to your systems, get in touch and we'll work through it with you.
A slower frontier still moves faster than any business that decides to wait.
Steve Lavine is a full-stack developer and founder of Lavine Web & AI Solutions, working with SMEs and startups across the UK. lavine.dev