AI Slowdown? Only if the world agrees, and how to keep up if it doesn’t

Calls to slow the pace of AI development are no longer coming solely from academics, campaigners, or commentators. Increasingly, they’re coming from the people building the technology itself.

The latest wave of debate was sparked by the resignation of former Anthropic researcher Jacob Coxon over concerns about AI governance and the potential dangers of ‘self-improving superintelligence’. This time, big names have joined the discussion. Renewed warnings from Anthropic CEO Dario Amodei, and reports that Sam Altman and Elon Musk share his concerns about the current pace of development, have added weight to the conversation.

The question is whether slowing down is actually possible.

In my view, it is, but only if the world’s major powers agree to do it together.

The problem with slowing down alone

AI development has become a global strategic competition.

Companies are racing to build more capable models. Governments see AI as an economic, technological, and geopolitical priority. Investors expect growth. Customers expect innovation. The incentives all point in one direction: faster.

That creates a problem for any organisation that believes greater caution is needed.

A publicly accountable company can’t just decide to move significantly slower than its competitors without consequences. If one organisation pauses development while others continue, market share, investment, talent, and competitive advantage can quickly shift elsewhere.

The same challenge exists at an international level.

If one country introduces restrictions that slow AI development while others continue at full speed, it risks putting its businesses and economy at a disadvantage. Governments know this, and businesses know it too.

As a result, even leaders who support a slower pace of development may feel unable to act unless legislation creates a framework that applies more broadly. Legal, commercial, and competitive pressures all encourage organisations to maintain momentum.

That is why meaningful restraint is unlikely to emerge through voluntary action alone.

International cooperation is imperative

The only realistic path to an AI slowdown is an international agreement between the major AI powers, particularly the United States and China.

Without that cooperation, any attempt to slow progress will remain fragmented and temporary. No government or company wants to be the one left behind while competitors move ahead.

History offers some useful parallels. International agreements have previously been used to manage technologies with significant global implications. AI presents different challenges, but the principle remains similar. Shared risks require shared rules.

If such agreements were ever pursued, enforcement would be critical.

One interesting aspect of AI, compared with nuclear technology, for example, is that the bottleneck is not scarce raw materials. With nuclear weapons, monitoring focuses on access to materials such as uranium. With advanced AI, the bottleneck is increasingly compute.

That may actually make oversight more practical than many people assume.

A future regulatory framework could distinguish between data centres used to train frontier models and those used to run existing models for customers. If oversight focused on the facilities responsible for training the most powerful systems, the number of sites requiring inspection would be relatively limited.

None of this would be simple, but it demonstrates that discussions about governance don’t have to be purely theoretical.

The security concerns driving the debate

For many advocates of a slowdown, the central concern is security.

They worry that increasingly capable systems may eventually exceed our ability to govern them safely. While there’s considerable debate about the likelihood and timescales involved, concerns about advanced AI are being taken seriously by some of the people closest to the technology. Still, I believe that a measured response is important.

There’s little value in either panic or complacency. Responsible leadership requires acknowledging both the opportunities and the risks without drifting into sensationalism.

Yet focusing exclusively on long-term security risks may cause us to overlook another challenge that has already started to emerge.

The workforce challenge is more immediate

While global safety concerns dominate many headlines, I believe the impact on work deserves equal attention.

Today, most organisations are still using AI as a form of thought partnership. Employees use copilots and assistants to support research, planning, communication, and decision-making. Productivity improves, and people become more effective.

Those benefits are significant, but they’re largely one-time gains.

The bigger transformation comes when organisations move beyond assistance and towards agentic automation. At that stage, AI systems begin performing increasingly complex tasks with human oversight rather than human execution.

That shift could dramatically change workforce requirements.

It’s not that people are incapable of adapting. History has repeatedly shown us that workers can learn new skills and move into new forms of work. The real question is whether education, training, and reskilling systems can keep pace with the speed of change.

If they can’t, we could face a troubling paradox: high levels of unemployment alongside large numbers of unfilled vacancies because the available jobs require skills that many workers don’t yet possess.

Preparing for change

Organisations have a responsibility to prepare for this transition.

Workforce planning should happen alongside AI implementation, not after it. Employers should identify future skill requirements early and invest in helping existing employees move into emerging roles. In many cases, retaining and reskilling talented people will be more effective than competing in costly recruitment markets for scarce expertise.

At an individual level, the best protection against this staggering pace of change is developing the habit of learning.

The subject matters less than the habit itself. Just as regular exercise supports long-term health, continuous learning builds resilience. People who are comfortable acquiring new skills will be better positioned to adapt as technology evolves.

The opportunity we should not forget

Amid discussions about safety and governance, it’s important not to lose sight of what AI could achieve.

The technology has the potential to help address challenges that have frustrated humanity for centuries. Teachers could provide more personalised support. Healthcare professionals could make better-informed decisions. Scientists could accelerate discovery. Climate researchers could analyse complex systems at unprecedented scale. Public services could become more effective and responsive.

The potential benefits are extraordinary. The goal shouldn’t be to stop progress, but to ensure progress happens at a pace that societies, governments, businesses, and workers can realistically absorb.

If the world’s leading AI nations can find common ground, a managed slowdown may be possible. If they can’t, competitive pressures will continue to push development forward at maximum speed. In that case, the challenge will be to collectively create the conditions for that advancement to remain safe, sustainable, and beneficial for everyone.

Ready to build your AI skills? Explore our training courses and learn how to work effectively and safely with emerging AI and automation technologies, automation, and AI-powered tools.