AI Should Slow Down. But Nobody Wants to Be the Only One Who Does

Something increasingly strange is happening around AI.

The people warning that development may be moving too fast are not outsiders who barely understand the technology. They include people building frontier AI systems themselves.

And yet the race is not stopping. If anything, it is accelerating.

We are discussing the need for brakes while everyone keeps a foot on the accelerator.

That contradiction deserves attention.

What is happening now

Reuters reported on September 16, 2026 that Meta CEO Mark Zuckerberg argued competition and liability give AI companies sufficient incentive to act individually on safety, rejecting the idea that the entire industry must coordinate a slowdown.

His position differs from recent calls by Anthropic CEO Dario Amodei and others, including OpenAI CEO Sam Altman and xAI leader Elon Musk, for slower capability growth and stronger safeguards.

Associated Press reported the same day that leading AI voices had found unusual common ground on safety, while translating that concern into coordinated action remains difficult because of commercial, geopolitical and regulatory pressures.

The point here is not to decide which executive is right.

The important fact is that people at the center of the industry are openly debating whether AI capabilities are advancing faster than safety mechanisms can keep up.

If it is dangerous, why not simply slow down?

In many industries, recognizing a serious potential hazard leads naturally to a pause for safety checks.

Cars, medicines and factories all work that way.

AI has a different structural problem.

If one company slows down, its competitors may continue. If one country becomes more cautious, another country may continue developing faster.

Slowing down for safety can therefore become a competitive disadvantage.

This is not simply a question of whether individual executives are good or bad people. It is an incentive problem.

Even if several participants believe that a slower pace would be safer, each participant has reason to keep moving if someone else might continue.

The result can be a race in which everyone keeps running.

Can safety depend on corporate goodwill alone?

AI companies are not ignoring safety. They conduct evaluations, restrict access, test models and increasingly use external reviewers.

The harder question is whether those mechanisms are sufficient.

Companies have strong incentives to make systems safe. A major failure can destroy trust, expose them to liability and drive customers away.

But those same companies also have strong incentives to release products quickly and remain competitive.

Inside the same organization, two forces therefore coexist: the pressure to proceed carefully and the pressure to move faster.

This tension is not unique to AI. Food, medicine and finance face similar conflicts, which is one reason societies created standards, audits, regulation and independent oversight rather than relying exclusively on corporate promises.

AI is increasingly confronting the same question.

AI is no longer just a convenient chatbot

Another reason the issue matters is the shift from systems that merely answer questions toward agents that can take actions.

Reuters reported on September 16 that Huawei expects autonomous AI agents to account for more than 90% of global AI token traffic by 2035. That is Huawei’s forecast, not an established future fact, but it illustrates the direction in which major technology companies expect the market to move.

AI systems are increasingly able to research, reason, call tools, write code and operate software.

As capability expands, so does usefulness.

But the potential impact of mistakes also expands.

I am pressing the accelerator too

It would be easy to criticize AI companies from a distance. But that would miss part of the problem.

I use AI extensively in my own work: research, writing, software development and business automation.

Tasks that once required hours can sometimes be completed in minutes.

Once an organization experiences that speed, returning to the old way of working becomes difficult.

And if competitors are using AI to increase productivity, how many companies can realistically say, “We are concerned about safety, so we will simply stop using it”?

The same incentive structure appears at the user level.

It is not only AI laboratories that find it difficult to lift their foot from the accelerator. The rest of us do too.

The question is not merely whether to stop AI

Framing the debate as either stopping AI completely or allowing unlimited development creates an unnecessarily extreme choice.

More practical questions are available.

What decisions should AI be allowed to make? Who retains final responsibility? Where must human approval remain mandatory? Can failures be traced after they occur? Can independent parties evaluate safety claims?

The performance race is unlikely to disappear easily.

That is precisely why brakes and guardrails may need to be designed before the accelerator becomes even more powerful.

If the current AI race consists of people warning about danger while collectively increasing speed, perhaps the most important risk is not only the technology itself.

It may also be the human structure that makes everyone feel they cannot afford to be the first to slow down.

Sources and references

• Reuters, “Meta’s Zuckerberg says AI labs have enough incentive to build safely”, September 16, 2026 https://www.reuters.com/business/metas-zuckerberg-says-ai-labs-have-enough-incentive-build-safely-2026-09-16/
• Associated Press, “AI rivals found rare agreement on safety. Putting it into practice is harder”, September 16, 2026 https://apnews.com/article/b61f28b6212338e88c0baec31f661701
• Reuters, “China’s Huawei forecasts billions of agents will dominate AI traffic by 2035”, September 16, 2026 https://www.reuters.com/legal/litigation/chinas-huawei-forecasts-billions-agents-will-dominate-ai-traffic-by-2035-2026-09-16/