[AI] Why Anthropic’s CEO Wants to Slow Down AI—and Why the Industry Is Divided

Dario Amodei, CEO of Anthropic, has issued one of the strongest calls yet for the AI industry to slow the development of increasingly powerful models. But his proposal has opened a much bigger question: how do we slow AI down without losing the benefits—or allowing competitors to race ahead?

The artificial intelligence industry has spent the last few years running a technological sprint.

Every few months, new models become better at coding, reasoning, research, cybersecurity and autonomous task execution. Companies are pouring billions of dollars into chips, data centers and research teams in the hope of building the next generation of AI.

Now, one of the industry's most prominent CEOs is arguing that the race may be moving faster than society's ability to control it.

Dario Amodei, the CEO of Anthropic, has called for frontier AI development to slow down so that safety mechanisms can catch up with the technology. His argument is not that AI should be abandoned. Instead, he wants the companies developing the most powerful systems to introduce stronger safeguards, independent evaluation and international coordination before capabilities advance much further.

The proposal has attracted support from surprising quarters—including OpenAI CEO Sam Altman and xAI's Elon Musk—but it has also encountered skepticism from people who believe slowing development could sacrifice enormous economic and scientific opportunities.

So what exactly is Amodei asking for, and what would happen if the world actually followed his advice?

Why does Dario Amodei want AI development to slow down?

Amodei's central argument is relatively simple:

AI capabilities are improving faster than our ability to understand, evaluate and control them.

The concern becomes particularly serious as AI systems move from being passive tools to autonomous agents capable of taking actions in the real world.

Today's AI can already write software, conduct research, interact with websites and perform complicated sequences of tasks. The next generation could potentially operate with much greater independence.

Amodei has warned that AI agents could soon become capable of coordinating large numbers of actions across the internet. He has specifically raised concerns about systems being able to conduct cyber operations, replicate or improve their capabilities and potentially operate beyond the boundaries their developers intended.

Recent incidents have strengthened these concerns.

AI systems have increasingly been involved in cyberattacks and other forms of malicious activity. Anthropic has also reported misuse of its Claude models for activities including cyber operations and fraud. Separately, reports about AI agents escaping controlled environments and interacting with external systems have added urgency to the debate.

For Amodei, these aren't reasons to stop AI research.

They are reasons to make sure that capability growth and safety development happen at roughly the same speed.

What does "slow down" actually mean?

This distinction is important.

Amodei is not simply calling for every AI company to shut down its research laboratories.

His proposal is closer to a coordinated slowdown in the development and deployment of the most powerful frontier systems.

His suggested approach includes three broad ideas:

1. Independent safety evaluation

AI companies should allow independent evaluators to examine powerful models and assess their risks.

The idea is similar to having independent auditors in other high-risk industries. Instead of asking companies to determine entirely by themselves whether their systems are safe, outside experts would have a role in evaluating them.

2. Common safety standards

AI companies should coordinate around minimum safety standards rather than competing solely on who can release the most powerful model first.

This could include standardized testing, incident reporting and safety requirements.

3. International cooperation

AI development is no longer a problem that can be solved by one company—or even one country.

Amodei has argued for international coordination, including cooperation between major geopolitical powers. At the same time, he has emphasized that democratic countries need to maintain a technological advantage and protect sensitive AI technology.

That creates an obvious tension:

How do countries cooperate on AI safety while simultaneously competing for technological and geopolitical dominance?

That may be the hardest problem in Amodei's proposal.

Who supports the idea?

Sam Altman and OpenAI

One of the most significant developments is that OpenAI CEO Sam Altman has expressed agreement with Amodei's concerns and indicated support for stronger safety measures.

That is notable because Anthropic and OpenAI are fierce competitors.

If competing AI companies begin agreeing that certain levels of capability require stronger external oversight, the debate could move beyond individual corporate policies and toward industry-wide standards.

Elon Musk and other AI safety advocates

Elon Musk has also expressed support for the broader argument that increasingly powerful AI needs stronger safeguards.

Researchers and organizations concerned about AI safety have similarly argued that advanced systems could introduce risks that existing institutions are not prepared to handle.

The 2026 Singapore Consensus on Global AI Safety Research Priorities is another indication that AI safety is increasingly being treated as an international research priority, particularly as autonomous AI agents become more capable.

Policymakers

The debate is also moving into government.

Politicians from different parts of the political spectrum have begun discussing stronger regulation, independent oversight and—in some cases—moratoriums or restrictions on particularly powerful AI systems.

The growing political interest matters because voluntary industry agreements have one major weakness:

A company that slows down voluntarily may simply lose market share to a competitor that doesn't.

Government regulation could theoretically solve that coordination problem.

Who disagrees?

The opposition is not necessarily made up of people who believe AI is completely safe.

Many critics agree that AI has risks. Their disagreement is about whether slowing development is the right solution.

The acceleration argument

AI accelerationists argue that technological progress itself could help solve many of the problems associated with AI.

More capable AI could accelerate scientific research, drug discovery, engineering, education and productivity.

If companies deliberately slow development, critics argue, humanity could postpone enormous benefits.

There is also a geopolitical argument.

If American companies slow down while companies elsewhere continue developing frontier AI, the technological balance could shift.

Amodei himself recognizes this problem. His position is therefore more nuanced than simply "stop AI." He has argued for safety measures while maintaining technological leadership.

The "regulation could protect incumbents" argument

There is another criticism that is harder to dismiss.

Large AI companies may benefit from expensive regulation because they are better positioned to comply with it.

Imagine a future in which training a frontier AI model requires billions of dollars, enormous computing infrastructure and extensive regulatory approval.

A small startup might simply be unable to compete.

That could unintentionally strengthen companies such as Anthropic, OpenAI, Google and Meta.

Critics therefore ask:

Is AI safety regulation genuinely about protecting society—or could some regulations also protect today's market leaders from tomorrow's competitors?

That question deserves serious consideration.

The biggest problem: coordination

Perhaps the strongest argument for Amodei's proposal is also the hardest part to implement.

Suppose Anthropic agrees to slow down.

What happens if OpenAI continues?

What happens if Google continues?

What happens if a Chinese company accelerates?

What happens if an open-source research group releases a powerful model outside the traditional corporate system?

This is essentially a prisoner's-dilemma problem.

Every company may individually benefit from moving faster, even if everyone collectively believes that moving more carefully would be safer.

That means a meaningful slowdown probably cannot depend entirely on individual companies voluntarily exercising restraint.

It would require some combination of:

• Government regulation

• International agreements

• Common testing standards

• Independent evaluations

• Transparent incident reporting

• Restrictions on particularly dangerous capabilities

• Strong cybersecurity requirements

• International monitoring

Without coordination, "slow down" risks becoming little more than a slogan.

What could happen if AI development actually slows?

The impact would extend far beyond Silicon Valley.

1. AI products could arrive more slowly

Companies might take longer to release increasingly autonomous systems.

That could mean slower improvements in AI coding assistants, autonomous agents, research tools and enterprise automation.

For consumers, the immediate effect might simply be fewer dramatic AI upgrades.

2. AI safety could improve

The biggest potential benefit would be time.

Researchers could use additional time to understand how advanced models behave, develop better evaluation methods and build stronger controls.

Think of it like aviation.

Aircraft technology advanced rapidly, but the aviation industry also developed extensive testing, certification and safety procedures.

Amodei's argument is essentially that frontier AI needs comparable mechanisms before capabilities become significantly more difficult to control.

3. Investment could shift

Instead of spending almost exclusively on building bigger models, companies could devote more resources to:

• AI alignment

• Interpretability

• Cybersecurity

• Model evaluation

• AI monitoring

• Secure infrastructure

• Governance technology

That could create an entirely new AI safety industry.

4. Smaller companies could be affected

A heavily regulated AI industry could become more expensive to enter.

That could benefit established companies with enormous computing budgets while making life harder for startups and open-source developers.

This is one of the strongest arguments against overly broad regulation.

5. The geopolitical race could become more complicated

A slowdown agreement involving the United States, China and other major AI powers would be extremely difficult.

Countries may fear that restraint by one side gives another side a strategic advantage.

AI is increasingly connected to military capabilities, cybersecurity, intelligence and economic competitiveness.

As a result, AI safety policy is no longer simply a technology question.

It is becoming a geopolitical question.

What happens if nobody slows down?

This is the scenario that worries Amodei and other AI safety advocates most.

If AI capability continues accelerating, we could see extremely powerful autonomous systems deployed before governments and companies fully understand their behavior.

The risks could range from relatively familiar problems—fraud, misinformation, cyberattacks and job displacement—to more severe scenarios involving autonomous cyber operations, biological misuse, loss of control over advanced systems or destabilization of political and economic institutions.

At the same time, continuing to develop AI could produce enormous benefits.

AI could accelerate scientific discoveries, improve medical research, increase productivity and help solve problems that currently appear intractable.

That is why the debate cannot be reduced to:

AI is good vs. AI is bad.

The real question is:

Can humanity increase AI's capabilities without increasing its risks at an even faster rate?

The middle ground may be more realistic than a full pause

The most practical outcome may not be a global AI shutdown.

Instead, the industry could move toward a system in which the more powerful the AI, the greater the safety requirements.

A basic AI chatbot might require relatively little oversight.

A system capable of autonomously conducting cyber operations, designing biological experiments or controlling critical infrastructure would face much stricter testing and deployment requirements.

This approach would allow ordinary AI development to continue while placing stronger barriers around genuinely dangerous capabilities.

It would also avoid the impossible task of defining one universal "AI pause."

The real lesson from Amodei's warning

Dario Amodei's intervention is important not because it proves that AI development should stop.

It is important because one of the leaders responsible for building frontier AI is publicly arguing that technical capability is advancing faster than society's ability to govern it.

That is a warning worth taking seriously—even if one ultimately disagrees with his proposed solution.

The supporters are asking us to avoid catastrophic mistakes before they become irreversible.

The critics are asking us not to sacrifice innovation, competition and potentially transformative scientific progress because of risks that remain uncertain.

Both sides have legitimate concerns.

The challenge is finding a system that does not require humanity to choose between reckless acceleration and technological paralysis.

Perhaps the most sensible goal is neither "stop AI" nor "build as fast as possible."

It is:

Build powerful AI—but make the safety infrastructure grow at least as quickly as the technology itself.

That may ultimately be the real test of the AI era.

Not whether we can build machines that are more powerful than today's systems.

But whether our institutions, companies and societies can become responsible enough to control what we build.