The battle over artificial-intelligence safety is splitting Silicon Valley in an increasingly visible way.

Anthropic CEO Dario Amodei has called for a slower pace of frontier AI development.

OpenAI CEO Sam Altman has backed greater coordination.

Nvidia CEO Jensen Huang has rejected the idea that the industry needs a coordinated slowdown.

And now Meta CEO Mark Zuckerberg is offering a different path: keep building, but let independent evaluators test whether the systems are safe.

Zuckerberg said AI laboratories should rely on independent evaluators and advisers rather than waiting for an industry-wide agreement to slow development. He described outside evaluation as an industry best practice and argued that companies already have strong incentives to make their systems reliable and aligned with users.

That places Meta closer to Nvidia's position than Anthropic's.

But it does not mean Zuckerberg is dismissing AI safety.

Quite the opposite.

Meta says it has already delayed the release of its Muse AI model for several months while working on safety and security issues.

The disagreement is therefore becoming less about whether AI needs safety controls and more about who should control them and how much they should interfere with development.

Zuckerberg's core argument is economic.

If an AI agent behaves unpredictably or ignores its user's intentions, customers will not trust it.

And if customers do not trust an AI product, they will stop using it.

That gives companies a direct financial incentive to improve what AI researchers call alignment — the ability of an AI system to behave in accordance with human instructions and intended goals.

Zuckerberg has argued that trust and alignment could become competitive advantages.

In other words, safety does not necessarily have to be imposed from outside.

It can become part of what companies compete on.

That is a fundamentally different philosophy from the proposal coming from Anthropic.

Anthropic CEO Dario Amodei recently published a lengthy argument calling for the world's leading AI laboratories to slow the development of frontier systems.

His concern is that AI capabilities may be improving faster than the industry's ability to understand and control them.

Amodei has pointed to recent incidents involving AI systems operating in unexpected ways and warned that future AI agents could become powerful enough to cause major damage if safeguards fail.

His proposal includes independent safety evaluation and greater international coordination.

That means Zuckerberg and Amodei actually agree on one important point:

External scrutiny is valuable.

Where they differ is whether the entire industry needs to slow down collectively.

Zuckerberg's answer is essentially no.

Individual laboratories can decide to slow down when necessary while continuing to compete.

Zuckerberg's strongest example is Muse.

Meta decided to delay the release of the AI tool while engineers worked on safety and security.

But the company did not demand that competitors pause their own projects.

Zuckerberg said Meta simply made the decision internally because it believed delaying the product was appropriate.

That distinction is central to his philosophy.

A company can slow down one model without asking the entire industry to stop.

The approach preserves competition while allowing safety decisions to be made at the product level.

Critics of coordinated slowdowns could argue that this avoids the strategic problem of giving competitors an advantage simply because one company exercised caution.

The most important part of Zuckerberg's proposal may therefore be independent evaluation.

AI companies could continue training and developing models while allowing outside organizations to examine those systems before release.

Those evaluators could test for dangerous capabilities.

They could examine cybersecurity risks.

They could look for deceptive behavior.

They could assess whether models follow instructions consistently.

And they could identify situations where a system behaves differently under pressure than developers expected.

The theory is similar to other high-risk industries.

Aircraft manufacturers build planes.

Independent safety organizations test them.

Financial companies operate markets.

Auditors and regulators monitor them.

Medical companies develop drugs.

Independent trials evaluate their safety.

AI is increasingly moving toward a similar model.

Meta's position becomes particularly interesting when viewed against the industry's current race.

There are a limited number of frontier AI laboratories.

But there could eventually be thousands of organizations building specialized agents.

That means relying entirely on internal safety teams may not scale.

A larger ecosystem of independent evaluators could potentially provide a continuous layer of external scrutiny.

Zuckerberg has specifically argued that the industry needs a broader and more diverse ecosystem of independent evaluators and advisers.

Such a system could also reduce the risk that companies become the sole judges of their own safety.

There is a difficult practical question.

Who exactly qualifies as independent?

If an evaluator receives most of its funding from AI companies, can it truly act independently?

If it depends on industry access for its research, could commercial pressure influence its findings?

If governments fund evaluators, could political priorities affect their assessments?

These are not minor technicalities.

The credibility of the entire model depends on independence.

A safety evaluation that is effectively controlled by the company being evaluated would be less meaningful.

That is why any future evaluation system would need clear standards around funding, access, publication and conflicts of interest.

Meta's chief AI officer Alexandr Wang has argued that AI laboratories need stronger governance frameworks and independent oversight as models become more capable.

He has also criticized a race toward recursive self-improvement as one of the potentially riskier paths for advanced AI.

That creates an interesting position for Meta.

The company is not arguing that safety can be ignored.

It is arguing that safety should be built into competitive development rather than used as a reason to create an industry-wide pause.

One of the most important concepts in the current debate is recursive self-improvement.

The concern is that future AI systems could potentially become capable of contributing significantly to the development of their successors.

If that happens, model improvement might accelerate dramatically.

Amodei views this possibility as one reason the industry needs to move cautiously.

Meta's response is that the industry should avoid directing enormous quantities of computing power toward a race for self-improving systems.

Zuckerberg has said Meta favors directing the majority of its compute toward serving users rather than aggressively pursuing recursive self-improvement.

That is an important strategic distinction.

Meta can continue investing heavily in AI while choosing which kinds of AI research receive the greatest resources.

AI companies have enormous incentives to keep moving.

Better models attract more users.

More users generate more revenue.

Stronger models can create competitive advantages.

And investors are rewarding companies that can demonstrate meaningful AI capabilities.

That means an industry-wide slowdown is extremely difficult to organize.

Suppose one company agrees to pause.

Its competitor could continue training.

The competitor might then release a better model and capture customers.

The company that paused would be punished economically.

This is the classic coordination problem at the heart of the debate.

Jensen Huang has pushed back against calls for coordinated AI restraint.

He has argued that market forces and responsible development can provide strong incentives for companies to build safer systems, while warning that excessive regulation could allow competitors to move ahead.

Meta's position is similar in one major respect.

Both companies believe developers should remain responsible for their own products rather than relying primarily on government-imposed limits on development speed.

But Meta is putting more emphasis on independent testing.

That distinction could make Zuckerberg's proposal a potential middle ground between acceleration and slowdown.

The political environment also favors continued AI development.

President Donald Trump has repeatedly rejected calls for broad AI slowdowns and emphasized America's need to remain ahead of China.

That creates another incentive for U.S. companies to keep advancing.

If regulatory restrictions are perceived as too aggressive, U.S. laboratories could potentially fall behind foreign competitors.

But if restrictions are too weak, safety concerns could intensify.

The challenge is therefore not simply technological.

It is geopolitical.

The debate could have major implications for AI-related companies.

If governments ultimately favor coordinated slowdown policies, companies supplying AI infrastructure could face changes in expected demand.

If policymakers favor independent evaluation, AI spending could continue while new businesses emerge around testing, cybersecurity, compliance and model monitoring.

That could create an entirely new industry.

The AI market would no longer consist only of model developers, chipmakers and cloud providers.

It could include a large ecosystem of independent AI auditors, evaluators, safety laboratories and certification companies.

Zuckerberg believes it can.

His argument is that users will naturally prefer AI systems they trust.

Companies that release unreliable or dangerously misaligned agents will lose customers.

That would turn safety into a market advantage.

But the opposite question remains:

What happens if a dangerous capability becomes commercially valuable?

History contains examples of industries where the most profitable behavior is not always the safest behavior.

That is why independent oversight may become important even if market incentives are strong.

That debate has largely moved on.

Almost everyone agrees that increasingly powerful AI systems need safeguards.

The disagreement is over the mechanism.

Anthropic wants the frontier paced.

OpenAI supports stronger coordination.

Nvidia argues against broad slowdowns.

Meta wants independent evaluators and believes companies can make safety decisions individually.

Google DeepMind is exploring industry-wide safety standards.

These approaches are not identical.

But they reveal something important.

The AI industry is beginning to build a new layer around its technology.

Not just models.

Not just chips.

Not just data centers.

Trust infrastructure.

Independent evaluation could become one of the most important businesses of the next stage of AI development.

And Zuckerberg's latest comments suggest Meta would rather invest in that infrastructure than agree to put the AI race itself on hold.

The technology keeps accelerating.

The debate over who gets to judge whether it is safe is accelerating with it.