Elon Musk is calling on the world’s leading artificial intelligence companies to subject their models to independent testing by competitors before releasing them to the public, proposing a form of industry peer review as concerns grow over the safety of sophisticated AI systems.

Speaking at the All-In Summit in Los Angeles on Monday, Musk said SpaceX’s xAI, OpenAI, Anthropic, Google, Meta and “three or four of the leading Chinese companies” should allow rival developers to run a common “test harness” against their models and identify potential safety problems before deployment.

“So, you know, instead of grading your own homework, you would at least have competitors grading your homework and raising the alarm if they see concerns,” Musk said.

The proposal comes as the AI industry faces a more intense debate over how quickly frontier models should be developed and whether voluntary safeguards are sufficient. Leaders of Anthropic, OpenAI and other AI companies have recently warned about the risks posed by powerful systems and called for a slower pace of development.

Musk and OpenAI CEO Sam Altman were among the technology executives who backed Anthropic CEO Dario Amodei’s proposal for slowing frontier AI development, creating an unusual degree of agreement among rivals that have otherwise been engaged in an aggressive race for users, computing capacity and market share.

Musk’s latest proposal shifts that discussion toward a specific mechanism: forcing AI developers to expose their systems to scrutiny from companies with competing commercial interests.

“The odds that you will find issues are dramatically greater,” Musk said, while acknowledging that the peer-review approach would not be a perfect solution.

The idea broadly resembles Amodei’s proposal for “embedded evaluators,” in which independent third parties would be given sufficient access to assess the safety of frontier models and verify that companies are following their commitments. Musk’s proposal goes a step further by explicitly bringing competing AI developers into the testing process.

AI Safety Debate Collides With Commercial Competition

The renewed safety debate was partly triggered by warnings from AI researchers about the possibility that future systems could pose catastrophic risks.

Jacob Coxon, a former researcher at Anthropic who had also worked at OpenAI, announced that he had left Anthropic and accused leading AI laboratories of “gambling with our lives.” His comments were followed by a warning from Evan Hubinger, an alignment lead at Anthropic, who said he agreed with Coxon and personally estimated that AI could kill all humans with a probability of more than 10% within the next decade.

Those warnings have added urgency to a debate that had largely centered on whether AI regulation should be imposed by governments or developed voluntarily by the companies building the technology.

The Trump administration has pushed back against calls for broader restrictions. President Donald Trump posted on Truth Social on Monday describing fears about AI as a “hoax” and a “scam.”

National Economic Council Director Kevin Hassett told CNBC on Tuesday that the private sector is the “right place” to address concerns surrounding AI. He said the government was monitoring the industry and would use “law enforcement when necessary to make sure that the firms are acting responsibly.”

The disagreement exposes a major tension in AI governance. The companies developing frontier models argue that they are best positioned to understand and manage rapidly evolving technical risks, while policymakers and researchers have raised questions about whether firms facing intense competitive pressure can be relied upon to police themselves.

Musk’s proposal is an attempt to address part of that problem without placing the testing mechanism entirely in government hands. If OpenAI, Anthropic, Google, Meta, xAI and major Chinese developers were required to test one another’s models, companies would have an incentive to search for vulnerabilities that their competitors might otherwise overlook.

But the commercial incentives are complicated.

Musk acknowledged that the other AI companies competing with his businesses have not agreed to his proposal. Each company has reasons to protect proprietary model information, while giving competitors access to sophisticated testing environments could reveal weaknesses, capabilities, or other information with commercial value.

The proposal also raises questions about who would control the testing framework, what constitutes a safety failure, and whether companies would be required to disclose problems discovered in a rival’s system.

Musk said the mechanism should be implemented quickly. “What I’m suggesting here is it’s a step in the right direction and it’s something that we do quickly,” he said. “I think it’s probably something that China would agree to.”

That last point is significant because the AI safety debate is unfolding alongside an increasingly explicit competition between the United States and China over advanced AI.

Trump has said that slowing the development of U.S. AI systems could allow China to gain an advantage. Amodei has also acknowledged the geopolitical dimension, describing the question of whether to slow development as the “toughest dilemma.”

China’s Foreign Ministry, meanwhile, dismissed the push by AI companies for a slowdown as “fear mongering,” according to a Reuters translation of remarks made Monday.

Musk’s Companies Face Their Own AI Scrutiny

Musk’s call for industry self-regulation also comes as his companies face legal and regulatory disputes over AI.

xAI has challenged AI-related legislation in U.S. states, including California’s AI Training Data Transparency Act, known as AB 2013, and a Minnesota law banning so-called nudify applications.

At the same time, SpaceX’s AI business is facing probes and lawsuits following the use of its Grok image-generation tools to produce and distribute non-consensual sexual imagery, including material depicting child sexual abuse.

Those controversies make Musk’s proposal particularly consequential. Peer review can increase the probability that dangerous behavior is identified before deployment, but its credibility depends on the willingness of companies to expose their own systems to scrutiny and act on findings that could delay a product or impose additional costs.

Musk’s AI empire has also expanded rapidly. SpaceX acquired his AI business xAI in February and completed a $60 billion acquisition of AI code-generation startup Cursor in August. The combined company is working to make SpaceXAI’s Grok models and tools more relevant to developers, putting them in direct competition with products from OpenAI, Anthropic and Google.

That competitive overlap is precisely what makes Musk’s proposed model both potentially useful and difficult to implement. A rival may be well positioned to discover a weakness in another company’s model, but it is also a direct commercial competitor that could benefit from exposing that weakness.

The broader debate is now moving beyond the question of whether AI companies should slow down. The more practical question is whether the industry can create a system in which developers are required to expose powerful models to credible scrutiny before those systems reach millions of users.

Musk’s “test harness” proposal offers one version of that model. Amodei’s embedded evaluators offer another. Both seek to solve the same problem: AI companies have strong incentives to move quickly, while the consequences of a serious failure can extend well beyond any single company.

Meta CEO Mark Zuckerberg has rejected calls to slow the development of sophisticated artificial intelligence models, arguing that companies that fail to make safety and alignment a core part of their products will ultimately lose ground to competitors.

Zuckerberg’s position places him closer to Nvidia CEO Jensen Huang than Anthropic CEO Dario Amodei, whose call for AI companies to deliberately slow the pace of capability development has triggered a broader debate about whether the industry can safely manage autonomous and powerful systems while racing to commercialize them.

“There is a lot of debate about slowing progress on capabilities until alignment catches up,” Zuckerberg said in a post on X and other social media platforms. “My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models.”

Alignment refers broadly to the work of making AI systems behave consistently with human intentions, values, and safety requirements. Zuckerberg’s argument is that alignment should become a competitive advantage rather than a reason to pause progress.

That belief is gaining interest as AI companies move beyond chatbots toward agents capable of taking actions, interacting with external systems and completing tasks with less direct human supervision. In that environment, an AI model’s usefulness depends not only on what it can accomplish, but also on whether users and businesses can trust it to operate within defined boundaries.

Zuckerberg said AI labs that fail to “focus on alignment will fall behind,” suggesting that safety could increasingly function as a product differentiator in much the same way that model performance, speed, and cost do today.

Liability Becomes Part of the Safety Equation

Zuckerberg’s case for continued development is also grounded in commercial incentives. He said that AI companies already have a powerful reason to prevent their models from causing harm because they face potentially significant liability if their products behave dangerously. That makes corporate responsibility, rather than additional government intervention, an important mechanism for managing AI risks.

Meta, he said, has already demonstrated that approach internally.

The company “delayed shipping” its Muse AI technologies because of safety and security concerns, Zuckerberg said, adding that Meta made the decision voluntarily rather than because regulators forced it to do so.

“We just did it as part of our day-to-day work because it was clearly the right thing for people and for us,” he said.

This represents a fundamentally different approach from Amodei’s argument that the entire industry should slow the rate at which AI capabilities improve until safety techniques can keep pace.

Amodei renewed that position at Salesforce’s annual Dreamforce conference, where he noted that slowing development can itself be a way for the industry to establish better standards and demonstrate responsible behavior.

“The way to lead the industry forward, to set an example, to say that everyone can always be better,” Amodei said.

The disagreement is therefore not simply about whether AI should be safe. All three executives acknowledge the importance of safety. The dividing line is how safety should be achieved while the technology continues advancing.

Amodei has emphasized the possibility that competitive pressure could cause companies to take risks they would otherwise avoid. Zuckerberg and Huang are putting greater weight on market incentives and corporate responsibility to produce safer systems.

Huang Rejects More AI-Specific Regulation

Huang made a similar argument at Dreamforce, telling Salesforce CEO Marc Benioff that AI developers should take responsibility for the products they build rather than rely on new government rules specifically designed to constrain AI risks.

Huang later reiterated that position in an interview on CNBC’s “Mad Money,” describing additional AI-specific laws and regulations as “just completely unnecessary.”

“We have plenty of laws. We have plenty of regulations that govern the reliability and the functionality of products,” Huang said.

His position reflects the interests of an AI infrastructure industry that is expanding rapidly and requires enormous investment in chips, data centers, and computing capacity. Additional regulation could raise compliance costs or slow the deployment of new systems, while existing product-safety and liability rules could potentially be applied to AI products without creating an entirely separate regulatory framework.

Amodei’s argument, however, is that frontier AI introduces risks that existing product rules may not adequately address, particularly as systems become capable of acting autonomously and potentially assisting in the development of more advanced AI.

Those risks are becoming harder for the industry to avoid.

OpenAI CEO Sam Altman, who later joined Benioff at Dreamforce, acknowledged that commercial and geopolitical competition creates a genuine safety problem. He said safety and monitoring should take precedence over features, while recognizing the fear that companies may fail to act cautiously because they are competing against one another.

“I think it’s great for our industry to say we want to come together and we want to be able to coordinate and make sure we have enough time to do this safely,” Altman said. “But when there’s any implication that because of the commercial pressures and the race, some company or between countries, some countries might not do the right thing, I think that’s when people get very scared.”

Altman’s comments occupy a middle position in the debate. He supports coordination and stronger monitoring but also recognizes that asking individual companies to slow down can be difficult when competitors continue advancing.

That tension is likely to become more significant as AI agents become a larger part of the industry’s commercial strategy. A model that merely generates text can be monitored differently from an agent that can access software, communicate externally, or execute tasks on behalf of a user.

For Meta, Nvidia and other companies betting on continued AI expansion, the commercial proposition is that safety can be engineered into the products without stopping the underlying capability race. However, the central concern for Anthropic is that capability improvements may outpace the industry’s ability to understand and control powerful AI systems.

The emerging debate is less about safety versus no safety than about where the burden of managing AI risk should fall: on companies through alignment, testing and liability; on governments through regulation; or on the industry collectively through coordination and deliberate limits on development.

Zuckerberg’s intervention adds one of the technology industry’s most powerful companies to the argument that safety itself can become part of the competitive race. If that view proves correct, the companies that build the most trusted AI systems may not need to choose between moving quickly and managing risk.