Amazon says AI models should be tested before release—but stops short of a slowdown

Amazon has entered the AI-safety debate with a clear middle position: powerful models should be released only after rigorous testing and strong safeguards, and industry should work with government on protections. But Amazon does not support a general, coordinated slowdown in AI development. 6

Its core argument is that safety and progress are not mutually exclusive. That places Amazon closer to a “test before deployment” model than to proposals that would make the pace of capability advances conditional on safety progress.

What Amazon is advocating

Amazon told Reuters that models should be released when they are “ready and safe to use,” which it said requires rigorous testing and strong safeguards. It also stressed the need for collective action between industry and government. 6

That is a consequential distinction. Amazon is not arguing for a pause in training or for across-the-board limits on model improvements. Instead, its stated threshold is tied to release readiness: evaluate a system, apply safeguards, and deploy it only when it is safe enough for use.

The approach leaves open difficult questions that the debate has not resolved: who sets the testing standard, what evidence establishes that a model is safe, and whether testing run by developers themselves is sufficient.

How this differs from Anthropic’s slowdown proposal

Anthropic CEO Dario Amodei has urged frontier AI companies to slow the rate at which they improve model capabilities so safety measures have time to catch up. His proposal calls for independent evaluators with deep access, industry-wide standards, and international coordination. 12

Amodei’s position is therefore more demanding than Amazon’s. Rather than treating safety primarily as a release gate, he argues that the development pace itself may need to be reduced when safeguards lag capability advances. His warnings describe potential future misuse and loss-of-control scenarios; they are arguments for precaution, not proof that an autonomous AI takeover has already occurred. 111

Put simply:

• Amazon: continue progress, but do not ship models until they have passed rigorous testing and are safe to use. 6
• Anthropic: safety must also influence how quickly frontier capabilities are advanced, supported by independent oversight and coordinated standards. 12

Where other AI companies fit

OpenAI, Google DeepMind and xAI leaders publicly expressed support for the direction of Amodei’s proposal. OpenAI’s Sam Altman said he would match Anthropic’s commitment to outside evaluators with employee-level access to review safety practices. 4

But public support for the underlying safety goal does not automatically settle the practical details. A coordinated slowdown would need a shared definition of dangerous capabilities, credible verification, and agreement among competitors operating under commercial and geopolitical pressure. Reuters reported that major AI leaders remain split over both the urgency of the risk and the appropriate response.

Meta CEO Mark Zuckerberg, by contrast, has favored market-led safeguards rather than a coordinated slowdown.

The central policy dispute: voluntary testing or enforceable oversight?

Amazon’s position aligns naturally with tools such as pre-release evaluations, red-teaming, security controls, and risk reporting. These are already recognized elements of AI risk management: the 2026 International AI Safety Report identifies threat modeling, capability evaluations, transparency disclosures, and incident reporting as among the relevant practices.

The unresolved issue is whether these measures should remain voluntary. U.S. Senate negotiators have discussed legislation that would require AI firms to demonstrate “reasonable precautions” against harm. The reported proposal could allow the commerce secretary to demand evidence and use government auditors to test company products. 20

That type of duty-of-care regime would move the debate beyond corporate promises. It would not necessarily require a blanket pause, but it could create an external authority capable of challenging a company’s claim that a model is ready to deploy.

Cooperation has an antitrust and governance problem

A shared safety compact can make testing and reporting more consistent, but it also raises concerns over who gets to write the rules. FTC Chair Andrew Ferguson said people should be “deeply suspicious” when AI companies seek antitrust exemptions while also lobbying for new regulation, though experts told Reuters that existing antitrust law may allow coordination to avert catastrophic risks. 17

Cohere CEO Aidan Gomez has likewise warned that a safety arrangement controlled by a small set of dominant labs could entrench incumbents. He has argued for independently verifiable safeguards, transparency, and broader participation rather than allowing leading firms alone to define the standards.

Those critiques do not disprove the safety case for coordinated action. They highlight a separate design requirement: oversight needs to be credible and resistant to regulatory capture.

A difficult political backdrop

The U.S. policy environment may make Amazon’s industry-government formula difficult to implement. The Trump administration has urged a hands-off international approach to AI regulation, while President Trump has argued that existing guardrails are sufficient. The administration also told AI developers it would not subject open-weight models to voluntary safety testing.

Meanwhile, Bill Gates has argued that governments are not prepared for AI-driven changes and should cooperate internationally on risks including employment disruption, addiction, and cyberattacks. 18

The takeaway

Amazon’s contribution to the debate is not a call to slow AI down. It is a call for a tougher deployment threshold: no release until a model has been rigorously tested and protected by strong safeguards. 6

That is less restrictive than Anthropic’s proposal to pace frontier capability gains when safety cannot keep up. But it still pushes the debate toward a practical question with major consequences: whether companies can credibly police their own release decisions—or whether independent evaluators and public authorities must be able to verify, challenge, or block them. 120