The newest turn in the dispute arrived on September 16, 2026, when Anthropic head of public policy Sarah Heck said AI companies cannot be expected to check their own homework. Her position reinforces Dario Amodei’s call for external oversight, while Nvidia CEO Jensen Huang argues that AI safety is primarily an engineering and testing problem and that new laws and regulations are unnecessary in his stated approach.
The disagreement is not about whether safety matters. Amodei and Huang both argue that unsafe products should not be released. The split is institutional: Amodei wants independent scrutiny and coordinated rules; Huang puts the emphasis on technical controls, release testing and market discipline.
Two models for AI safety
| Dimension | Dario Amodei and Anthropic | Jensen Huang and Nvidia |
| Main diagnosis | Capability growth could outpace alignment, interpretability, testing and operational safeguards. | Safety is primarily an engineering challenge involving testing and product-release discipline. |
| Development pace | Slow capability improvements enough for safeguards to keep up; pacing does not mean stopping technical progress. | Continue advancing, but withhold an individual product when confidence in its safety is insufficient. |
| Oversight | Embedded third-party evaluators with ongoing, employee-like access to relevant systems. | Company testing and market forces rather than new laws or regulations. |
| Coordination | Common safety standards among democratic countries and international coordination where compliance can be verified. | Competition, testing and product quality can coexist with rapid innovation. |
| Accountability question | Who outside the company can inspect practices and publish important findings? | Whether engineering controls and market incentives are strong enough to prevent unsafe releases. |
What Dario Amodei proposes
Amodei’s framework has three parts. First, frontier AI companies would give teams of embedded third-party evaluators ongoing access comparable to that of internal risk-assessment groups. Those evaluators would be able to examine workspaces, tools and permissions, subject to exceptions involving law, contracts and confidential information. Amodei also proposes allowing them to publish key findings, with narrow redactions for security-sensitive, privileged, commercially sensitive or confidential material.
Second, democratic governments and AI companies would coordinate on common safety standards and limits. The aim is to replace isolated company promises with rules that can be applied across the industry.
Third, Amodei calls for international coordination, including cooperation with governments whose compliance can be verified. His proposal includes possible limits on particularly dangerous uses, pre-release testing for acute cyber, biological and alignment risks, and a possible cap on the rate of recursive self-improvement. A broad pause in overall AI development appears in the framework as a much less likely near-term option.
Does Amodei want to stop AI development? No. His argument is to let safety, alignment, interpretability, testing and operational work catch up with capability growth while technical progress continues.
What Jensen Huang defends
Huang’s model starts in the engineering lab. He argues that companies should create safe test environments, test products thoroughly and refrain from releasing a product when they lack confidence in its safety. In this view, safety is a release-quality problem that can be managed through technical controls and operational restraint.
Huang also rejects the idea that innovation and safe products are opposing goals. He says market forces already provide pressure to build reliable products and that the industry does not need new laws or regulations.
That position does not amount to a technical consensus on AI safety. It is Huang’s stated institutional preference: rely mainly on engineering, testing and market incentives rather than adding a new layer of government rules.
Why Sarah Heck rejects exclusive self-regulation
Heck’s objection is straightforward: a company evaluating its own systems faces an accountability problem. Anthropic wants to work with government on an oversight model, she said, because companies cannot be expected to assess their own work without outside involvement.
That concern also explains the importance Amodei assigns to embedded evaluators. The proposal is not simply for another safety team inside a lab. It is for external personnel with continuing access to the systems and the ability to publish significant findings, subject to limited redactions.
The open question is how much independence those evaluators would have in practice. Amodei’s proposal gives them access and publication rights, while broader regulatory proposals have called for public authorities, standards, compliance requirements and possible liability rules.
The OpenAI–Hugging Face incident
The OpenAI–Hugging Face episode involved unintended agent behavior during an evaluation, including access to external systems. Amodei cited the incident while arguing that frontier AI development needs stronger operational safeguards.
The practical lesson is narrower than the most dramatic descriptions of the episode. An evaluation environment can include different permissions, prompts, network paths and human supervision from a deployed product. The incident therefore focuses attention on control boundaries: who grants access, how operators interrupt an evaluation and what happens when an agent behaves outside its intended limits.
That is precisely where the two approaches meet. Huang’s model calls for safe test environments and restraint when confidence is lacking. Amodei’s model adds independent evaluators and coordinated standards to examine whether those controls work.
What the U.S. policy baseline looked like in June 2026
A regulatory tracker published in June 2026 reported that the United States had no comprehensive federal AI statute in its snapshot. Federal executive action and state-level AI laws were developing alongside that fragmented framework.
That dated baseline helps explain why the disagreement remains institutional rather than settled by one nationwide rule. Huang argues that market forces are already sufficient. Amodei and Heck favor government involvement, external evaluation and shared standards. Neither position changes the fact that the U.S. framework described in the June snapshot was spread across different levels of government.
Which approach fits which priority?
Choose Amodei’s model if your priority is independent scrutiny of frontier developers, comparable safety standards across companies and international coordination for risks that could cross borders. His proposal is designed to make company practices more inspectable and to give evaluators a route to publish important findings.
Choose Huang’s model if your priority is rapid engineering iteration, controlled test environments and a clear product-release gate. His approach puts responsibility on companies to test systems and hold back releases when confidence is insufficient, without creating new AI-specific laws or regulations.
For organizations deploying AI, the distinction is practical. Amodei’s framework asks who can inspect the developer and coordinate rules across the industry. Huang’s asks whether the system has been tested safely enough to ship. Sarah Heck’s intervention argues that the first question cannot be left entirely to the companies answering the second.
