The latest fight over frontier AI is no longer just about how quickly models should improve. Since Dario Amodei published his pacing proposal on September 12, 2026, the debate has shifted toward whether laboratories should coordinate safety work—and whether that coordination might require government mediation or an antitrust waiver. FTC Chair Andrew Ferguson, Jensen Huang and Mark Zuckerberg have warned that the approach could weaken competition or protect established companies.
Amodei’s proposal describes slower capability development so evaluation, monitoring and human control can keep up. It does not call for a blanket halt to AI research. The reported backlash concerns the legal and competitive structure around that safety effort, not a government decision to stop development.
Amodei’s proposal reaches the antitrust question
The framework associated with Amodei has three parts: embedded third-party evaluators, coordination among frontier laboratories and international coordination. The evaluators would have ongoing, employee-like access inside leading labs, giving them a continuing role in reviewing safety work and reporting incidents.
The second part is where competition policy enters the picture. The proposal calls for common safety standards and coordination among major laboratories, with reported language pointing to government mediation or waivers of antitrust restrictions. That is a proposed legal mechanism, not an enacted rule.
OpenAI has supported common testing, independent assessments, cybersecurity protections, incident reporting, national preparedness and tracking of recursive self-improvement. Elon Musk has suggested that competitors review one another’s models for one or two weeks before release. Demis Hassabis has supported an industry-funded standards body modeled on the Financial Industry Regulatory Authority.
The underlying argument is a collective-action problem: a laboratory that slows down alone could lose ground to rivals that continue pushing capability forward. Coordinated safety standards would aim to prevent that race from outrunning oversight. The trade-off is obvious—and thorny. The same coordination that makes shared safeguards easier could also affect how companies compete.
What supporters and critics want
The disagreement is mainly about who should set the pace and how much authority should sit outside each company.
| Policy question | Collective safety coordination | Company-by-company responsibility |
| Why slow development? | Safety evaluation and human control may be falling behind capability growth. | Each laboratory can adjust its own pace when it judges a model unsafe. |
| How should safety work happen? | Common testing, independent evaluation, incident reporting and shared standards. | Internal engineering, testing, liability incentives and existing legal oversight. |
| What role could government play? | Mediate cooperation or provide an antitrust framework for cross-lab discussions. | Apply existing law without creating a special industry exemption. |
| Main competition concern | Firms may race ahead when a rival believes a capability is unsafe. | Incumbents could use shared standards to raise entry barriers or reduce competitive pressure. |
Sam Altman has supported slower frontier development while allowing technical progress to continue. Anthropic’s approach gives independent evaluators a continuing role inside the development process. By contrast, Zuckerberg has said each laboratory has both the responsibility and the incentive to move at the pace required to train its models safely.
Huang has argued that market forces and existing law are sufficient, describing safety and testing as engineering problems. Ferguson, who chairs the Federal Trade Commission, warned that combining new regulations with an antitrust exemption could create barriers to entry and insulate incumbents from challengers. David Sacks has also opposed suspending antitrust law or creating a special regulatory-approval process.
Why critics distinguish safety sharing from competition limits
Jonathan Kanter, former head of the U.S. Department of Justice’s Antitrust Division, argues that AI companies do not need an antitrust exemption to deliver safe products. In his view, companies can share safety information and adopt common standards under existing competition law; agreements designed to reduce innovation or throttle competition would be a different matter.
That distinction is central. Sharing information about vulnerabilities, incident response or evaluation methods can address a common technical risk. An agreement about how quickly competing companies may improve their products would affect the market itself. The policy dispute is over where one activity ends and the other begins.
The proposed evaluators and shared standards would therefore have to be designed carefully. If large laboratories control the standards, smaller competitors could face higher compliance costs or rules shaped around incumbent resources. That is the competition concern raised by critics; it does not establish that the companies intend to suppress rivals.
The consumer lawsuit remains an allegation
On September 19, a consumer lawsuit alleged that Anthropic, OpenAI, SpaceXAI and Google coordinated to slow AI development and reduce the value of paid subscriptions. The allegation has not established illegal collusion or concrete consumer harm by itself.
Public statements supporting slower development are not automatically an unlawful agreement. Legal analysis presented by Renée DiResta said that proving both an illegal coordination arrangement and specific consumer harm would be difficult. The lawsuit’s allegation and the broader policy proposal are related in public debate, but they are not the same claim.
The distinction matters for readers and regulators alike. Amodei’s proposal is a public framework for pacing capability growth and coordinating safety work. The lawsuit makes a separate allegation about conduct by named companies. One does not, by itself, prove the other.
The unresolved policy choice
The argument now turns on whether safety coordination can be limited to technical safeguards without becoming an agreement that governs competitive behavior. Supporters see common evaluation and incident reporting as ways to prevent laboratories from racing beyond effective oversight. Critics want companies to use existing law, independent engineering and normal liability rules rather than receiving an antitrust carve-out.
Michael Kratsios has said that companies that believe their systems are unsafe can voluntarily stop or slow development. That position leaves each laboratory responsible for its own decision, while Amodei’s proposal seeks a framework intended to prevent unilateral restraint from becoming a competitive disadvantage.
The policy challenge is to make safety cooperation possible without giving the companies that dominate frontier AI control over who may compete and how quickly the field can advance.
