If an AI company believes its next model needs more safety work, should it delay the release itself—or wait for rules that require its competitors to do the same?

That distinction sits at the center of David Sacks’ argument about frontier AI development. In a Bloomberg Tech interview, he said OpenAI and Anthropic should prioritize reliability and predictability if they believe that is necessary, without making those improvements contingent on government concessions or industry-wide agreements.

His position separates two decisions that often get bundled together: a company’s responsibility to make its own products safer, and the government’s responsibility to set rules for the market. The distinction is useful, but it does not resolve every question about AI oversight.

Slowing Down Does Not Have to Mean Stopping Research

Calls to slow frontier AI—the most capable general-purpose models—can describe several different actions. A lab might postpone a release, limit access to a risky capability, spend more time testing, or redirect researchers from increasing model power toward improving reliability.

Sacks framed the debate largely as a choice about those internal priorities, rather than a demand to abandon development. His argument was that companies already have commercial reasons to build systems that behave consistently: customers, especially enterprises, do not want software that takes unpredictable actions.

For buyers, that distinction matters. A model can improve on benchmarks without becoming suitable for a sensitive business workflow. Better evaluation, tighter permissions and more dependable behavior may be more valuable than another increase in raw capability.

Existing Laws Matter—but Liability Is Not Automatic

Sacks emphasized that AI does not operate outside the law. He pointed to rules covering fraud, privacy, cybersecurity and child safety, while arguing that potential product liability gives developers a strong incentive to avoid unsafe releases.

The broad point is important: using AI does not make otherwise unlawful conduct permissible. But the legal consequences of a particular failure depend on the jurisdiction, the claim, the parties involved and the facts. Whether a developer is liable for a model’s output or a downstream use is not settled simply by labeling the system a product.

There is also a difference between compensation after harm and prevention before deployment. Liability can encourage caution, but its effectiveness depends partly on whether injured parties can identify the failure, establish responsibility and pursue a remedy.

The Competition Problem Cuts Both Ways

Sacks described OpenAI and Anthropic as dominating frontier AI and argued that their position gives them room to emphasize safety. That was his assessment of the market, not an independently established finding; leadership can look different depending on the capability, product category or metric being measured.

His objection to coordinated restraint was more fundamental: companies should not seek protection from competition as the price of making their products safe. He rejected the idea that an antitrust waiver should be a prerequisite for internal safety improvements.

That concern deserves scrutiny. Rules with high fixed compliance costs can favor established companies over smaller challengers. Agreements among leading firms can also raise competition questions.

But the counterargument is not trivial. A company that delays a release may lose customers or investment to a less cautious rival. Common minimum standards can reduce that pressure. The policy challenge is to distinguish legitimate safety coordination from arrangements that unnecessarily restrict competition. Shared testing methods and an agreement to limit development are not the same thing.

Audits Need More Than an Independence Label

Sacks expressed support for transparency and audits, while questioning whether evaluators with financial or personnel ties to a lab could be genuinely independent.

A useful audit framework should answer practical questions:

• Who funds and selects the evaluator? Conflicts should be disclosed and managed.
• What can the evaluator inspect? Conclusions should reflect the access actually provided.
• What is being tested? General capability scores do not substitute for assessments of specific risks.
• What happens after a failure? Findings need a defined path to remediation and follow-up.

Independence also requires technical expertise. The goal is not simply to exclude anyone with industry experience, but to prevent relevant relationships from compromising the assessment.

Global Competition Does Not Settle the Safety Question

Sacks warned that broad restrictions could allow China to gain ground. That is an argument against poorly designed constraints, but it does not establish that every safety requirement would weaken competitiveness. More dependable systems may also support adoption and trust.

The most useful takeaway is narrower than either a blanket pause or blanket deregulation: labs can act on risks within their control now, while policymakers assess where voluntary action and existing law leave gaps. Corporate responsibility and public oversight are not mutually exclusive. A credible approach needs both clear obligations and evidence that those obligations are being met.

This article was inspired by David Sacks: Anthropic, OpenAI Should Slow AI If Deemed Necessary from Bloomberg Tech. Please visit the original video for the creator’s full presentation and context.