Worries that artificial intelligence could cause catastrophic harm are no longer confined to science fiction and late-night podcasts. Calls for new safeguards are multiplying, Bloomberg reports, but the people who would design those rules cannot agree on what they should look like.

The Administration's Playbook: Use What Already Exists

President Donald Trump's preferred approach came into focus on October 10, 2026. Rather than create a new federal regulatory regime built specifically for AI, the administration favors enforcing existing liability laws to manage the technology's risks. That model relies on established legal doctrines such as negligence and product liability, with courts drawing the lines case by case instead of a dedicated regulator writing rules in advance.

The stance has backing from prominent names. Treasury Secretary Scott Bessent supports it, as does David Sacks, the former White House AI czar.

That position builds on a White House meeting held September 29–30, 2026. There, Trump endorsed a non-binding accord among major AI companies, including Anthropic, OpenAI, Nvidia, Meta and Google. The accord centers on two commitments: robust internal processes at the participating companies and independent audits.

The key word is non-binding. Participating firms are agreeing to hold themselves to a standard, not submitting to one enforced by law — compliance rests on the companies themselves, and a participant that steps away from the commitments faces no statutory penalty for doing so.

Experts Can't Agree, and One Dissent Comes From Inside Big Tech

On October 2, 2026, Google's James Manyika said AI needs regulation that goes beyond self-policing, warning that companies keeping themselves in check is insufficient given how critical the technology has become. The critique carries added weight because it comes from Google, one of the companies covered by the September accord — a counterpoint to the voluntary approach voiced from inside the group it relies on.

A RAND report from March 2026 helps explain why the broader debate has stalled. The study gathered the views of 24 AI and policy experts and found little optimism that strict regulation is achievable. Among the options assessed, the experts rated incentives for risk disclosure as the most promising path, while strict regulations landed at the bottom as the least feasible. Voluntary measures and disclosure incentives were seen as the more realistic tools.

The States Aren't Waiting

On September 18, 2026, California Governor Gavin Newsom issued an executive order launching a review of AI safety plans. The review also covers so-called "kill switch" mechanisms — ways to shut down an AI system if it starts behaving dangerously. The order reflects rising concern about risks driven by AI. It also marks a contrast with the federal posture: unlike a non-binding accord, safety requirements that emerge from a state review could carry the force of law.

What This Means for the AI Industry and Investors

For companies building frontier AI, the administration's approach offers breathing room. Leaning on existing liability law rather than new rules means firms may face less immediate pressure to overhaul how they operate.

Relying on liability law, however, means the boundaries get defined after harm occurs, often through lawsuits, rather than being spelled out in advance.

If California and other states pursue their own safety requirements, firms could face a patchwork of rules that varies by jurisdiction, which is typically more expensive to navigate than a single national standard.

Several things are worth watching. One is whether the independent audits promised under the accord actually take place, and what they reveal. Another is what California's review produces, particularly on kill switch requirements. A third is whether dissent like Manyika's spreads to other executives, which would make the argument that self-regulation is sufficient harder to sustain.