Tech Tactics

The race to build ever more powerful AI systems is outpacing the world’s ability to govern them. Competition may even be actively undermining AI safety as cutting edge technology and accuracy is constantly in search of token optimisations and monetary savings. When Anthropic recently walked back on its industry-leading safety guarantee for new model releases, the company acknowledged that unilateral commitments made little sense while competitors were pressing ahead without similar constraints. OpenAI has faced comparable pressures, reportedly trimming pre-deployment safety testing windows to stay competitive. The pattern reveals an uncomfortable truth, asking companies to police themselves when their rivals are not, creates a structural race to the bottom.

That insight has given rise to an unconventional policy proposal gaining traction in legal and governance circles. Rather than relying on government regulators who typically lack the technical expertise and move too slowly or trusting companies to self-impose meaningful standards, the idea is to make AI firms regulate one another through a formal, government-certified peer oversight system.

The model draws on a long-established mechanism from financial services regulation. Under the proposal, private governing bodies certified by a government authority would set binding safety rules for member firms. Companies that opt-in would receive a form of liability protection, but that shield would be collective; if one member’s safety failure caused the entire governing body to lose its certification, every firm inside it would suffer the consequences.

The elegance of the idea lies in its incentive structure, if your competitor cuts corners on safety and costs you your operating licence, you have a very powerful reason to care about their practices. Such pledges amount to little more than reputation management; voluntary disclosures can encourage superficial compliance without substantive change, functioning essentially as marketing rather than accountability.

While Washington has moved steadily away from federal AI oversight as is evident in President Trump’s Executive Order 14179, issued in January 2025, which revoked the previous administration’s flagship AI safety directive and directed policymakers toward a pro-innovation, pro-competitiveness agenda. With a comparatively smaller federal role the burden of governance has fallen piecemeal onto states and, increasingly, on the other side of the Atlantic, on the EU.

Yet Europe’s much-heralded regulatory blueprint is proving harder to implement than its architects anticipated. The EU AI Act came into force in August 2024 and was intended to be fully applicable two years later, with a phased rollout assigning obligations to different categories of AI systems at staggered intervals. That orderly timetable is now in serious jeopardy.
As recently as late April 2026, the European Parliament, the Council of the EU, and the European Commission met for a second political trilogue on a proposed Digital Omnibus package intended to simplify the Act’s implementation and they were unable to reach an agreement. A further trilogue has been scheduled for later in May. Should negotiations collapse before the August 2026 deadline, the original, unamended obligations will snap into place automatically.
By pushing back high-risk AI rules while keeping the law non-retroactive, the bloc risks leaving some of the most sensitive AI applications permanently beyond oversight’s reach.

One of the more contested flashpoints has been the Code of Practice for general-purpose AI models, the guidance document that was supposed to give developers of large foundation models a clear framework for compliance. Reports indicate that disagreements over its scope and enforceability, particularly around transparency and systemic risk, have stalled its finalisation well beyond its original deadline.

The cumulative picture, a deregulatory US, a fragmented European framework struggling under its own weight, and a global absence of any coordinating mechanism is precisely what makes the peer-regulation proposal so timely. It does not require an international treaty. It asks only that governments certify oversight bodies with sufficient membership requirements to make defection costly, and then step back.

Although there are calls for government-mandated oversight conducted by professional AI auditors, arguing that it could deliver accountability without stifling innovation. The peer-regulation model goes further, embedding accountability within the competitive dynamics of the industry itself, turning the profit motive that currently undermines safety into a mechanism that enforces it.

Whether governments have the political will to structure such a system, and whether AI firms would accept meaningful peer scrutiny of their most valuable trade secrets remains an open question. But with the US’s state-level patchworks proliferating and Brussels  negotiations faltering, the pressure for a workable alternative has rarely been greater.