Washington chose a handshake on AI safety. The UN says that isn’t enough

Two very different answers to the same question — who keeps frontier AI in check? — landed within a week of each other.

The White House: a voluntary accord

On September 29, President Trump announced that the leaders of Anthropic, Google, Meta, Nvidia, OpenAI and xAI had signed a voluntary accord, the Joint Commitment on Frontier Responsibilities, after a luncheon at the White House, NPR reported. Signatories included Anthropic’s Dario Amodei, Google’s Sundar Pichai, Meta’s Mark Zuckerberg, Nvidia’s Jensen Huang, OpenAI president Greg Brockman and xAI founder Elon Musk.

The accord pledges stronger safety controls, including internal and external reviews. It is, by design, self-policing: no new law, no regulator with powers to compel.

Other governments: make testing mandatory

That approach contrasts with a joint statement from 21 countries and the European Commission, published by the Dutch government. It asks for transparent safety protocols with mandatory pre-deployment testing and independent evaluation, common standards with shared reporting of serious incidents, and exploration of a new international institution able to set standards and verify them.

The UN: time is running out

On October 5, speaking at an AI and human rights summit in Oxford, UN High Commissioner for Human Rights Volker Türk warned that the world is running out of time to put guardrails on AI, UN News reported. He called for:

• a prohibition on AI systems making life-or-death decisions;
• mandatory human rights safeguards and due diligence;
• updated laws on liability, cybersecurity and data protection;
• countries that host AI companies to hold them accountable.

“We can no longer afford to wait or rely on goodwill.” — Volker Türk

What does it mean for builders?

Whether or not binding rules arrive soon, the direction is clear: evidence will be expected. Teams shipping AI features should start now — keep a record of the evals you run before each release, decide who reviews high-risk changes, and set up a simple way to log and report incidents. If regulation lands, you will already have the paperwork; if it doesn’t, you will have a better product.

For one example of a lab sharing a powerful model with vetted users first, see how Google is rolling out Gemini 4 Argon to cyber defenders.