Labs Launch Safety Body as Global AI Governance Fractures

The week's AI news converges on a single fault line: who, if anyone, will govern the world's most powerful technology — and the gap between those demanding action and those resisting it is widening fast.

Major AI Labs Form Independent Self-Regulatory Standards Body

OpenAI, Anthropic, and Google DeepMind are collaborating to establish a jointly funded, independently run body that would evaluate frontier AI models before their public release — modelled loosely on FINRA, the US financial industry self-regulator. Under the proposed structure, powerful new models would undergo standardised safety assessments by outside evaluators before deployment, adding a layer of accountability between a company's internal testing and public access. The initiative follows a series of public safety warnings from researchers at these same organisations, and signals that the labs themselves recognise external scrutiny is now expected. Whether the body will have genuine independence from the companies financing it remains the central question; self-regulation proposals in other industries have often struggled on exactly that point.

Trump Calls AI Safety Concerns a 'Hoax'

At a public event on September 16, Donald Trump dismissed existential AI risk as a "hoax" and argued that the only AI oversight the US needs is a "strong president." His remarks put him in direct tension with researchers at OpenAI and Anthropic who have publicly warned that increasingly capable models could cause serious harm without adequate safeguards. Trump also restated his opposition to federal AI legislation, preferring industry self-regulation — a position endorsed by some major technology investors but at odds with most published AI risk assessments and with the safety standards the labs themselves are now endorsing. The statements deepen an already significant US political divide over how, and whether, to govern frontier AI.

UN Chief Warns of a 'Race to the Bottom' on AI Safety

UN Secretary-General António Guterres on September 16 warned that competitive pressure among nations is driving governments to relax AI safety requirements rather than strengthen them, describing the dynamic as a "race to the bottom." Speaking in New York, Guterres called for urgent international agreement on shared principles to protect civilians from AI-enabled harm, pointing to autonomous weapons and AI-assisted surveillance as areas requiring immediate multilateral rules. He noted that no binding global framework for AI currently exists, and that the gap between AI capability and governance infrastructure is growing faster than member states are moving to close it. The Secretary-General has placed AI governance on the UN agenda since 2023; his direct language this week reflects mounting urgency as major powers deepen their investments.

AI Rivalry Expected to Dominate Xi-Trump Washington Summit

When Chinese President Xi Jinping and Donald Trump meet in Washington in the coming weeks, artificial intelligence is expected to be among the defining agenda items. Chinese firms including DeepSeek and Moonshot AI have closed a significant portion of the performance gap with leading US models over the past year, complicating earlier US assumptions about maintaining a clear technological lead. Talks are expected to cover access to advanced semiconductors — where US export controls have been repeatedly tested and circumvented — as well as allegations of intellectual property copying and the possibility of bilateral agreements on AI oversight. The outcome could reshape both the international AI governance landscape and the global rules around the technology's most powerful hardware.

US Congress Adds Kill Switches to AI Bill; Floor Vote Slips to 2027

US lawmakers are debating the addition of mandatory pre-deployment safety audits, emergency shutdown mechanisms for high-risk AI systems, and a permanent prohibition on developing artificial superintelligence to the Frontier AI Safety Act — the most comprehensive federal AI bill currently under consideration. The new provisions reflect growing concern in Congress about both near-term misuse and longer-term capability risks. However, the House committee chair responsible for the bill has indicated it is unlikely to reach a full floor vote before 2027, citing opposition from industry groups and disagreements among committee members about scope. The delay leaves the US without a federal AI law as individual states — California foremost among them — continue passing their own rules, creating a growing patchwork of requirements in the absence of national standards.