Governments on both sides of the Atlantic are converging on a new framework for artificial intelligence safety, one that would require labs building frontier models to submit to third-party testing before deployment. Officials briefed on the draft say the proposal, expected to be tabled at the 2026 AI Safety Summit, mirrors elements of the EU AI Act while incorporating red-line commitments first floated at Bletchley and Seoul. The goal, according to one senior negotiator, is to make "safety audits as routine as financial audits" for any system above a defined compute threshold.

Industry response has been cautiously cooperative. Several major model developers have privately signaled willingness to accept mandatory pre-release evaluations in exchange for regulatory clarity, though they continue to push back on provisions that would require the disclosure of proprietary training data. A coalition of smaller open-source labs, meanwhile, warned that an overly rigid regime could consolidate power among a handful of well-capitalized players, arguing that transparency obligations should scale with capability rather than company size.

Outside Europe and North America, the picture is more fragmented. China has published its own interim rules for generative AI and is in active talks with the G20 on mutual recognition of safety testing, while India and Brazil have indicated they will introduce domestic legislation in the first half of 2026. African Union officials, speaking in Addis Ababa last week, said they intend to table a common position by the summit, emphasizing that the global south must be at the table when thresholds and definitions are set. Analysts say the diversity of approaches is itself a risk: without interoperability, companies could route development to the lightest-touch jurisdiction.

Civil society groups are pressing for stronger enforcement teeth, including whistleblower protections for researchers and an independent international secretariat to monitor compliance. "Standards without inspection are just suggestions," said one longtime AI policy advocate. The next twelve months, both sides agree, will determine whether the world treats AI safety as a shared public good — or as a competitive race in which the most cautious player loses.