The United States government has begun delaying the public release of the most powerful artificial intelligence models, subjecting them to case-by-case security reviews that have no fixed criteria. In June, Anthropic’s Fable and Mythos models were temporarily blocked for foreign users over cybersecurity concerns, and OpenAI’s GPT-5.6 Sol launch was staggered at Washington’s request. The improvised process has drawn criticism for thin expertise and opaque decision-making, but it marks a decisive shift: frontier AI is now being treated as a dual-use technology requiring pre-release gatekeeping.

Viewed from Washington, the immediate trigger is the capacity of new models to autonomously discover software vulnerabilities. The same anxiety is reshaping policy in Beijing. According to Reuters, China’s Ministry of Commerce has met with Alibaba, ByteDance and Z.ai to discuss restricting foreign access to advanced models, including options ranging from outright publication bans to compulsory API licences. Officials are also examining whether the theft of proprietary AI technology should be reclassified as a national-security offence. The discussions follow Anthropic’s public accusations that several Chinese firms used fraudulent accounts to distil capabilities from its Claude model.

These regulatory moves are unfolding against a race for capital. Chinese memory-chip maker CXMT has set terms for a Shanghai IPO that could raise up to $9.8 billion, which would be the country’s largest mainland tech share sale. AI startup DeepSeek is preparing its own STAR Market filing, targeting a submission this year, while simultaneously raising fresh private capital at a valuation of roughly $74 billion. Its founder, Liang Wenfeng, now holds an estimated net worth of $36 billion. In the United States, OpenAI’s Sam Altman has signalled a price war, offering to undercut Anthropic’s Claude Fable 5 by a further 75 per cent, as both companies pursue divergent state-level lobbying strategies: Anthropic pushes for progressively stricter safety bills, while OpenAI promotes a uniform “reverse federalism” framework.

Amid the commercial and regulatory manoeuvring, a cluster of preliminary studies is raising questions about the cognitive effects of generative AI. One American-British study of 1,222 participants, still under peer review, found that using AI tools for arithmetic and reading comprehension improved short-term performance but diminished long-term results and persistence when the tools were removed. A 2025 MIT study reported that students using generative AI to write essays displayed weaker critical thinking. Researchers describe the phenomenon as “cognitive offloading” and caution that large-scale, long-term data are not yet available.

Google DeepMind CEO Demis Hassabis has proposed a public-private AI safety body modelled on the US Financial Industry Regulatory Authority, staffed by independent experts and funded largely by industry. The body would review frontier-class models up to 30 days before release, with the process initially voluntary before becoming a hard requirement for the US market. Microsoft CEO Satya Nadella endorsed the framework, writing that the goal is to promote innovation “while avoiding any one model drop that breaks the world.” Hassabis aims to have the body operational before the end of the year; its establishment, or the alternative of continued ad hoc export controls, is the next milestone to watch.