AI export curbs blur the line between security and mercantilism

Are restrictions on frontier models really driven by national-security concerns?

Not in our lifetime has any technology accelerated at the pace that artificial intelligence (AI) is doing so today. When OpenAI’s ChatGPT first broke into public awareness in early 2023 or so, it was clearly a vast improvement over previous text chatbots; yet, in the three years since then, it and its competitor models have metamorphosed into something unrecognisable. 

We still do not know what the effect on the macro-economy of these models’ effectiveness will be. What does appear clear, however, is that — as a public policy and economic-security problem — they do not represent a problem for the 2030s, but something far more immediate. 

Many were forced to confront this when the United States’ (US’) Department of Commerce reportedly told the AI lab Anthropic to cut off access for non-US citizens to its most advanced deployments, Fable 5 and Mythos 5. Naturally, this was a completely unimplementable order — if nothing else, Anthropic itself has a large number of non-US citizens working on those models. The order thus amounted to telling a major company to switch off its latest offer. 

After some negotiations, export controls were lifted on the deployment with more safety guardrails, namely Fable 5; Mythos 5, however, remains restricted although specific partners are allowed access. Fewer of its potentials are switched off for safety reasons. In return, those vetted operators who have been allowed to use it have to agree to far stricter terms that limit how they use it. 

Given that we know very little about the actual capabilities of these models — though rumours abound that Mythos proved capable of breaking high-level crypto-security used by the US intelligence establishment — it is also possible that this was a bit of strong-arming by the administration of President Donald Trump, which has taken a dislike to Anthropic after the latter attempted to impose ethical restrictions on the use of its technology by the Pentagon. 

But, if so, it has certainly served as excellent marketing material for Anthropic. Mythos has lent itself to a great deal of myth-making. 

But there is nevertheless a real question here. If Mythos or some successor deployment — perhaps from Anthropic, perhaps from another American lab — does indeed demonstrate the sort of extraordinary jump in capabilities and danger that we have only seen before a couple of times, such as with the development of nuclear weapons, what will Washington’s reaction be? And how will other countries respond? 

It seems clear that the US administration believes that it can both analyse swiftly enough the dangers that a model being released to the world might pose to its security, and that if it acts to restrict it then it will be able to do so effectively. The labs, they believe, will comply; the technology will not leak; it will not be replicated or copied. These are fairly stringent assumptions. 

The only real competitor to the US in terms of creating frontier models is the People’s Republic of China. (The European front-runner, France’s Mistral, has lost a great deal of ground in recent months.) Chinese models were initially being touted as being largely open-source, and easy to deploy and adapt. But that will not last forever. It was reported over the last fortnight that Beijing’s Ministry of Commerce had convened sessions with major technology companies to examine whether and how restrictions, especially export bans, could be imposed on their own top-tier models. According to one Reuters report, officials also want to limit who can invest in related startups within China. This followed the thwarting of Meta’s $2 billion bid for the Chinese startup Manus, and official orders to Moonshot AI that it obtain government approval before receiving foreign venture capital. Legal scholars gathered a few months ago to try and frame a new AI-focused economic-security law that might allow Beijing to limit frontier systems to domestic release, or to withhold them from the public altogether. 

The open question is whether this intensifying climate of restrictiveness in the two major AI geographies is driven by economic nationalism or by genuine national-security concerns. We know that most restrictions on trade in past years supposedly driven by security worries were actually mercantilism in another guise. But it might be too soon to assume that the same is true for these restrictions on AI. Certainly, the news that Beijing — which has so far used its cheaper models as a way of extending its economic power — might be reversing course suggests that genuine security concerns might exist.