Steve Dempsey

IMAGINE WAKING UP on Monday morning to discover that the US Government had told Microsoft to disable Office 365 for everyone outside the United States. No Word. No Excel. No Outlook. No Teams meetings. No Azure-based sites or applications.

Sure, there’d be some relief. But there’d be way more outcry.

Government departments would stop functioning. SMEs would grind to a halt. Bigger companies would struggle too, maybe even more. A huge chunk of the workforce would suddenly find that their productivity is a matter of American foreign policy. It may sound like a farfetched scenario. But replace Microsoft’s services with frontier artificial intelligence and it is not hypothetical.

In June the US Commerce Department ordered Anthropic to cut off foreign nationals’ access to its two most advanced systems – Fable and Mythos. The reason? National security. The result? Anthropic’s newest systems were available only to approved American companies and government agencies. Foreign employees working for those organisations were also given access. Everyone else was barred.

And then, earlier this week these export controls were lifted. Fable and Mythos are now available to all.

But it’s also clear to all that American AI policy is a mess.

And this is a mess that impacts any organisation that wants to use the latest models from US companies.

But how can states and companies build tomorrow’s economies when the US is blocking and releasing frontier models inconsistently and with opaque rules that are invented and interpreted on the fly?

What shape does regulation take?

That said, regulation is clearly vital. Advanced AI models are increasingly viewed as strategic technologies, comparable to cryptography, semiconductors or nuclear technology – and concerns about model theft, cyber attacks and state-sponsored industrial espionage are real. But the current ‘jerk the steering wheel approach’ not only creates uncertainty, it also smacks of protectionism.

There’s also a real question about whether the type of regulation the US is imposing on AI companies is effective. The Chinese are great at copying other countries’ intellectual property; and AI isn’t any different. Anthropic recently claimed that Alibaba deployed a systematic distillation attack using 29m fake accounts. Distillation attacks are a way of copying an AI model through masses of interactions designed to map the capabilities of an AI model.

But even without this industrial-level skullduggery, technical geoblocks are porous. Over the past year researchers and tech enthusiasts have developed increasingly sophisticated workarounds to access models that the US has blocked.

While workarounds exist for Chinese tech companies, what about the rest of us? If AI is as critical as its many proponents say, shouldn’t we all have access?

A report from Ibec this week estimated that 64% of jobs will require significant reskilling because of the impact of artificial intelligence. It’s not too optimistic to think that every knowledge worker could rely on it in the same way modern accountants now rely on spreadsheets. If access to those systems becomes subject to export controls, geopolitical disputes or licensing decisions, productivity itself becomes vulnerable.

For Ireland, this is familiar territory. We have spent decades discussing energy security worrying about gas pipelines, and north-south interconnectors. Perhaps AI should now belong in the same category as energy. Other governments are asking similar questions. Austria has even floated the idea that the European Union hosts Anthropic infrastructure inside Europe as a way of protecting European users from American export restrictions.

And that’s the real question here. How does any Government or company outside the US consider investing in models that could be turned off on a whim?

Local hosting, as suggested by the Austrians, is one option. Reliance on more open source models is another. Diversifying investment across a range of model providers also sounds like a good plan.

Related Reads

Maybe we should be giving children more options than 'bad' or 'really bad' online lives?

Interoperability is key too. Critical public services should be able to move between models without disruption, and European cloud infrastructure should be capable of hosting multiple frontier systems, regardless of where they were built.

More funding for frontier labs outside the US or China is a winning notion. But it would be expensive. We’re talking hundreds of billions. Why? Because AI is less like the early internet, where smaller teams could build world-leading companies, and more like civil aviation or nuclear energy, where only a handful of organisations can afford to build the underlying systems.

Most importantly, given we don’t really know where we are on the AI hype cycle (are we on the peak of inflated expectations or the slope of enlightenment?) AI should be seen as an issue of resilience rather than technological prestige. Governments should procure AI the way sensible investors build portfolios: spread the risk, avoid single points of failure and always have another option.

It’s worth noting that AI doomers have been saying for years that AI will destroy humanity. Wouldn’t it be ironic if this were true, but not in the way they’re thinking?

What if AI will prove to be our downfall, not through some Terminator-style war between man and machines, rather by initiating societal collapse through inconsistent foreign policy and poor procurement practices?

Steve Dempsey is a media and technology expert and commentator. He is also director of advocacy and communications with the Irish Cancer Society.