The European Union often makes bad laws. This isn’t just my opinion – holders of this view include ardent Eurosceptics such as Ursula von der Leyen and Mario Draghi.
This is relevant for Britain, because we have a Prime Minister committed to bringing the country closer into Europe’s legal orbit. Whether that is a good or a bad idea is therefore a deeply practical question – what would the law do, what would the UK get, and what would we give up in return?
Sadly, a close look at the proposed Cloud and AI Development Act (CADA), which Burnham wants us to be aligned with, shows it is EU policymaking at its very worst. Lots of noble, lofty words about achieving European AI sovereignty. But little recognition of the fundamental reality that Europe is not, and has no prospect anytime soon of being, at the frontier. And then some very bad ideas of the kind that are compounding our economic problems in Europe.
What is in CADA?
There’s a lot in CADA that I could get into – Net Zero for data centres, prioritising the kind of open source technology that primarily comes from China. But though these are important, I think the most important is that CADA would place strict limitations on the kinds of AI that governments can actually use.
In true eurocrat style, this involves the creation of a new layer of bureaucracy, with each member state charged with creating its own “national competent authority” that can accredit AI products for use (Article 17). But the real problem is what it takes to be accredited in the first place.
All AI products used by EU governments would have to meet “assurance level 1” (Article 30), which requires, among other things, that:
- The company itself is “established in the Union”,
- “Where the [AI company] is subject to the control of a third country… there are no laws and practices in that third country… that require the [AI firm] to report information on software vulnerabilities to authorities… prior to those vulnerabilities being known to have been exploited”. (Clause 1, Annex II)
The conditions get progressively stricter as the use cases get more sensitive, such as /security, defence or justice. For instance, Level 2 requires:
- The firm’s infrastructure, assets and personnel are all inside the EU
- “if the [AI company]… is subject to the control of a third country… they demonstrate… that the possibility of disruption of the service… by a third country… is prevented”. (Clause 2, Annex II)
Assurance levels 3 and 4 exclude ownership or influence by third countries altogether.
What might this mean in the real world?
This all sounds reasonable, doesn’t it? But let’s think about what this means in the world as it is. All the best frontier models are currently developed and operated by American companies, staffed by largely American employees and built on American infrastructure. In Euro-speak, the US is very much a “third country”.
In the first instance, a plain reading of the criteria for levels 3 and 4 would simply exclude American firms – i.e. the world’s best AI companies - from working on the most important government problems altogether.
But even levels 1 and 2 are problematic. Demanding that “there are no laws and practices… that require the [AI firm] to report information on software vulnerabilities to authorities” seemingly runs against the regulation that the AI industry is itself asking for.
We have to be clear-eyed: the most important AI regulation is likely to be via domestic US law, governing these firms’ activities in the US. And the frontier labs themselves are pushing for mandatory reporting, which will probably be to the US government:
- OpenAI has called for “a mandatory evaluation process” run by the Centre for AI Standards and Innovation as the “US federal government’s primary institution for frontier AI safety”.
- Anthropic CEO Dario Amodei has said that “frontier labs should partner with government to formalize the idea of permanent embedded evaluators”
- Demis Hassabis has proposed a “new Standards Body”, which would be “federally overseen” with a view that frontier models would be “required to pass [testing] to be deployed”.
Any company subject to regulation in this way would seemingly be ineligible for even the lowest assurance level in CADA.
The one thing Britain has done to play itself into these important debates was Rishi Sunak’s decision to create AISI, which is the only serious non-US player in these debates. UK government’s have fought to keep AISI access to frontier models. It would be quite easy for measures which might strengthen AISI oversight to fall foul of this new EU law.
Excluding the possibility of disruption is also impossible for US firms. Earlier this year, Anthropic’s latest models were temporarily withdrawn after the US threatened to impose export controls. There is no way for any US company to guarantee that this will not happen again.
This is a bad idea
My view is that CADA is likely to be extremely bad for European public services, because European governments will effectively be cutting off their access to the world’s best AI services. The combination of the law and political pressure will push us away from using the best, but will not create an EU sovereign capability any more than when Europe tried to create an “EU Google”, known as “Quaero”. Reader, it did not go well.
Their own impact assessment recognises that its plan will come at a price, noting “any additional costs linked to providing sovereign services – compliance costs linked to audits, certification costs, the use of EU-based infrastructure, higher labour costs for EU-based staff – would be passed on to customers at higher prices”.[1]
But that assumes that Europe gets the same AI technology, albeit at a higher price. Recent history suggests that it simply will not happen.
The question for Britain is what we need to agree to in order to be accredited as an ‘associated’ third country under CADA (Article 18), which would allow UK tech to be accredited up to assurance level 3. Would it involve our own government signing up to the same standards, so that Europe’s anti-competitive wall now extends around us?
These are the practical problems that need to be addressed about what greater alignment actually means. Imagine a future Labour government tries to bring us back into the EU. If the government is already using US-based AI services, would re-entering mean adopting CADA and ending those deals? What would that mean for public services? Would that actually be in the best interest of the British public?
Sadly, I don’t think this Labour government has answers.
[1] Unbelievably, there is also a lengthy section espousing the benefits of protectionism, stating that there will be “domestic value added through import substitution”, arguing that it would “stimulate additional investment in cloud and AI infrastructure within the EU”. That’s right – the EU’s view is that banning what are quite literally the most technologically advanced and high value-add tools in the history of the human species will actually be good for the European economy.
Incidentally, this is basically the Chinese government’s approach to industrial policy: ban imports and then force your taxpayers to subsidise uncompetitive domestic firms until they have caught up with the rest of the world. The problem with the comparison is that the Chinese have far tougher bans on foreign products than the EU is proposing, and critically that the volume of subsidy is absolutely enormous in China, to the point of drastically suppressing household incomes. I somehow doubt the EU would be willing to do that if that’s takes to catch up with the frontier (which China can’t currently do anyway).