Reverse centaurs: EU approach to AI risks dystopia for human workers

The EU’s approach to AI changed substantially last June, but, to most, it went unnoticed. Instead of regulating the risks of AI, the European Commission now seems preoccupied with making AI happen. Whether you are a fan of AI or not, as a human, that should worry you. The Cloud and AI Development Act (CADA), published in June as a flagship of the EU’s Tech Sovereignty package, illustrates this new push. Were you to read Titles II and III of CADA, you might notice something odd: Almost none of the obligations are aimed at companies enabling AI. Instead, they’re aimed at the EU’s national governments. Member states must write national cloud and AI strategies, set up ‘Centres for AI’ and designate at least one “data centre acceleration zone”. They must remove data bottlenecks. The commission and member states must also make sure designated frontier AI projects get access to all the computational capacity they need. Gabriela Zanfir-Fortuna of the Future of Privacy Forum calls it arguably the EU’s first comprehensive tech regulation to impose “positive obligations” to enable innovation and the widespread adoption of AI. LinkedIn AI-hype beasts might love it, but the commission’s choice here is more important than it has been treated – or rather ignored. Most of the argument around CADA has been about sovereignty: hyperscalers, data centres, assurance levels, whether Europe can finally build some of its own infrastructure. But something else is happening below all this: widespread adoption of AI is no longer being treated as something that might happen, and which governments therefore need to regulate. It is becoming an objective of supranational government policy. Scifi author, journalist and digital activist Cory Doctorow’s new book The Reverse Centaur’s Guide to Life After AI provides an excellent framework for thinking about how the EU is manifesting its AI future. The premise of the book revolves around two simple concepts. The first is that a person using a machine can be considered a centaur: a human brain controlling a strong body. A person working in service of a machine is a reverse centaur – a horse’s brain using the human's dexterity or insight for what it cannot do. Doctorow illustrates this with a (painful) example regarding journalism. A media company doesn't necessarily use AI to make 10 journalists 20 percent better at their jobs. It instead discovers that three journalists with AI can be expected to produce roughly what 10 journalists produced before. The three who remain haven't been liberated from drudge work. The output targets of the missing seven have been thrust upon them, while also checking the machine's work and taking responsibility when it goes wrong. This brings us to the second premise: It’s less interesting to look at what AI can do, than who it does it for, and who it does it to. Who AI works for, and who it works upon, are two questions that don’t feature all that prominently in the current direction of European AI policy. The AI Act asks whether systems are safe, transparent and accountable, which are obviously important questions that are moving in the right direction. The new competitiveness agenda asks: how do we get European organisations to use more AI, faster? Why should that be the question? The tech industry has spent the past few years telling us that mass AI adoption is inevitable, and European policy is starting to absorb that assumption - despite increasing evidence that the transformative productivity gains are yet to be found. In his book, Doctorow describes this sort of argument as a ‘vulgar Thatcherism’: there is no alternative. In our European version: there is no alternative, so we had better build the data centre acceleration zones. So let’s take it at face value, and assume, for the sake of argument, that the commission is right: the models become very good, European firms become much more productive, Europe builds its own AI industry and reduces its dependence on American hyperscalers. Hooray! But it still leaves a basic problem. Blind spot: labour cost The one big reason why AI is attracting the terrifying amounts of capital it is is simple: its promise is to displace or replace a large part of the labour force. For a company, labour is a cost. But for an economy, labour is also income. For an individual firm, replacing workers with AI can make perfect sense. If it can produce the same amount with fewer employees, costs fall, productivity per worker rises and margins improve. The company has become more competitive! And done so strategically autonomously! Cue a new deadpan video on X by commission president Ursula von der Leyen. But now, imagine thousands of European companies do the same thing. If AI really does allow companies to employ substantially fewer people, or puts downward pressure on wages, households have less income. They spend less. Internal demand weakens. The single market as a whole gets poorer. A company can become more competitive by reducing its wage bill, but that does not mean that Europe becomes more competitive when every company reduces its wage bill. Europe, after all, is not a company. Worse still, because of the strange division of responsibilities inside the EU, Brussels has a formidable collection of tools for accelerating a technology: single-market legislation, competition policy, state-aid rules, infrastructure spending, standardisation, funding and now permitting coordination. CADA combines pretty much all these tools. But at the same time, the tools for dealing with large-scale labour displacement do not sit with Brussels. Unemployment insurance, welfare systems, most taxation, collective bargaining, wage-setting and much of labour-market policy remain member state responsibilities. In other words: the commission can tell 27 governments that widespread AI adoption is necessary for European competitiveness, while leaving those same 27 governments to work out individually what happens if one of AI's principal competitive advantages turns out to be needing fewer people. Perhaps I’m being pessimistic. Perhaps the technology does discover drugs, creates new products, opens new markets, makes previously impossible services possible and generates entirely new industries. I hope some of that happens. I truly do. But the trouble is that an individual company doesn't need any of that to happen for its AI investment to pay off. Nor, maybe more importantly, do their investors and shareholders. If a company spends €10m on AI systems and can subsequently remove €20m from its annual payroll, the business case is finished. No cancer cure required. Replacing labour is a considerably easier proposition to put in a spreadsheet than predicting the value of some future market. That means there is a non-zero chance that European industrial policy is being constructed around AI's most ambitious macroeconomic promises while the firms receiving the technology have a much simpler incentive: do the same work with fewer expensive people. Fun aside: European technological sovereignty doesn't fix this. Suppose a Dutch company buys its AI from a French provider and discovers that it can produce the same output with 800 employees instead of 1,000. The AI is European, the data centre is European, and the revenue stays in Europe. European productivity has increased. Strategic autonomy achieved! However, the Dutch government has 200 people looking for work. It may have to pay their unemployment benefits and retraining costs, while receiving less income tax. And those 200 households have less money to spend in the single market. It is obviously better, at least from a European industrial-policy perspective, if the AI company receiving the money is French rather than American. It doesn’t tell us anything about who ultimately benefits from what the machine does. Spoiler: it’s the shareholders. Sure yeah, that’s what taxes and redistribution are for, one might argue. But it so happens to be that those are some of the policies Brussels has no control over. It also overlooks the main premise of the commission’s U-turn from risk-controller to tech-promoter. Europe’s AI gamble The EU and its single market is not only a collection of companies who need to implement technology to make costs come down to compete globally. It’s a society, with 450 million people who buy things from those companies. Even more so, the fact that these people can buy things we make is one of our biggest strengths. If AI really does produce labour savings on the scale its investors expect, where those savings go is part of the concept of competitiveness, and not some social-policy question to worry about once the important work of becoming competitive is finished. AI could well make individual European companies more competitive. But if the mechanism by which it does so is substantially reducing the amount those companies spend on people, what happens when every European company does it? Doctorow's reverse centaur is a useful image here in more ways. In his version, the machine decides the pace and direction of travel while the human being attached to it does the work, absorbs the friction and takes responsibility when something goes wrong. There is something eerily similar in the way Europe is organising its AI policy. The commission is using the powers it does have to decide the direction we go in: more compute, more data centres, faster permitting, more investment, more adoption. If that produces large-scale labour displacement, many of the consequences will be borne further down the body: by national governments with their welfare budgets and labour markets, and ultimately by the people whose jobs and incomes are affected. Those governments can try to redistribute the gains afterwards. They can retrain people, change tax systems, strengthen collective bargaining or pay unemployment benefits. But they have (or pretend to have) remarkably little control over the European policy decision that helped create the problem in the first place. With CADA, the commission is shaping the EU like a reverse centaur. The commission is choosing where the body is going. The citizens supply the legs. And if the legs buckle, it will be their problem.