Comments
My response complains that this argument rests on assumptions[2] which are left implicit in the narrative, when instead they should be made transparent. There is much that we do not understand well about how Europe can respond to accelerating AI development and power concentration, and we need open discussion about those uncertainties and competing responses.
In that spirit, here is the overarching question I would love to see discussed more[3], rather than less, amongst all those who engage with Europe 2031’s call for more serious EU AI policy endeavours:
How do we overcome the coordination problem of global AI governance? How do we prevent extreme power concentration within a few leading companies (and individual company CEOs)? What role can Europe play in responding to these challenges?
This is a brief essay about my concerns regarding Europe 2031 and the way it may be read and received. I was a bit reluctant to write this because I understand and share the authors’ frustration with European inaction, which is in part fueled by internal division and endless bickering. But I’m also reluctant to not write down what worries me: the publication of Europe 2031 may have net negative consequences if we don’t receive it critically and if we use it as an end rather than a start to discussions around the EU’s geopolitical AI strategy. I hope that this essay helps us think more critically and intentionally about the futures we envision and about how we work towards them. It is very tentative, not presenting final solutions but raising questions and prompting us to look more carefully before locking in onto any one strategy. I mean for this to be a critical but constructive contribution to the effort Europe 2031 has started, not an adversarial attack.
I spent roughly 3-4h drafting, discussing with friends, and formatting the essay below. It could be much better, more thoughtful, and more incisive in the arguments it is making (especially in its identification of assumptions in Europe 2031 and its outline of possible alternative perspectives). I am keen for criticism, refinements, and expansions of the ideas I start to develop below; as I emphasise above: the goal is engaged discussion, including discussion that takes apart what I'm saying!
Before launching into my critique, let me note what I appreciate the scenario for: it points at what I feel are valid and needed concerns about EU officials and institutions being indifferent to AI and its possible geopolitical implications, about the incredible lethargy of European bureaucracy, and about people with authority unjustly dismissing thoughtful input from lower-ranked colleagues. I share a sense of frustration with all of these, as well as a desire for things to improve. And I think Europe 2031 may be an effective prompt to bring some such improvements about, which I feel grateful for.
Why, then, do I worry about net-negative outcomes from the publication of the scenario? Because I think that it unreasonably presents one kind of AI risk and one bucket of solutions to mitigate that risk, making the implicit case that this is the one risk we need to be most concerned about and hence the one bucket of solutions we need to push for. That implicit case is - or should be - deeply controversial, based not on solid evidence but on a bunch of value judgement and empirical hunches. Making such value judgements and empirical guesstimates is fine; inevitable, even, if one is to venture into the space of policy recommendations at all. But presenting them hidden underneath a well-written and captivating story, rather than making the case openly and explicitly and with epistemic humility and transparency where warranted, is not inevitable.
The one risk that is presented in the scenario runs roughly as follows: frontier AI is presently being developed mainly by a few US companies, with Chinese competitors reasonably close behind whilst European-developed models remain far from the frontier. That makes European end users dependent on access to foreign AI models. As AI integration starts to dominate growth and productivity potentials around the world, dependence on foreign AI models makes Europe a vassal of whoever it can rent these models from. Such dependence basically disenfranchises Europe as a sovereign entity because the providers of frontier AI models can derive crushing leverage from Europe’s need to deploy frontier models. A disenfranchised Europe is a Europe without a viable future, and certainly one that is incapable of defending its values and European citizens’ interests in the decades to come.
The bucket of solutions, if I read the scenario correctly, mainly revolves around the building of compute capacity on European territory: data centres primarily, possibly also a greater portion of the supply chains needed to build the semiconductors that go into those data centres, with the goal of providing model access and model development capacities through European infrastructure.
Europe 2031 tells a story that is not without merit, nor one built on demonstrably false premises. But it is a bit arbitrary: other stories would be possible that would seem equally meritorious and would equally rest on not-demonstrably-false premises. Alternative stories could, for instance, depict a dystopia where a massive European data centre build-out sacrifices regional standards and distracts from global governance efforts to slow down AI development, while at the same time doing little to reign in the power of the United States (which continues to house leading frontier AI companies) over Europe. Or, to pick one other among many possibilities, such a story could portray a future where access to the newest AI models is only marginally important because progress in frontier models easily leaks out to help improve replica models (similar to what we seem to be seeing with Chinese model development at the moment), which means that industrial productivity, innovation, and research capabilities are nowhere near as badly affected by US frontier model export controls as Europe 2031 insinuates.
The choice of which of these stories to tell is arbitrary rather than random or coincidental because there is a specific worldview that underlies the narrative adopted by the authors of Europe 2031: one that privileges deregulation’s short-term gains for innovation over its longer-term risks to societal and environmental sustainability; one that accepts competition as an inevitability and holds that collective action problems, at least at a global scale, simply cannot be overcome, no matter the costs of failing to overcome them; and one that has accepted the risks of rapidly evolving AI development and thus absolves even self-avowedly responsible actors from the burden of fighting such recklessly fast development.
Building more data centres with fewer regulatory hurdles in the EU is not the one sensible response to our emerging AI-shaped geopolitics. It is a heavily assumption-laden response: It assumes that more data centres really will make the EU more autonomous and immune to US coercion, even though leading AI companies will continue to sit in the US (thus giving the US leverage even if it does not have a near-monopoly on compute). It assumes that more data centres in the EU will not speed up capability development, or that sped-up capability development is not a risk worth fighting (despite dire scenarios about the possible ramifications of AI models that improve faster than society can adapt and faster than safety researchers can build guardrails). It assumes that material assets are the only real instrument of power, the only truly secure option to protect the EU’s sovereignty and global influence. And it assumes that the conflict exists between countries (or country blocks), rather than between power-amassing monopoly companies and everybody else.
I can’t help but feel that following the recommendations laid out by the authors of Europe 2031 is giving away much of what we should be fighting for: a chance to mitigate extreme risks of AI; a chance of preventing extreme power concentration in the hands of a few recklessly accelerationist companies; and a chance to think critically about pathways to AI development where technical progress does not outpace the wisdom and resilience we need for a future where technology actually serves flourishing, rather than dooming it.
In other words, the questions I’d love to see contested and discussed, in cooperative but non-groupthink-y ways:
In the spirit of wanting to spark and advance discussion, I’m collecting a few resources that speak to the questions and arguments above: EU_AI_geopolitics_resources. If you know of or encounter others, please do drop them in the comments or submit them via this form: https://forms.gle/1tW1VAE9tA9QrxSKA. I’ll endeavour to keep this a living resource repository going forwards.
***
Lastly, as a PhD researcher studying the epistemics of foreign policy in a context of inconclusive evidence, I’m always keen to meet and convene policy analysts working on some of the most tricky knowledge problems we face in trying to tackle global challenges. If you are one of those, and might be interested in discussion exercises or joint projects with other researchers trying to make sense of Europe’s role in global AI governance, do reach out: [email protected].
For a summary of the scenario itself, focus on the first few sections. The last section of the linked summary outlines a range of policy recommendations, some of which go beyond what is discussed in the scenario itself (as far as I recall/understood). That last subsection fits neatly with my call for transparent argumentation and broad discussion. My argument in this essay is that this transparency and honest argumentation is lost in the scenario itself.
This post is a quick write-up, so no claim to exhaustiveness, but I identify some of the assumptions that strike me as particularly contentious and debate-worthy below.
A longer list of (sub-)questions in the concluding section below.
