High Stakes for Middle Powers: Why I work on German AI Policy

In the pursuit of superlinear effects, caring about the frontier AI policy of middle AI powers like Germany is underrated: A lot of AI benefits are contingent on medium power policy, their successes and failures can quickly cascade, and medium power voices matter greatly on global governance.

Why I work on German AI Policy

I spend some of my time working on German frontier AI policy. That goes against the strong trend of AI policy folks seeking the largest jurisdiction in pursuit of superlinear effects of their work. So sometimes, I get asked why I spend time on Germany. I think some of my responses are broad enough to matter even beyond Germany’s borders. I argue:

Policy contingency of AI benefits is higher in middle powers. Policy work matters a lot.
Beyond its own borders, Germany will likely serve either as a blueprint or a warning tale
Germany is highly relevant to some important international dynamics. Without good policy work, it could mess them up.

AI policy work in middle powers can insulate against failure modes and to shape reactions to external wake-up moments. Policy organisations and the many large AI diasporai hailing from middle powers could take note.

Most People Don’t Live In The US

Frontier AI policy talk nowadays is largely a tale of two cities: There’s the US as a principal nexus, hosting all frontier labs, the leading manufacturer, most major policy organisations, and holding most of the cards when it comes to comprehensive regulation of model development. There’s also China, viewed as a main adversary, treated as a central motivator for many different maximalist demands from open-source crackdowns to all-out racing and prohibitive export controls. There’s also the UK’s impressive efforts to secure the third spot on the podium – but beyond that, prevailing opinion holds that no policy professional worth his salt should spend much time working on other national jurisdictions. But a lot of people live in other countries, too. Many of these are what I’d consider middle AI powers - economies with some AI presence and plenty of AI compatibility, but no current plcae at the frontier. These countries and their governments might not all have an outsized effect on how the development of frontier AI shakes out, and so it’s quite justified for a lot of people to work on US policy. But the policies of other nations still have a substantial effect on how this development affects their respective citizens – which do make up a majority of the world’s population.

In fact, I suspect that many practical outcomes of AI progress for countries that are not global leaders is far more contingent on successful policy, because they are not the default ground zero for any avenue of adoption. By successful policy, here and throughout this piece, I try my best not to import any value judgement: I mean neither particularly AI-safety-focused policy nor particularly hands-off dynamistic policy, but simply policy that sufficiently takes into account the recent and likely future rapid trajectory of AI capability growth.

For instance, by default, the US population will probably somehow partake in the everyday benefits of adopting US frontier labs’ models: Invariably, US models will be released and deployed in the US. For a country like Germany to benefit from US-driven AI progress, a lot more policy has to go right: They need to have continued access to them and must not lose it via geopolitical conflict, overregulation or affordability constraints due to economic downturn. The same argument applies e.g. to the security umbrella provided by non-civilian use – which is by default guaranteed to the US, but will only extend across the Atlantic under specific circumstances. Or to the availability of computing power to become sovereign or at least run your own inference on – which will presumably be provided in the US, but could well be strongly constrained elsewhere.

A recent, much-read piece published by the Hoover Institution makes all that abundantly clear, where even meaningful participation of Five Eyes countries is framed as highly uncertain. To be clear: It should not be taken for granted that US-aligned allies get unfettered access. This is not a matter of demanding access or railing against a bullying Trump administration — it’s a matter of making grown-up policy choices and leveraging strenghts to make the kind of participation we want mutually beneficial.

*This* *should send chills down the spine of every middle AI powers’ government.*

Many more examples come to mind, which are all to say: Life as a middle AI power is set to be acutely difficult and will require genuinely good policy work.

Germany is one of these middle powers, of course, and with 82 million people, the world’s third largest economy and fourth highest industrial output, there’s simply a substantial share of ‘the world’s activity’ here. Even by virtue of that, I’d think the way that frontier AI policy is done there should matter to some extent. (And of course, similarly, what is done in similar countries, too.) But my argument goes beyond that, too.

Quick Aside: Why not EU?

You might think that this is all fine and true, but could be addressed by EU-level action and policy work, where the EU AI Office has gathered a strong suite of technical and policy talent. I think the EU AI Office does great work on everything it has a mandate for, but I’m skeptical that this mandate will expand to next years’ frontier AI politics as they relate to Germany (and likely France and Italy). That is for broadly three reasons:

First, insofar as AI systems become increasingly relevant to ‘hardcore’ issues of national security, the EU will be consulted on them less due to lack of confidentiality & lack of loyalty to idiosyncratic national security interests.

Second, insofar as AI systems become increasingly central to economies and societies, they and their regulation are likely to become more politicized – which is to say there will be greater national political interest to influence them directly and less incentive to turf them to a more technocratic, but less political EU.

Third, between deeply unpopular legislation and purported chilling economic effects, some national governments are already concluding that the EU has been given too much leeway and hasn’t used it too well. This sentiment is further enhanced by a perceived overall ‘vibe shift’ supercharged by an incoming Trump administration and its threat to back US tech corporations against EU regulation. When given the choice, major national governments might choose to do their frontier AI policy inhouse instead.

Aspirational and Cautionary Tales

Beyond Germany’s borders, its frontier policy matters in two further ways that apply to Germany more than to some other middle powers. The first relates to aspirational and cautionary tales. Among the middle AI powers, Germany is one that many smaller countries will look at - the jury’s still out as to whether that will be in imitation or rejection. If done well, a strong and successful adaptation to AI progress could serve as a blueprint for the actions of other middle powers, effectively scaling up German policy wins at least throughout the group of similar-ish countries in which a lot of the world’s GDP and industrial capacity still lies. The UK already offers one such blueprint – on how to participate from a position of AI-specific strength, provided world-leading talent and a technology sector that largely outperforms the rest of an ailing economy.

Germany could offer another blueprint: On how the integration of AI can help leapfrog an atrophied digital environment, on how it can be integrated into a predominantly industrial economic model, and perhaps by a power that’s just a little bit less entirely aligned with the US. In many senses, Germany is well-positioned to do so. First as an economy: For pre-AI-rise levels, it had decent data center capacity; it has a comparatively highly industrial economy that might be a less immediately susceptible to near-future AI labour market impacts and a lot better at leveraging these systems’ benefits; and it still produces world-leading research contributions on many levels that would stand to benefit from AI-based acceleration. And second as a political environment: It has the theoretical overall state capacity to pull off decisive moves, the policy environment and associated political stability to be receptive to outside advice, and the political window to take decisive action this year. If we find good policy interventions – and make good choices where not to intervene –, the findings might be worthwhile elsewhere.

But Germany could also fail to adapt to AI developments. It could keep squabbling about its many other economic and political problems, largely inherited from the last decade or two, and only find the capacity to adapt to a new wave of changes once it’s already washed over its economy and society. It might push the country into surrendering its aims of sovereign competitiveness, push it further into blocks and geopolitical subservience; encourage political backlash and suggest pessimism and luddite techno-skepticism upon the prospect of having no path to maintain comparative standing in a world with ubiquitous advanced AI, etc. I don’t mean to overstate how much other countries care about Germany, but still: the global perception of a country with as strong a starting position faltering at that adaptation could also have ripple effects: Disabuse slightly weaker middle powers of their plans of sovereignty, invite them to blindly align with one geopolitical master or another, or encourage the same sort of political backlash. In short, it seems valuable to provide models for middle power successes, and it seems risky for a well-positioned middle power to falter and fail. Germany is one rare place where policy work might affect which of these stories will be told.

Internationally, Germany Is Somewhat Influential Still

Last, Germany still is a major voice in somewhat important fora: It’s one of the top NATO contributors, a de-facto co-leader of the EU, one of the G7 and one of the most well-positioned G20 members. In these roles, Germany can be quite impactful, from seriously endangering a lot of reasonable international policy to being a genuine driving force for good compromise.

Seriously endangering international progress is not all that unlikely: The fora listed above are some of the precious few avenues to ensure meaningful international cooperation on critical issues, from non-proliferation of dangerous capabilities to limits on military uses and more. Many of the very sophisticated international frameworks that have been suggested would be in danger if Germany wasn’t up to speed on AI. For a recent example: The EU AI Act almost got derailed by a lobbying push that hinged on motivating the German government to protect a nascent frontier model development sector that did not exist in any relevant sense. That plan almost succeeded due to lack of state capacity and expertise on frontier AI. To be clear, I think there were reasonable arguments to refuse the current version of the AI Act on economic grounds, as were provided e.g. by France – but that is emphatically not the conversation that happened in Germany. No matter what you think about the AI Act: If you care about international agreements, you probably don’t want a major international player to be far enough from the policy frontier to be this susceptible to shallow motivated campaigns from any interest group.

If things changed, Germany’s diplomatic position might very well be an asset to global AI policy. Very generally, Germany has historically performed a conciliatory function between power blocks before, e.g. in the late 70s and 80s between the USA and USSR in relation to nuclear weapon stationing and in the 2010s with regard to Russia. Today, for a western-aligned country, its relationship and economic integration with China is fairly strong, to an extent that could make it an important player and intermediary in some scenarios of high-stakes diplomacy around frontier AI development, deployment and military use. Even if it were my field of expertise, making more precise predictions around the geopolitical dynamics around upcoming AI diplomacy would be difficult given the uncertainties around the technical trajectory. For now, I simply want to say: Germany could matter here. If it does, it’s probably much better if it has well-informed frontier AI policy.

The Situation In Germany

This is not a post about German AI policy, per se. But to give some ideas on what actions all the arguments above might imply, I still provide a very short outlook of the kind of challenges it’s facing.

On policy, the only legislation concerning frontier AI so far is an early draft on EU AI Act implementation – i.e. a bill chiefly concerned with downstream enforcement of strategic considerations made elsewhere that largely only affects consumer market issues. Neither broader government strategy documents nor the plans outlined in the party platforms for the upcoming election next February suggest any change to that. It would be no exaggeration to say that there is no frontier AI policy originating from Germany. But that is partly because parties have only recently gotten interested in addressing the issue on a national level – and I’d expect at least a more prominent role for AI in the next administrations’ economic policy, and maybe some reactive pieces of frontier AI legislation depending on external shocks.

On capacity, the German civil service has long-standing issues with attracting young, dynamic talent that do not stop with AI. Germany’s broader ecosystem for political advice is similarly calcified, with a lot of incumbency-based structures and few pipelines for policy advice keeping the government more insulated from disruptive advice than elsewhere. But recent plans to relax and reshape structures around digital policy, as well as some breakthrough advisory organisations, have begun to upset this at least a little.

On clustering, there is a lot of latent category errors happening: Predominantly, AI is either understood as part of a big cluster of broadly relevant ‘future technologies’ one ought to be excited about; as part of a set of digital issues that also include social media platforms and state digitisation; or as part of a set of civil liberty threats that also include narrow AI systems from facial recognition to predictive algorithms.

Standalone awareness of the issue is exceedingly rare and becomes rarer still because these misconceptions shape the ecosystem. There are parliamentarians and organisations for each of these issue clusters, and each of them are eager to claim jurisdiction over AI based on their idiosyncratic categorization – and with limited general political interest, appreciation for the singular character of recent AI progress fails to materialize.

What’s Next?

Based on all of that, my model for the future of German frontier AI policy is not that a concerted push will lead to a meaningful shake-up in the short term. I am more confident that latent presence can prevent a lot of the genuinely bad effects I outlined above. Beyond that, I think the main impact lies in being ready for the external wake-up call. I am confident that critical moments are coming up quickly, whether they’re prompted by really good AI news or really bad AI news: at some point, awareness will reach the upper political echelons.

And at that point, I would like people to be ready: Ready with good ideas, ready with established channels of communication, ready with an established track record of caring about what happens in these middle powers. To my mind, that’s the best way to ensure that whenever they do happen, middle powers’ reaction to AI progress are shaped by a realistic view of what is happening, and aren’t susceptible to paralysis or overreaction.

In very practical terms, that opens some avenues for participation that I wish were better traveled and that I suspect are similarly open to other middle powers as well: The troves of highly qualified diaspora Germans can reach out back home and start engaging - I suspect similar groups exist for many other middle powers, too. The landscape of policy organisations could extend much more prominently to the national politics of middle powers. And the broader funding ecosystems could take engagement in these countries into account. The frontier AI policy of Germany – and its fellow middle powers – matters a little bit more than most people think.

I try to post updates on my writing, but not on much else, on X / Twitter.