To be or not to be, is that still the question?

The topic of AI sovereignty has been in the news a lot lately. Back in June Trump demanded Anthropic and OpenAI put restrictions on distributions of models to the EU. Although not the origin of the European “digital sovereignty” discussions it was definitively an acceleration point. As is the nature of bureaucracy, the powers that be had already spent many meetings doing a lot of talking, and talking, and talking, with little forward traction. Trump’s tantrum forced the conversation to move forward, somehow. Coincidentally, at the same time that the Trump administration was putting together their directive to prevent the use of Fable 5 and Mythos 5 by foreign nationals the EU was unveiling their latest tech sovereignty package, including the announcement of Mistral’s Compute strategy. This may seem to many like a nothing burger, this emerging picture has made for some interesting conversations. On one hand the argument that moving away from the EU’s reliance on foreign tech is a step in the right direction, on the other hand EU tech is behind and catching up will take a collective European mind shift.

Working in the tech industry in Europe has been an interesting experience, my background is in the US, here things move a LOT slower. The idea of move quick, make mistakes, innovate, ask questions but not for permission, is not how the process works here. This has brought some comfort with data security that I didn’t realize how comforting it is to have but this has come at the expense of innovation – leaving the EU in the dust.

Mistral is a good example of this: a young company with a competitive uphill climb. They are the main LLM provider in the EU, providing mediocre open source frontier models that have little chance of competing against the big tech dogs (evaluations don’t even put them in the top 10). I have to admit, I’ve never used their models, electing to go for better performing models from Meta or Moonshot. Mistral has a long way to go to catch an edge, so their pivot to compute makes sense. But this leaves less than to be desired in terms of “sovereignty”. Their own announcement hints at a hard sovereignty shell with a soft reliance center; with multi-year contracts partnerships with Z.ai and Nvidia which could end up as a lot of process bureaucracy wrapped in some fluffy messaging. It’s safe to say I am not confident that EU can loosen their own self-deprecating policies enough to allow Mistral to catch up and if it can’t the sovereignty they champion is just a lot of smoke and mirrors.

Europe has long played it safe, watching the weakening of a capitalist country will do that to you. It’s no secret that America’s obsession with consumerism has widened the pay gap, destabilized the economy and rendered the administration unreliable. Steps needed to be taken to protect the privacy, security and property of the EU and its citizens. What was not considered was the implications this would have on the advancement of technology within the EU. I’m also not convinced it was ever a consideration. My experience working in Germany, for example, is that it has been very successful with establishing processes that have helped define it as a manufacturing powerhouse. From Birkenstock to BMW the country’s reputation for quality is well deserved. But, with regards to process, one size do not fit all, and to remain relevant in a rapidly changing world, one must adapt. Innovation is not static, by definition. Unfortunately for Germany, and the wider EU, they have stopped adapting believing instead that the world must adapt to them. News flash, the world is moving on. Unknown risks cannot become a sufficient argument against experimentation.

When the automobile was invented, it was not introduced to the public fully equipped with seatbelts, crash testing, road rules or licensing systems. Those things developed along the way. They emerged because people used cars, observed recurring failures and developed responses. People couldn’t have designed the complete automotive safety system before living and experiencing the automobile.

Another example comes from my introduction to disruptive technologies. I worked with a company that designed IoT systems. They owned the devices, designed the protocols, built the platforms. The company spent a lot of money, hired the right people. They, largely, failed. The experienced offered an especially relevant lesson. Early deployment revealed serious problems: insecure defaults, weak authentication, abandoned devices, botnets and unclear responsibility for updates. Most of all, an inflexibility and conviction that didn’t allow for reflection, redirection or the freedom to fail. The correct conclusion was not that connected devices should never have been deployed (think the connected toaster, refrigerator with automated grocery lists). It was that experimentation needed to be accompanied by mechanisms for correction.

That were the more useful principles for AI are born. Do not demand certainty before deployment. Instead, demand observability, reversibility and accountability during deployment. That means allowing rapid innovation while building:

Logging and monitoring

Controlled permissions

The ability to update, disable or recall systems

Clear responsibility when harm occurs

Incident reporting and shared learning

Progressively stronger requirements as deployment becomes large and more consequential

The differs substantially from trying to specify every acceptable behavior in advance. It treats regulation as an adaptive system that learns alongside the technology. Excessive precaution can assume that future generations will be incapable of responding appropriately and timely to new problems. An oddly pessimistic view of both technology and society, a stifling pessimistic view. We should not be reckless, but neither should we be so presumptuous as to believe that today’s policy makers are the most capable and must therefore solve every possible future misuse before allowing progress.

The sensible balance is not “innovation vs. safety”, but rather innovation with the capacity to detect mistakes and recover from them. The truly dangerous alternative is either extreme: completely uncontrolled deployment, or a system so afraid of uncertainty that it hands technological leadership to countries more willing to learn by doing.

Europe’s danger is that it may regulate hypothetical future systems very carefully while becoming dependent on actual systems built elsewhere. In that situation, it sacrifices innovation without necessarily gaining meaningful control or safety. That is considerably worse than allowing bounded experimentation and improving the rules as exivence develops. Almost every genuinely important technology begins with unknown risks. Waiting until society understands a technology completely usually means waiting until someone else has developed it, deployed it and shaped its standards.