Present discussions about European Digital Sovereignty underline the fact of an almost complete dependence of Europe and in particular of Germany on US technology – regarding

• shear computational power (supercomputers, data centers),
• classical IT, office and ERP applications (dominated by Microsoft, Google, Meta, …)
• as well as AI models and related computational capacities
• and, last but not least, the combination of robotics and defense systems with AI.

The standard recipe is a call for money to build more computation centers on European soil.

Actually, most of such centers presently under construction are planned to be eventually used by US companies. This is a sign of schizophrenia of European and German politicians or – even worse – a total lack of understanding of what happens in the fields of IT/AI and of what digital sovereignty really requires. Certainly not more data centers of Google, Microsoft or Meta in Germany. At least, this should not be a top priority item on an agenda for European independence of US monopolists.

Related concerns of European experts working in the fields of IT and AI culminated last week in the publication “https://europe2031.ai/“. The paper and the scenario described in it was triggered by the present failure of European governments to take and scale counter-measures against a coming US-dominance in the little remaining interval of time – if there is any at all. Visitors of this blog have probably read the paper, already.

From my perspective as a former IT consultant, I want to add three points which the scenario in the “Europe31” paper does not cover to the extent these points deserve.

The first point is “education“:

AI is predominantly the result of certain types of SW, the development of artificial neural networks, the theory of artificial neural networks and advanced ideas about learning and learning algorithms. But, and this is totally underestimated., it i also based on a lot of mathematics (in particular regarding Linear Algebra, analysis of data distributions in multiple dimensions, differential geometry, theory of graphs, optimization problems and advanced statistics).

Top hardware is also an important aspect, but as China has shown with its recent CPU-based supercomputer LineShine not a predominant factor. You can build sufficient supercomputer power also based on previous generations of hardware. What is much more important is the bundling and focusing of efforts – as well as a support of solution oriented thinking. Regarding these points we have a lot to learn from China.

And: As in any high-tech field, real progress in AI was and is achieved by clever ideas.

Proof: Consider the impact of transformer based neural networks on the development of present LLM-based AI systems. One idea, enormous consequences.

Clever ideas in turn require educated people.
Consequence: European AI Sovereignty requires at least as much investment in the education of people as in hardware.

Now, compare this with the fact that the present conservative government of Germany plans NOT to rise the state’s budget for helping students (Bafög). Although the level of support has already for a long time been below minimum social standards guaranteed to other groups of citizens. The obvious stupidity of such a planning remembers strongly of the Kohl era when physicists were told that Germany did not need theoretical researchers, but application oriented engineers. As if this had ever been a contradiction … 15 years later the same advisors who had told Kohl their “wisdom” suddenly found out that Germany did not receive enough Nobel prizes.

Well, as a relatively old person, you automatically think of the 5 apes in action, which unfortunately have become so characteristic for conservative German governments during the last 3 decades:
(1) Don’t hear and listen (to warnings of scientists), (2) don’t see (but ignore obvious developments in high tech realms), (3) don’t speak (about problems and long term strategy or e.g. about the self-inflicted dependency on Russian gas during the Merkel period and not having invested e.g. in digitization and AI), (4) don’t understand (because a substantial lack of MINT scientists among politicians), (5) don’t pick up ideas beyond a conservative ideology of the last century (let the market alone do its job).

As an elderly, rather conservative person I do not utter this criticism lightheartedly.

But the real task is much bigger than supporting young academic students – we need a proper preparation of the present and the next generations to handle advantages and risks of AI properly. According to European standards. And not in 10 years, but now. In school, in companies and in daily life.

A democratic state should not leave this playing field to money driven interests of business managers. Education is a fundamental task of the state – and not of some market. Regarding the total failure of the German market and of German managers: A look into the present existential crises of the German car companies should provide a clear warning of what happens when people from the last century make plans for the technologies of tomorrow.

Regarding AI education: Where are the thousands of public adult of education centers for IT and AI required to bring the average German up to date (at least up to yesterday)? Where are the 10ths of billions required to support German universities regarding AI projects across all fields of science?

The second aspect is Open Source: We should not repeat the mistakes we Europeans have made over the last decades regarding closed source and using SW of de facto US monopolists. We need an Opensource environment with a quick exchange of ideas if we in Europe want to grasp our last chance for a development of our own AI systems on short timescales.

The usage of closed source during the last decades has created the extreme dependency of the German administration and of German companies on US tech giants. Most remarkable is the permanent ignorance of warnings coming from experts already in the first decade of this century on national and EU levels regarding the growing dependence of governmental institutions and public services across Europe, but especially in Germany on US technology. Plus a lack of initiatives to invest in European solutions in IT, communication and satellite technology. Or to invest in digitization, at all.

Since 2002 a proper digitization of governmental and administrative public services has been discussed in multiple waves by German governments – and where are we now? Place 22 or worse on European scale of competencies and capabilities? With a full addiction to Microbesoft … predominantly due to the ignorance of conservative and often enough rather old politicians.

Another point which favors open source is the (unfortunately) continuing existence of national barriers and interests within the EU. Egocentric projects and endeavors (look at the FCAS project!) can best be overcome by states which strongly support a free exchange of information regarding base technologies. I leave it at that. There has been so much written about this point – and certainly I do not need to convince people working with Linux about the advantages of Open Source SW.

The third point is quality:
The third reason why we need an European path regarding AI is quality and the control of errors and mistakes the present generation of LLM-based AI will always make (by design, so to say).

LLM-based AI as a pattern extracting and pattern application machinery based on probabilities (additionally weighed and biased by the chosen sources of information for the LLM-training). Of course, stupid algorithms based on probabilities make errors and hallucinate in the range of any correlations seen during training. On a fundamental level this is due to the unavoidable statistical elements inherent in the process of generating answers to prompt requests.

You may limit the impact of statistics by secondary processes (e.g. reasoning steps) – but you cannot fully eliminate errors of present day neural networks. This is a trivial fact. The real problem, however, is that errors and mistakes do happen on a very elementary level within chains of thoughts about complex problems – but one may not see the mistakes directly. E.g, when a mathematical dialog with an AI goes over 30 or more pages. You do not believe that errors happen on basic levels of mathematical reasoning?

Well, my personal experience with AI regarding questions in physics, mathematics is two-fold:

AI can help you to collect ideas, you may not have thought of in the context of your questions and prompt requests.
However, when it comes to the question of complex coding or of mathematical proofs, each generation of LLMs has disappointed me. By in parts blatant errors the LLMs made from time to time.

An example: You ask for a Python program which should plot the gravitational potential of a disturbed, but basically spherical mass distribution with a certain 2-dim density profile and with a clear border at some radius and vacuum outside. On first sight the Python program generated by a GPT-4 system looked OK. Clear separation of tasks, parameter and interface description and a sequence of functions for the math of non-relativistic gravitational fields, …. However, when executing the program, the plot I got was totally wrong because of a standard continuity condition at the border of the mass distribution had not been applied by the AI. A very basic error, which the GPT-4 variant admitted after new prompts, apologized for multiple times and corrected after some pressure. The error happened with another system (of the GPT-4 generation) even for a one dimensional mass distribution.

Another example: I wanted a proof for an assumed relation between the principal curvatures of a certain multidimensional surface in comparison to the curvature of the level sets of this surface – in particular at those points on the surface where extreme values of curvature occur. I started with asking for a proof for finding the extreme points of the level sets from first principles. The proof covered many pages. The result was correct, but hmmmm, it was known from the beginning (and is well documented and normally proven by other more advanced methods). The thought line of the proof was the really interesting thing. But the line of argumentation of GPT-5.5 had a blatant error in its middle regarding a claimed, but obviously wrong relation based on a simple binomial formula. The value of the whole proof which was sold by the AI as based on a “remarkable cancellation” was zero.

In a related example regarding curvatures of multi-dimensional surfaces the AI diagonalized a wrong matrix for another “clever” proof. The AI detected its mistakes in hindsight, only, when asked for details of certain mathematical steps – which clearly were wrong.

For yet another example, see also here.

You do not expect such elementary errors in much more complex contexts – but they do happen much too often. In some cases a discussion with the used AI-variant about its errors indicated that it had looked for “short cuts” and omitted some necessary consistency checks (which – I quote the AI – “I should have done from the beginning”). Actually, there is a general tendency of current LLMs to shorten complex considerations or to hop over intermediate steps, which due to some unclear reason appear trivial to the AI. Very often, elementary and blatant mistakes happen such points, where the AI appears to lean onto some dubious patterns it has learned, but (of course) not on real insight – despite the fact that the wording may sound like driven by insight.

In most of the cases I have seen, the errors were unfortunately also hidden under the surface of long answers – especially regarding mathematical proofs. The mistakes were not really obvious – and I only saw them during a second close analysis of each single step. In all cases named above the AI frankly admitted its errors in hindsight and sometimes even gave me a reason for their occurrence.

But what do you make of an insight in hindsight like the one GPT 5-5 told me about after a thorough analysis of a rather basic mistake:

“The deeper lesson: Your check illustrates something very important mathematically: When manipulating equations with square roots, one must be extremely careful about: squaring, factorization, cancellation, branch structure. A single unjustified algebraic rewriting can silently change the solution set. You caught exactly such a mistake.”

You have to be careful when doing mathematical transformations … really? I would say careful transformations are a minimum requirement for mathematical reasoning … But the AI appeared to just have noticed it the first time.

This is the quality offered to people with freely available LLM-versions. Whether expensive Platinum versions would really do a better job is questionable. The nature of the errors do not indicate a lack of computational power, but a general lack of quality and respective checks. In addition I noted a tendency to invent steps to fulfill the questioner’s expectations.

To say it clearly: I do not expect insight of a stupid algorithm. But I expect sufficient checks of so called “reasoning systems” – in particular on a basic level of math. Presently, you sometimes do not get the necessary quality – at least not with what e.g. OpenAI offers on the European market. As I said: Blatant errors on an elementary level can happen any time – and often unexpectedly during a “conversation” about some complex problems. You need to check at least mathematical output carefully – and then the advantage of saving time by an AI is lost by a large percentage.

Conclusion: We need a different approach in Europe, one which is focused on quality and intrinsic error checking. We need clear warnings when capacity limits regarding context memory, computational time and the involvement of reasoning steps are reached.

In particular: We do not want AI-systems which makes blatant errors in questions potentially concerning life and death – e.g. when creating target lists for the military or developing formulas for precise tracks of ballistic rockets (be it in defensive or offensive scenarios). Or stupid errors in the analysis of data on climate change or of data of surveillance systems. And we certainly want to avoid errors in scientific processes.

The German industry once got strong because its governments had understood that monopolies seldom lead to real quality. But to break the development of monopolies the administration of a state must sometimes sponsor new ideas and create an environment in which these ideas can grow until they are mature enough for an international market. Having understood this has always been an advantage in the competition with the US.

Another success in was based on yet another fundamental insight: The top four priorities in a country with few other resources than the brains of our children and grand children must be education, education and education accompanied by sufficient social security. These are good, old and maybe conservative recipes.

Unfortunately, our present government seems to have forgotten both the virtues of financially securing education on all levels of school, universities, services and production, the virtue of strongly supporting a free exchange of knowledge about basic technologies and science and the virtue of quality and enforced quality assurance.

Or why do we accept the AI products of Microbesoft et al. when we undoubtedly have European capabilities to replace them by something better? I strongly doubt that our young students are not eager to prove this. But unfortunately, German politics obviously prioritizes the covering of costs for a system of privileges for so called “Beamten” (sworn in employees of the state’s administration) more than spending money for the an overdue adequate support of our students in comparison with average costs for a very modest living.

Enough for today. There are, of course, many other aspects of AI which Europe could use to its advantage. But hearing conservative German politicians defending cuts of already minimum social expenses and in particular expenses on the education of our young people makes me – a rather conservative person – sick. Not supporting students is irresponsible and certainly no sustainable measure on the way to general digital or AI sovereignty.