Digital sovereignty, data spaces and large language models

What does “digital sovereignty” mean in practice and how does it relate to the Swiss LLMs that are currently being developed? The article shows: Digital sovereignty requires national LLMs.

Digital sovereignty is a political goal for Switzerland that is practically unattainable in its entirety. The question therefore arises: To what extent should it be achieved?

Switzerland is dependent on other countries for key components of the digital ecosystem. But the EU is not much better off. And whether the USA has digital sovereignty is a much-discussed question. So how much digital sovereignty do we want?

If we formulate the goal more moderately, there are three questions. We can quickly tick off the first one. Is Switzerland open to digital blackmail? Answer: In principle, yes. It just depends on how drastic the threats are. But that also applies to Germany, which is ten times bigger. The question is always what is more expensive: resisting the blackmail or giving in to it. If the costs are similar, you should resist the blackmail. Even if they are slightly higher. But somewhere there is a limit beyond which blackmail will be accepted because resistance is too expensive. The exception is when giving in would go too far.

Possible means of blackmail are: Cutting Switzerland off from important digital technologies or from the international financial transaction system, in addition to the usual analogue blackmail such as the threat of immensely high tariffs. The possible blackmail targets: The renunciation of the regulation of market-dominant digital platforms (this has already been proactively offered by Switzerland in the customs dispute with the USA), the adaptation of laws (for example the BGEID, in such a way that national Swiss eIDs can be issued by large foreign corporations) or even the withdrawal of already existing laws (for example the Data Protection Act or at least data portability in the Data Protection Act).

The aforementioned means of blackmail are many times more expensive for Switzerland than the EU Commission’s pinpricks against Switzerland in the past (e.g. exclusion from the EU’s research framework programme). If Swiss banks only have limited access to customers abroad, that is one thing – many are already doing without American customers. But if Swiss banks can no longer do international business because they are digitally cut off, that would be the end of them.

However, these and other horror scenarios are very unrealistic, which is why the impossibility of complete immunity to blackmail should not cause any further concern. Those who plan for the improbable usually fail because of the predictable. Two questions therefore take centre stage (which also play a central role in the costs of blackmail): Firstly, can Switzerland develop digitally customised solutions for itself, or does it have the eternal disadvantage of having to use solutions from other countries that the Swiss population does not fit in with? That sounds cynical, but it’s normal everyday IT life, even if it’s only anecdotal evidence. I have seen the IT managers at a university complain that the rector’s directive did not fit with their ERP system.

Apart from such cultural peculiarities of the IT scene: when Swiss researchers buy medical records in the USA, they have exactly the same problem. The data does not fit the Swiss population, because lifestyle, genetic distribution and healthcare are different in Switzerland than in the USA. Although we have various data science methods to reduce the problem, it remains strange that we are working with American data here due to data protection. This is ethically questionable to say the least – and it would be easy to justify if the American government put an end to this activity. If Switzerland does not allow Swiss data to be exported to the USA, then no American data should be exported to Switzerland. That would only be fair – except that there is hardly any personal data available for research in Switzerland. Switzerland is lagging far behind China and the USA in the area of secondary data utilisation.

The second question is: can we preserve our values, or must we allow the values of those who are technologically superior to us to be imposed on us? Objectively speaking, the question is a strange one. This is because many people in Switzerland (and in other European countries) hope that technology can be used to combat their political opponents. The idea is to use technology to create fact-checkers against fake news. In plain language, this means that technology should be used to spread credible fake news, as fact-checkers are known for spreading fake news themselves (and have already been convicted for this in Germany). But technology should also guarantee freedom of speech. In plain language, this means that technology is being used to drive forward the destruction of liberal democracy. Because much of what has been developed digitally in the name of freedom in the past is being used very successfully by anti-democratic political forces and criminals – under the label of freedom of speech.

This illustrates that the undermining of liberal, semi-direct democracy is also being driven forward domestically with a great deal of idealism via technology projects. But the question is, is there no alternative to the use of foreign technology based on foreign values or not?

The very specific versions of Christian thinking that are popular in Silicon Valley show that this question is not harmless. There, digital entrepreneurs compare themselves to Jesus Christ (Daub, 2020) (Thiel, 2014). They sincerely believe that they must redeem the world – including from the evil of liberal democracy. Of course, this is a legitimate view, but when it is invisibly coded into algorithms, it becomes an instrument of power. That is why there should be alternatives to digital services and tools from abroad that implement our values rather than values that are alien to us. In Switzerland, these include, in particular, openness to debate with other opinions, political neutrality of information services (unless otherwise declared) and multilingualism. Open source solutions are particularly desirable in this context.

The first question discussed above makes it clear: We (Switzerland, other European countries, the EU, Europe) need our own data spaces! The Informatics and Law conference on 26 August in Bern City Hall will discuss this in detail and in principle. The second question suggests investing much more than before in open source. Particularly and especially in the field of artificial intelligence.

What is needed are indigenously trained AI models that reflect indigenous thought logic and indigenous values. Specifically and in particular in the area of large language models (LLMs). Switzerland needs large language models that have been developed according to Swiss values. Specifically, this means above all no data theft, i.e. no use of text without the consent of those who own the copyright. Furthermore, the values of Switzerland should be reflected (for example, the equality of all people) and the LLM should be geared towards multilingualism. This is because multilingualism is an essential element of Swiss culture.

ETH and EPFL will develop such national LLMs on the new ALPS supercomputer at the Swiss National Supercomputing Centre (SCNC) in Lugano so that they can then run as services from conventional cloud computing service providers. This requires experience and therefore takes time and electricity, but has the advantage that the experience gained can be used for many sector-specific specialised LLMs, while also consuming less electricity thanks to the efficiency gained through experience. The services of ALPS are also to be offered to private Swiss companies in the future – ideally supplemented by consultancy services from the universities. The corresponding activities were presented and discussed at length at the 7th Digital Summit organised by Digitalwitzerland on the Bürgenstock.

The first step will be the Swiss Justice Base Model, an LLM for Swiss law. It is obvious that a national solution is needed here. After all, apart from a few exceptions, law is a national matter. In addition to published, anonymised court judgements (a success of the open access movement), legal texts provided by publishers are also used for training the LLM with the consent of their authors. At the same time, principles of a fair use regulation will be developed that are suitable for Switzerland – or rather several versions of such regulations, between which the developers of LLMs may choose (because there are also different licence models elsewhere). Of course, only those fair use regulations will be selected whose compatibility with Swiss law is beyond any reasonable doubt.

Change managers have long been preaching that quick wins are crucial. The Swiss Justice Base Model should be such a quick win on the way to partial digital sovereignty!

References

Daub, Adrian(2020). What the Valley calls thinking (4th ed.). Suhrkamp.

Thiel, Peter (2014). Zero to one: notes on startups, or how to build the future. Chapter 14.