Making Sovereignty Measurable

Answers widely differ when organizations get asked about the meaning of “Sovereign AI”. For some it means a German or European model. For others, it means a European cloud, or hardware in a data center they can walk into. Two pieces of recent German research try to pin down some basic requirements for AI sovereignty. One comes from the people who buy AI for the public sector, the other from the people who regulate it. Both try to make the degree of sovereignty measurable. This post summarizes what they found and what it means if you are running AI on enterprise data.

Two German Studies Define What Buyers and Regulators Ask For

MÖVE (“Modelle für die Öffentliche Verwaltung Evaluieren”, roughly “evaluating models for public administration”) is a benchmark built by Bundesdruckerei, the German federal printing and security company, together with partners including the Federal Office for Information Security (BSI) and Fraunhofer AISEC. It exists because public bodies were choosing language models by headline or by country of origin. The first paper scores 39 models on German administrative tasks (summarizing, answering questions about documents, extracting topics) and on governance criteria: hallucination, energy use, provider transparency, and alignment with values from the German constitution. The public site is a living comparison that is updated as new models arrive, the methodology is described in a research paper.

PABeLA comes from the regulator’s side. It is a system built by the German Research Center for Artificial Intelligence (DFKI) together with the state media authority of Saarland (Landesmedienanstalt Saarland), and it is described on the DFKI project page. It turns legal and societal criteria, drawn from the German constitution, the State Media Treaty, the EU Digital Services Act and the EU AI Act, into automated black-box tests that a regulator can run against an AI service. The first two tests cover discrimination and political neutrality. The authors ran more than 150,000 test runs across model families (read the research paper for details on methodology).

What The Research Says

1. Keep models and data under your control. MÖVE centers on open-weight models, the kind that can run on-premise or on local hardware, and it uses the models in the federal IT provider’s internal AI application as a reference. The assumption behind it is that sensitive work should not have to leave your environment.
2. Choose models per task, not by origin or size. The authors saw buyers favor domestic providers on the assumption that a German model must be better at German. The results do not support that. No model wins across tasks, top performers differ by use case, and small models range from among the best to among the worst. Organizations need to be prepared to pick the right model for the right task.
3. Be efficient. Energy use per query varies by a factor of 63 across the models tested, far more than quality does. Small models often come close to top quality at roughly a tenth of the energy.
4. Be documented and auditable. Training data, bias mitigation and compute are poorly documented almost everywhere. Signing the EU’s voluntary code of practice for general-purpose AI hardly changed the transparency score. On the regulator’s side, PABeLA shows what auditors want instead: repeatable tests, structured verdicts, and a reason recorded for every single case.
5. Give grounded, neutral answers. Faithfulness to the supplied source material ranged from 0.20 to 0.83 across models. In PABeLA’s tests, models drifted politically when primed with one-sided context, and some recognized a discriminatory request and then carried it out anyway.

Where The Data Layer Comes In

All these requirements have one theme in common. They are about the behavior of AI steps that touch your data: where the model runs, which model you picked and why, what it cost, what it was asked, and what it answered. Most of that is decided in the data platform, not in the model.

This is where Exasol as a sovereign analytics database that runs on-premises, hybrid, private or in the public cloud fits:

• Models next to the data. Open-source models run inside the engine through user-defined functions and AI functions in SQL, so data does not have to leave your environment for AI to run on it.
• Model choice per use case. Because models sit behind plain SQL and functions, picking a different model for a different task is a configuration choice, not a platform decision.
• Right-sized, predictable compute. Small models and functions run in parallel next to the data, on CPU or GPU, under predictable pricing with no per-token meter.
• Governed LLM steps. Exasol’s agent control plane treats each LLM step as a governed function with versioned prompts, schema checks and an audit trail. That is the kind of evidence an auditor asks for.
• Answers grounded in governed definitions. Semantic Views give agents one governed definition of the business to query instead of raw tables, and refuse questions that have no well-defined answer.

Sovereignty Checklist

1. Write down your deployment boundary. Which data and which models must stay on-premise or in a given jurisdiction?
2. Evaluate on your tasks. Use public benchmarks such as MÖVE as a starting point, then test candidate models, including small ones, on your own documents and questions.
3. Plan for audit evidence from day one. Version prompts, validate outputs against a schema, and keep a record per call.
4. Ground answers in governed data. Give agents a defined model of the business rather than letting them improvise joins.

Conclusion

The studies present an approach that makes Sovereign AI demonstratable. Buyers want open models they can run where they choose, picked per task and judged on quality, energy use and documentation. Regulators want repeatable tests with a recorded reason for every result. Much of the evidence both groups ask for is produced where models meet data, which makes the data platform a central part of the sovereignty story: it decides where models run, how each AI step is governed, and whether answers rest on governed definitions. Organizations that treat sovereignty as a checklist they can evidence will be ready when procurement and audit ask. Exasol is built for that: a sovereign analytics database that keeps models, data and records together, under your control.

Sources

• Dalerci, Michael, Schaefer, Weinland. MÖVE: A Holistic LLM Benchmark for the German Public Sector. Bundesdruckerei, arXiv:2606.13111, June 2026. Live results: moeve.bundesdruckerei.de
• Schaffer, Kleer, Lessel, Feld, Bauer, Schmelz, Goedert. PABeLA: A System for Auditing Societal and Media-Regulatory Criteria in Large Language Models. DFKI and Landesmedienanstalt Saarland, Mensch und Computer 2026. Project: DFKI