Local, verifiable, under human oversight. How responsible AI is taking shape in Czechia

The question CZDEMOS4AI sets itself sounds straightforward: can generative AI be developed as trustworthy, publicly beneficial infrastructure? One that helps an information-overloaded society make sense of complex topics, decide on the basis of good data, and still keep the final say in human hands? The answer is anything but simple.

Led by the Institute of Political Studies at the Faculty of Social Sciences, Charles University, and principal investigator Professor Nik Hynek, the project runs until 2029 and is funded by the Technology Agency of the Czech Republic under its SIGMA programme. Its research teams met on 8 June 2026 to put their four emerging assistants in front of direct feedback from people outside the project for the first time. Hynek opened the conference by recalling the thread running through all four areas: an emphasis on ethics, legal clarity, and social responsibility.

The state as a network of AI assistants

The conference’s tone was set by the keynote from the Government Commissioner for AI, Lukáš Kačena. He laid out the priorities of Economic Strategy 2.0 – conditions for deploying AI in business and public administration, support for research, and access to infrastructure – pointing to the AI Factory in Ostrava and the National Centre for Artificial Intelligence at the Czech Technical University as its first results.

At the heart of his talk was a vision of an agentic state on the Estonian model. A citizen would bring their problem to an entry-point agent that understands the query and passes it on as a concrete task to the specialised agents of individual authorities. Each official would have their own AI assistant for working with documents and legislation – yet would still make the decisions themselves. And the individual agents would talk to one another over a shared protocol. The state would be built on its own language models, where every single decision is traceable, and therefore open to scrutiny.

Why do it locally

Why ordinary large models fall short of that goal was explained by Petr Špelda of FSV UK, head of the project’s technical platform, in a talk that introduced attendees to its technical framework. As he noted, the development of commercial models is not primarily aimed at the alignment that would guarantee answers consistent with the facts and norms of the Czech context. The state’s resilience, in Špelda’s view, rests on the resilience of its experts – so the project sets out to effectively involve specialists working in public administration in building its dedicated AI models.

Martin Popel of the Faculty of Mathematics and Physics at Charles University then showed how the model built by MFF UK and the company MAMA AI handles basic queries and produces answers drawing on internal sources, such as the Czech News Agency (ČTK) database. He also walked through how output quality is assessed on three levels: first it gathers feedback from relevant experts, then from members of the public taking part, and finally from AI itself.

Four assistants, four lessons

After the opening, the floor passed to the workshops, where visitors tried the individual assistants out for themselves.

The history-and-education avatar, led by Čeněk Pýcha of the Faculty of Arts at Charles University, is meant to help teachers convey topics in history education. The workshop served as its very first focus group and immediately ran into the gap between what an expert expects from an answer and what a layperson does – the public asks differently, and the model struggles to interpret the questions correctly. Those findings will now feed into revising sources and fine-tuning prompts.

The security-and-media assistant is a joint effort by Angelika Suchanová of FSV UK, Martin Čmejrek of MAMA AI, and ČTK. It targets situations in which verified information tends to get lost – the 2024 floods, the war in Ukraine or the Green Deal. It is built on ČTK’s independent data and is currently being tested on the floods. Early results suggest more specific answers than the usual tools give. The team stressed, though, that the aim is not to offer a competitor to the large models, but rather a replicable direction for development at national level.

The legal assistant, led by Ján Matejka of the Institute of State and Law of the Czech Academy of Sciences, is being built on a corpus of roughly 300,000 legal documents and is to be linked to the e-Sbírka, the state’s electronic collection of legislation. Matejka stressed that law leaves no room for probabilistic answers – everything must follow the rules and the legal order. A serious technical problem identified here is AI’s non-deterministic behaviour, where the model responds to an identical query with markedly different, even contradictory, outputs. The auditability of sources is therefore becoming a foundational principle for the entire legal strand.

The mental-health assistant, overseen by Roman Gabrhelík of the First Faculty of Medicine at Charles University, is intended to support professionals working with digital addictions. The workshop mapped what they need from such a tool: a quick survey put transparent development and access from multiple devices in first place, followed by working with vetted sources and tailoring to specific cases. It also opened up the question of the model’s tone of communication and the handling of clients’ sensitive data.

Ethics keeping pace

The conference closed on ethics. Anetta Jedličková, chair of the ethics board, welcomed the fact that the project’s ethical dilemmas are not lagging behind the technology but keeping pace with it. As principles running through every team, she named transparency and clarity, the verifiability of outputs, privacy protection, the suppression of bias, clearly defined accountability, and expert oversight of both development and deployment.

Above it all stands a single ambition for the project: not to replace human decision-making but to strengthen trust in it – with verifiable data and the right context.