Responsible AI in Prague

SciComm Meetup: Communicating Responsible AI

Thu, 25 Jun 2026 00:00:00 +0000

How do you talk about Responsible AI to someone who has no idea about how AI works, and how much can you simplify before the message stops being true? On June 25, we shared our thoughts on these issues at the SciComm Meetup: Finding Your Voice, an evening organized by the CTU Department of Research and Development Activities at Café Husovka in Prague. The meetup brought together students and researchers interested in publicly sharing their work, with talks from cybersecurity researcher Veronica Valeros, space ambassador Anna Krebsová, tech archaeologist Sara Polak, and the two of us on behalf of the Responsible AI in Prague Initiative.

Our contribution was a case study rather than a how-to guide. Over the past three years, we have presented about the ethics of AI and the need for its explainability to audiences ranging from conference attendees and interdisciplinary workshops to high school students at open days, festival crowds, and European youth workers. That gives us a good idea about how the very same talk changes when the audience changes.

We focused on two dimensions where the approach differs. The first was structure. For fellow academics, a talk typically moves from the importance of the issue, through the argumentation, to the interpretation of the results. For a general audience, the sequence looks different: first earn attention and relatability, then interact, and finally land a takeaway message. The contrast is easiest to see in how one motivates the need for AI explanations. For researchers, the regulatory requirement under the AI Act’s right to explanation can be a compelling opening, since they already know what the AI explanation means. For everyone else, we need to show what we mean by an AI explanation and create the need for it at the same time. In that case, a relatable story about an automated cheating-detection system that causes real harm through a false accusation is a much better fit.

The second dimension was risk, and this is where we think the interesting part lies. Academic communication is somewhat guarded against unsupported claims and lapses in scholarship. Public communication has to worry about calibration: the pull towards sensationalism on one side and oversimplification on the other, towards either fear or uncritical optimism. Both failure modes are easy to find in current AI coverage, and we brought a small gallery of headlines to prove the point. The worst-case outcomes are analogous for both types of audience. From an academic audience, we would hate to hear “explanations are useless, the EU is just adding regulation” or, equally wrong, “with explanations mandated, we will have safe systems tomorrow”. From a general audience, “I got nothing out of that, and I hate AI anyway” is as much of a failure as “AI will solve everything, just build more datacenters”.

The conclusion we offered is that scientific and public communication are not categorically different disciplines but a spectrum. The room environment, cognitive load, and attention span matter in a lecture hall too, and pruning jargon while assuming minimal prior knowledge rarely hurts an expert audience either. What changes between the two ends of the spectrum is mostly the entry point and the tolerance for detail, not the obligation to be accurate.

We are grateful to Karolína Pštross and the organizing team for putting the meetup together and for the discussion that followed, and to Petr Lebeda for the photographs accompanying this post. Judging by the number of researchers in the room who wanted to start communicating their work, the meetup deserves to become a regular event at CTU.

CEFIG International Study Visit: Discussing Responsible AI with European Youth Workers

Wed, 25 Mar 2026 00:00:00 +0000

On Wednesday, March 25, the Responsible AI Initiative hosted a study visit for 24 participants of an Erasmus+ Youth Training Course “AI Unlocked” coordinated by CEFIG International. The group, consisting of youth workers, NGO staff, and educators from various European countries, gathered to explore the ethical and safety dimensions of artificial intelligence.

The visit was part of a week-long training in Kolín focused on AI usage in education. Having already spent several days discussing digital security and practical AI tools, the participants came prepared with a solid foundation and hands-on experience.

During the session, we focused on several critical areas of Responsible AI:

Large Language Models vs. Traditional ML: We discussed the distinction between highly visible but often unreliable generalist LLMs and the narrow, domain-specific expertise of traditional machine learning.
The AI Act Framework: We explored the EU’s risk-based approach, distinguishing between “High Risk” systems used in recruitment or law enforcement and “Unacceptable Risk” systems involving social scoring or subliminal manipulation.
The Alignment Dilemma: A key point of discussion was the challenge of “alignment”—the process of tuning AI to human expectations. We examined whose values are being taught to these models, noting the influence of Western cultural norms and the role of the Global South workforce in data labeling.
Beyond the Black Box: We addressed the fundamental bottleneck of AI—its lack of transparency. We introduced the concept of Explainable AI (XAI), demonstrating how visual and importance-based explanations can help humans understand, verify, and potentially challenge AI decisions.

The presentation was followed by an interactive discussion where participants applied these concepts to their own work and asked insightful follow-up questions. We were thrilled to see such high interest in Responsible AI, and we hope these conversations will foster critical thinking regarding the next generation of AI systems.

Debates at One World: Debating In the Belly of AI

Sun, 22 Mar 2026 00:00:00 +0000

This March, we participated as expert discussants at three screenings of the documentary In the Belly of AI (V útrobách AI), held as part of the One World (Jeden svět) international human rights documentary film festival, one of the largest such festivals in the world, organized annually by People in Need (Člověk v tísni). This year’s edition ran from March 11 to April 24, 2026.

The film offers a critical look at the infrastructure underpinning today’s large language models. Through testimonies from experts and people directly affected by AI systems, it examines the less visible costs of the technology: low-wage data labeling work in the Global South, psychological damage suffered by content moderators, and the extractive mining of minerals that power our devices. These are framed against the ideology of longtermism and the priorities of the Silicon Valley elite.

While we appreciate the film raising these questions for a broad audience, we found its tone and some of its factual claims to be somewhat one-sided. The film mentions there are 145 to 435 million data workers, which sounds staggering, but on closer inspection, it encompasses all online gig work, including web developers, graphic designers, administrative assistants, and many others. Using such a broad category to illustrate the specific plight of AI data labelers inflates the picture considerably and muddies an otherwise legitimate concern.

A more subtle issue the film raises—perhaps without fully confronting it—is the widening divide between the Global South and the North. People in lower-income countries are currently doing significant amounts of data labeling and similar low-skill AI support work. But this work is inherently self-undermining: as AI systems improve, partly thanks to that very labor, the need for human labelers will diminish. The same dynamic is already visible in other routine service jobs (like call centers) that are being automated away. These are young people spending formative years on repetitive, unskilled tasks. What happens to them once the systems they helped build no longer need them, and when they have had little opportunity to develop more transferable skills in the meantime? The film gestures at exploitation but doesn’t fully reckon with this longer-term structural trap.

We participated in post-screening debates twice with student audiences and once with a general public audience. The student discussions were particularly engaged. The question that perhaps best captured the mood was a simple but important one: What can we actually do about this? Our answer was twofold. At the individual level, the effect is limited to being mindful of the tools we choose to use and how we use them. The more effective lever is political: electing representatives willing to put real pressure on technology companies, especially at the EU level, where regulatory frameworks with similar scope can take shape. And beyond the systemic, we also noted that some of the most direct harms, including cases where AI chatbots have allegedly contributed to users harming themselves or others, call for a more personal response: staying attentive to the people around us and not leaving people close to us navigate these technologies entirely alone.

AI in Research: The Dilemma Game

Wed, 14 Jan 2026 00:00:00 +0000

Should you submit a manuscript that is fully ideated and written by an LLM? What should you do if you get a seemingly AI-generated review of your paper? Dilemma Game is a discussion-based card game (and mobile app) encouraging honest dialogue about scientific integrity. Recently, our RAI team had the opportunity to work with the National Technical Library (NTK) team on a special edition dedicated to novel academic dilemmas posed by AI.

The Academic Services team at NTK organizes regular events for PhD students and other early-career researchers. One of their recurring endeavors is local facilitation of the Dilemma Game, a game developed at Erasmus University Rotterdam that provides academics with realistic dilemmas and uneasy choices. The game deliberately selects academic scenarios independent of players’ backgrounds, touching on topics ranging from publishing ethics to funding issues and team dynamics, with the aim of sparking a lively discussion in which participants try to persuade their colleagues of their truth.

Since the original game predates the recent boom of generative AI, we were happy to be approached, through reference of our founding member Lucie Stewart, by Eliška Skládalová from NTK with the ask to brainstorm a possible “AI update” of the game: creating new dilemmata focusing on issues with AI generation, authorship, or chatbot dependence. The collaboration resulted in ten brand-new scenarios and was ready for a test run at the event on January 14. Below is a small teaser for what the newly designed scenarios look like, learn more about the outcome in Eliška’s post.

Mixed with some scenarios from the original deck, the AI dilemmata prompted debates on the tradeoffs between textbook choices and practical complications. Participants differed in their interpretations of the scenarios and their ultimate decisions, as well as in their willingness to use AI tools in their own research. Still, it should be noted that participants of the Dilemma Game events are likely subject to certain selection bias (through their decision of attending an ethics-based game), and we, at the Responsible AI initiative, believe that a broader “survey” on people’s sentiment about generative AI usage in research would be beneficial, especially in the computer science circles which are often portrayed as highly techno-optimistic.

Explainable AI at FEE CTU Open Day

Fri, 28 Nov 2025 00:00:00 +0000

Should we follow AI predictions? What kind of explanation of AI’s processes would persuade us? What does explainable AI (XAI) really mean? And which XAI method should we use? On Friday (28 November 2025), we posed these questions to our audience at the Open Day of the Faculty of Electrical Engineering at the Czech Technical University in Prague.

Most of our listeners (and potential students) had heard about explainability for the first time, so it was crucial to introduce the topic in an approachable way and to showcase the need for explanation through a relatable example. After discussing the problem of black-box AI, we turned to a tried-and-tested scenario of cheating detection during exam proctoring and played a recording of slightly suspicious student behavior, leaving many participants unsure. Was the student cheating or not? What should I do when the AI detects cheating?

One of the most interesting outcomes for us was the significant division in opinions across the nearly 20 groups we presented to. Some people clearly saw cheating even in the most unlikely situations, while others were more sympathetic to the proctored student. Some said they would reexamine the video after hearing the AI’s decision, while others wouldn’t have batted an eyelid. The only answer that remained fairly consistent was that almost nobody wanted to overturn their decision solely based on the AI’s recommendation.

As regular readers of our website know, this is exactly our main case for XAI: the only reliable way to be persuaded by an AI recommendation is to uncover the reasoning behind it. We then demonstrated that the XAI landscape is vast and that one must carefully select methods based on the audience for the explanation and the application at hand. It was encouraging to see most listeners left convinced of the usefulness of XAI.

We hope that our listeners gained valuable insight into explainable AI and will take to heart our final pledge to consider the broader impact of technology, particularly if they decide to pursue a career in computer science and artificial intelligence at FEE CTU or elsewhere.

AI Days: Understanding AI-Enabled Injustice And Biases

Wed, 12 Nov 2025 00:00:00 +0000

When can we trust AI? What does it take to overturn complex AI decisions? How biased are such decisions made by today’s state-of-the-art AI models? And, is the accuracy of AI models all we need? We shared these and many other questions with the audience of the 2025 edition of AI Days, a two-week program spanning 36 Czech cities that featured both public educational events and researcher meetups.

Last year, we had a great experience with a presentation on AI ethics for technically oriented students. This year at the AI for Talents event, we took on another type of challenge: two talks (6. and 7. 11.) for a large audience of high-school students with diverse backgrounds. Aiming for an interactive experience, we asked the question: What kind of information would you like to have when an AI makes an important decision about your studies/career/life? In a realistic online test scenario with an automatic proctoring system, the audience had to decide whether to trust the AI’s decision or override it. And while the concrete results of this interactive experiment varied from one group of students to another, one lesson always persisted: some of the decision instances remained highly ambiguous. How to resolve the uncertainty then? We provided two types of explanations of the automated decision process: one based on visual cues and the other on the importance of extracted features. Yet, this move only introduced another dilemma: What is a good explanation of AI prediction? We briefly explained that this is exactly our current work in the RAI initiative and urged our audience to always think critically about AI-based decisions.

A couple of days later, we were invited to discuss the influence of AI on the LGBTQ+ population as a part of an event on 11. 11., asking a central question: “Is AI an ally or an enemy of LGBTQ+ diversity?” We briefly introduced how Large Language Models (LLMs) work and what makes them different from other, earlier types of AI. We then described the dependence of the models’ capability and “safety” on the data provided for training, describing the secretive nature of cutting-edge model training and psychological demands on workers in third-world countries. We finally discussed some downsides (and possible upsides) of AI models for LGBTQ+ people and minorities in general. We concluded the talk with a couple of suggestions on how current AI tools should (and should not) be used, based on our insights. Afterwards, we had a fruitful discussion with Juraj Hvorecký, a philosopher from CETE-P.

Our contribution to the AI Days would, however, not be complete without pitching our ideas to our peers at the Young AI Research Forum (12. 11.). We decided to take a bird’s-eye view of our work and pinpoint the arguments for pursuing explainability research, simultaneously preaching the necessity for interdisciplinary dialogue. While we were initially unsure about fitting in with this non-technical overview of our efforts, the topic perfectly complemented other talks by following the overarching topic of this year’s forum, interdisciplinarity.

On behalf of all RAI members, we thank the organizers of AI Days, especially the amazing team from prg.ai!

RAI in Bologna: A Look Back at The European Workshop on Trustworthy AI

Sun, 26 Oct 2025 00:00:00 +0000

The TRUST-AI workshop, held during the weekend (25.–26. 10.) before The European Conference on Artificial Intelligence 2025, set out to examine what “trustworthiness” in AI really means, and whether the research community can ever settle on a shared definition. Judging by the conversations, most participants had already discovered (the hard way) that agreeing on a common vocabulary across disciplines is roughly as easy as explaining your AI research to your grandparents. So our interdisciplinary proposal for assessing explainable AI (read the accepted paper) found itself in very sympathetic company.

Over one and a half days, the workshop offered a rich selection of perspectives: neurosymbolic explanations of AI behaviour, ethical analyses of high-risk AI systems, and even a developer-friendly compliance toolbox. By the time I presented our work on assessment dimensions for explainable AI on Sunday morning, no one needed reminding that trustworthiness (explainability included) is not just a fashionable trademark on AI papers, but something we actually need to operationalise. After all, thanks to the previous day’s groupwork session, where we were split into random teams and wrestled with a shared proposal that needed to be finalized in under two hours, the organisers’ gentle plea for a shared language no longer felt aspirational; it felt like a survival skill.

Despite the overflow of ideas, the audience remained impressively focused, at least judging by the lively Q&A that followed. Questions ranged from whether we should broaden our scope beyond the harm-oriented explanations required by Article 86 of the AI Act (towards more proactive, pre-emptive types of explanation), to deeper debates about the feasibility and scalability of assessing non-technical dimensions. In the workshop spirit, the discussion managed to validate our current direction while simultaneously pointing to the next round of challenges, just enough to keep things interesting.

RAI at Ethical AI workshop at ACM KDD conference

Sun, 03 Aug 2025 00:00:00 +0000

On the 3rd of August, in Toronto, the 4th Workshop on Ethical Artificial Intelligence took place. It was held in conjunction with the ACM SIGKDD’25 conference, and we had the pleasure to present our work. It was the latest iteration of the assessment framework for XAI methods through the lens of the AI Act. This work shall help practitioners choose suitable XAI methods for their use case (or even influence the decision about the choice of the predictive model).

The accepted paper and the oral presentation were well received, and we even put up a poster for the afternoon, which also sparked some discussions. Overall, it has been a fruitful event, with important topics discussed, such as the various theories of bias or the exponential growth in reported AI incidents (where staggering 55% are from third-party tools). In general, the reception of our work showed us that we are moving in a meaningful direction. We hope to finalize the work and ready it for a major submission in the coming months.

RAI at KU Leuven: Assessing Explainability for AI Safety Governance

Tue, 27 May 2025 00:00:00 +0000

At the International Conference on Large-Scale AI Risks (May 26-28), we presented the second iteration of our work on a safety- and governance-informed assessment framework of AI explainability techniques. The proposal, combining AI safety and AI regulation desiderata, proved to be well-received by the interdisciplinary audience of computer scientists, philosophers, lawyers, and policy makers, and complemented very well the diverse topics of the conference (see the book of abstracts), ranging from the moral status of AI to value alignment, economic impacts of AI, and safety governance. Check out our slides, or read the accepted abstract below:

AI safety has not only been a matter of academic but also of public interest, culminating in a call for an AI moratorium in March 2023 (Future of Life Institute, 2023). Governments around the world have since taken action to promote the safe development of AI. For example, the UK, the US, and Japan have founded National AI Safety Institutes (AISIs), and other countries have followed since (Allen & Adamson, 2024). AISIs are tasked with the safety evaluation of advanced AI systems, contributing to standards with technical expertise, and strengthening international cooperation (Araujo et al., 2024). The EU AI Office has largely similar roles to AISIs, with the additional mandate to support the implementation and enforcement of the AI Act.

The success of governance initiatives for AI safety depends on three components: i) specification of technical and legal means by which governance initiatives, such as the AI Act, shall be implemented; ii) stringent legal enforcement of regulation; iii) meaningful human oversight. However, safety criteria are impossible to fully specify in technical terms and to enforce at scale for current AI models that are deployed in dynamically changing, complex environments. We thus propose approaching i) standard specification, ii) auditing, and iii) continuous human oversight with explainable AI (XAI) methods (Došilović et al., 2018; Schwalbe & Finzel, 2024) that allow stakeholders to react flexibly as new safety concerns arise.

Arguing that not all existing XAI methods are equally beneficial for AI safety, we review their usage in safety-related case studies based on AI Act classification and propose a 5-dimensional framework for their assessment. The first two dimensions are understandability to subjects and auditors. While (non-expert) subjects require simple and succinct justifications, auditors can be expected to process more involved explanations utilizing, e.g., their understanding of statistics. The third dimension is veracity, which measures how accurately an XAI method represents the model’s behavior. The penultimate dimension is actionability, which evaluates the helpfulness of an explanation in resolving possible undesired behaviors of an AI model. Finally, there is scalability, ensuring that even the largest state-of-the-art models can be explained with an XAI method. We suggest that a joint evaluation of the presented (possibly interdependent) dimensions is an essential part of a holistic approach to AI governance, bridging the gap between technical development and regulation.

References

Gregory C. Allen and Georgia Adamson. The AI Safety Institute International Network: next steps and recommendations, 2024. Accessed: 2025-02-13.

Renan Araujo, Kristina Fort, and Oliver Guest. Understanding the First Wave of AI Safety Institutes: characteristics, functions, and challenges, 2024.

Filip Karlo Došilović, Mario Brčić, and Nikica Hlupić. Explainable artificial intelligence: A survey. In 2018 41st International convention on information and communication technology, electronics and microelectronics (MIPRO), pp. 0210–0215. IEEE, 2018.

Future of Life Institute. Pause Giant AI Experiments: an open letter, 2023. Accessed: 2025-02-13.

Gesina Schwalbe and Bettina Finzel. A comprehensive taxonomy for explainable artificial intelligence: a systematic survey of surveys on methods and concepts. Data Mining and Knowledge Discovery, 38(5):3043–3101, 2024.

RAI at POSTER: Explainability for AI Alignment

Thu, 22 May 2025 00:00:00 +0000

On May 22, we presented a poster summing up the current outcomes of our collaboration with Paula Gürtler at the 29th International Student Scientific Conference (POSTER) in Prague. The work titled Dimensions of Explainability in AI Alignment stood out during the poster session and the subsequent elevator pitch as an original, interdisciplinary proposal. It sparked interest among the conference committee, and the positive feedback received is an important validation of our efforts. Read the full paper or the abstract below.

Human-AI alignment is challenging due to limitations in both technical solutions and governance frameworks. Given the infeasibility of properly anticipating all potential misalignment risks, we see explainability as essential for continuous oversight, bridging the gap between AI systems, governance, and human intervention. Recognizing the multi-faceted character of the problem, we argue for a structured framework for evaluating explainability methods, moving beyond narrow technical metrics, to enhance future developments in AI accountability and alignment.