SciComm Meetup: Communicating Responsible AI

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.