I am a Full Professor at the University of Warsaw and serve as Director of the Centre for Credible AI at Warsaw University of Technology. I lead a research group working at the intersection of mathematics and computer science focused on deeper understanding of AI models.
My work asks a question benchmarks cannot answer: what is actually going on inside a model, and can we trust it where it will be deployed? Over the last decade this has meant new explanation methods, adversarial analysis of those methods, open-source tooling used by tens of thousands of practitioners, and a research programme I call Model Science.
AI is broken,
and we are here to fix it.
Hundreds of millions of people now depend on models that essentially no one understands. Benchmarks saturate faster than new ones can be written, and a high leaderboard position tells you remarkably little about whether a system will hold up in a hospital, a courtroom, or a bank.
Model Science treats a trained model the way other mature disciplines treat their objects of study: as something to be probed, described, catalogued and reasoned about empirically. Cognitive science combines complementary analytical perspectives; neuroscience learns from deep single-case studies that population averages hide; medicine separates diagnosis, treatment and prevention into distinct but connected institutions; agriculture built shared infrastructure so that results accumulate instead of evaporating. AI research has none of this yet.
With Wojciech Samek (ECAI 2025) and in the fuller 2026 agenda paper, I organise the field around four questions.
Does the model perform reliably where it will actually be deployed?
Benchmarking is necessary and nowhere near sufficient. This pillar covers context-aware evaluation protocols, red teaming, auditing, robustness under distribution shift, and the systematic study of edge cases. Our work here includes red-teaming vision foundation models and analysing predictive multiplicity — the uncomfortable fact that many equally accurate models disagree on individual cases.
What computational mechanisms generate the output?
Attribution methods, concept-level analysis, sparse autoencoders, counterfactual explanations, and interactive tools that let a human actually interrogate a model. The framing matters: explanations exist to question a model, not to justify it after the fact — a position I have argued at ICML.
How do we change a trained model's behaviour without retraining it?
Once we can name what a model has internalised, we can intervene on it. Feature-level editing, representation steering and alignment techniques let us move a deployed system toward the behaviour we want — cheaply, reversibly, and with an audit trail.
Which training decisions produced this behaviour?
The loop closes here: connecting data, objectives and fine-tuning procedures to emergent misalignment, catastrophic forgetting and reward hacking. Understanding a model is only half the point if nothing flows back into how the next one is built.
None of this scales without shared plumbing. Model Science calls for three things the field still lacks: dataset catalogues built on FAIR principles so that claims can be stress-tested across dozens of datasets rather than one; model catalogues that record lineage and inherited properties, because models are increasingly descendants rather than originals; and shared findings repositories — something like MITRE ATT&CK for model behaviour — so that a failure mode discovered once is known to everyone rather than rediscovered a hundred times.
Hierarchical sparse autoencoders that pull interpretable concepts out of CLIP-class models at a state-of-the-art sparsity/fidelity trade-off, and the question of what a "concept" in such a model even is.
Region-constrained visual counterfactuals, model-agnostic symbolic counterfactuals for time series, and global counterfactual directions — moving from "what mattered" to "what would have to change".
Explanations can be attacked. We showed partial dependence can be fooled by data poisoning, mapped the attack/defence landscape in a widely cited survey, and study when global feature effects can be trusted at all.
Interpretable time-to-event prediction, explanations for 3D segmentation models, radiomics for treatment-outcome prediction, and open clinical datasets — where a wrong model has a name attached to it.
Many models fit the data equally well and behave very differently. We use that set — rather than a single winner — to reason about stability, bias and the real uncertainty behind a prediction.
Detection obligations for generative models, resistance against manipulative AI, and the attribution of ideas in AI research. Published in Nature Human Behaviour, Ethics and Information Technology, and at ECAI.
MI².AI brings together researchers from the mathematics and computer science faculties of the University of Warsaw and Warsaw University of Technology. The group is deliberately mixed, statisticians, ML researchers, clinicians' collaborators, because the questions do not respect departmental boundaries.
A curated selection from 200+ publications, organised by theme rather than by count. The complete and continuously updated record lives on Google Scholar and DBLP.
Selection reflects methodological contribution and venue rather than citation count.
A method that only exists in a paper is a method almost no one will use. The DrWhy.AI ecosystem is built on one idea: a model should be interrogable through a single, consistent interface regardless of what framework trained it. The design has been adopted well beyond our group — DALEX alone has been downloaded millions of times and is taught in courses on several continents.
One explainer object, any model. Variable importance, partial dependence, break-down and SHAP attributions, residual diagnostics — the workhorse of the ecosystem. Published in JMLR.
Time-dependent explanations for machine learning survival models, including SurvSHAP(t). Published in Bioinformatics; the reference tool for interpretable time-to-event modelling.
Checks a classifier against a battery of group-fairness criteria at once, and offers pre-, in- and post-processing mitigation. Published in The R Journal.
Generates a serverless, shareable HTML dashboard from any explainer — the interface layer that turns explanation from a static plot into a conversation with the model.
An interactive arena for comparing several models on the same observations — built for the Rashomon-set way of thinking, where the interesting question is where equally good models disagree.
Reproducibility infrastructure: every model, plot and dataset gets an address you can cite and retrieve years later.
Earlier packages, still maintained and still downloaded.
My most-cited software contribution by a wide margin — a fixture of biomedical papers worldwide.
Correct handling of plausible values and replicate weights for PISA, PIAAC, TIMSS and PIRLS — the studies that shape education policy.
A family of goodness-of-fit tests that select their own complexity.
Mixture models that incorporate partial, uncertain label information.
PISA 2000–2012 and PIAAC packaged for R, so that anyone can check what the headlines actually say.
Half of these are textbooks and half are comics, and that is deliberate. The people whose lives are shaped by predictive models — clinicians, teachers, regulators, students — are not going to read a NeurIPS paper. All titles are free to read online.
A comic journey through AI transparency. Beta and Bit show why post-hoc explanations matter — and where they mislead.
Fifty-two pages, one story, and a complete reproducible ML workflow underneath it. Translated into several languages.
How to build plots that inform rather than impress — for audiences who did not ask for a chart in the first place.
Fractals and chaos theory, with worked examples in R, Python and Julia.
A comic about Zeno's paradoxes, infinite series, and the strange places where intuition fails.
With Tomasz Burzykowski. A step-by-step treatment of the main methods of explainable machine learning; published by Chapman & Hall/CRC and free online.
The Polish-language guide to R — basics, programming, statistics and visualisation. Now in its fourth edition.
Essays on the design and aesthetics of data visualisation.
An academic textbook on estimation and hypothesis testing, worked through in R.
Every course I run is built around the same conviction: students should leave able to break a model, not only to train one. Course materials — slides, code, projects — are public on GitHub, and the projects frequently turn into papers.
The flagship course. Students work through the modern XAI toolkit — attributions, counterfactuals, concept methods, sparse autoencoders — and then spend the second half of the semester attacking those methods on real models. Assessment is a written model audit, not an exam.
Running continuously since 2017. A reading and research seminar where thesis projects are developed in public, critiqued by peers, and — often enough — pushed to a workshop or conference submission.
| Course | Topic | Years |
|---|---|---|
| Data Visualisation Techniques | Grammar of graphics, perception, narrative visualisation | 2014–2019 |
| Warsztaty Badawcze / Case Studies | End-to-end applied projects with external partners | 2015–2018 |
| Advanced Programming and Data Analysis with R | Package development, performance, reproducibility | 2017–2018 |
| Projekt Zespołowy / Team Project | Team-built data products | 2017–2019 |
| Linear and Mixed Models | Fixed and random effects, applied inference | 2015–2017 |
| Statistics II: Exploration, Prediction and Verification | Statistical learning foundations | 2015–2017 |
| Programowanie i Wizualizacja z R | R programming and visualisation | 2016–2018 |
| R and Big Data | Out-of-memory and distributed data analysis in R | 2014–2016 |
| Data Products | Building tools other people can use | 2015–2016 |
I give roughly a dozen invited talks a year, to audiences that range from ML conferences to hospital boards to journalists. Three talks I am currently giving:
Why benchmarks are not enough, what the four pillars are, and what infrastructure the field would need to make findings about models accumulate.
Keynote · 45 min
A critical tour of what explanation methods actually deliver, how they can be attacked, and how to use them honestly in a high-stakes setting.
Invited talk / tutorial · 60–90 min
What goes wrong when a model meets a clinic — multiplicity, distribution shift, and the specific ways in which validation is usually too optimistic.
Domain audiences · 40 min
I am interested in collaborations on the verification and interpretability of foundation models, and in applications where the stakes make credibility non-negotiable — medicine above all. I supervise doctoral students through MI².AI and, through Solutions42.ai, work with organisations that need existing models validated rather than new ones promised.
If you are considering a PhD in interpretability, verification or steering of modern models, write to me with a short note on the problem you cannot stop thinking about. For talk invitations, please include the audience, date and format — there is no fee for public-interest and educational events.
Przemysław Biecek is a Full Professor at the University of Warsaw and Director of the Centre for Credible AI at Warsaw University of Technology, where he leads the MI².AI research group. His work centres on Model Science — the interpretability, controllability and verifiability of machine learning systems. He has authored more than 200 scientific publications, is listed among the top 2% most-cited scientists worldwide, and is a Fulbright IMPACT Award laureate. His research appears in Nature Machine Intelligence, Nature Human Behaviour, NeurIPS, ICML, ICLR, CVPR, ECCV and MICCAI.
Alongside academic work he founded the Smarter Poland Foundation, which promotes data literacy in schools and public debate, and in 2021 founded Solutions42.ai after fifteen years of building solutions for universities and companies including Samsung, IBM and Disney.
Academic lineage: Witold Klonecki → Julian Perkal → Hugo Steinhaus → David Hilbert. Second (biological) advisor: Stanisław Cebrat.