TRAIF 2026—Prof. Matthias Uhl

TRAIF 2026—Prof. Matthias Uhl

by Matthias Uhl · View the channel on PSI

Humanities, Social Sciences & Thought Transport & Mobility TTechnology, Engineering, Agriculture, Industrial processesTRTransport technology and trades
🎙 Matthias Uhl 👥 395 📅 October 20, 2026 ⏱ 20 min 👁 0 📄 original study 🧭 2026-10-20
Available in: English (current) Français

Keywords

generative AIbiasparticipatoryempirical studygender

Summary

The talk by Professor Matthias Uhl at The Responsible AI Forum 2026 presents a participatory approach to generative AI, focusing on how to identify shared values and operationalize bias. He argues that biases are systematic deviations from normative standards, which are often implicit. He discusses different ethical standards such as proportional equality, uniform equality, and substantive equality as possible benchmarks for AI representation. The main part of the talk describes an empirical study conducted with co-authors Florian Richa and Sebastian Kruger, which investigated lay people’s preferences for gender representation in various social categories like CEO, firefighter, nurse, kindergarten teacher, terrorist, and thief. Participants were recruited via a crowdsourcing platform and asked to choose the most appropriate set of images from 11 variations, with only the share of women varying. The study found a strong correlation between participants’ empirical beliefs about actual gender proportions and their normative choices, regardless of the participant’s gender, the moral connotation of the category, or whether the category was male- or female-dominated. Importantly, marginalized groups did not seek idealized representations. The speaker concludes that human-centered design should at least be aware of user preferences, even if they are overridden, and that this study is a first step toward incorporating lay intuitions into AI bias discussions. Limitations include focusing only on gender and using a US sample.

220 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information lies in it presenting a concrete empirical method to elicit lay people’s normative standards for AI representations, which is a novel and often overlooked perspective. The argumentation is solid: the speaker carefully defines bias, illustrates the difficulty of choosing a normative standard with examples like the Linda problem and ethical dilemmas, and then shows how the study operationalizes the concept. The correlation between empirical beliefs and normative choices is highlighted, with robustness checks across different categories, genders, and moral connotations. The claim that marginalized groups do not want idealized representations is particularly striking and supports the conclusion that users generally prefer realism over idealization. The limitation on gender and cultural scope is acknowledged, showing scientific humility.

Scientific Rigor, Source Quality, Title Accuracy

The speaker references the Linda problem by Kahneman and the omission effect in ethics, but does not provide specific citations in the talk. The description includes links to the Responsible AI Forum, the Institute for Ethics in AI, the alignAI project, and other resources, which provide context and potential sources. The title is generic (just the forum and speaker name) but the content is clearly aligned with the forum’s theme; the adequacy is neutral, neither misleading nor enhancing. The scientific rigor is apparent in the study design, random assignment, and regression analysis, though the talk lacks detailed methodological appendix. Overall, the sources are not explicitly cited, but the surrounding materials support the credibility.

249 words

Title / Content Match

The title 'TRAIF 2026—Prof. Matthias Uhl' is generic, but the content matches the context of The Responsible AI Forum 2026. The talk presents a participatory approach to generative AI, which aligns with the forum's theme.

Quality & Reliability

8/10

The talk presents a well-structured empirical study with clear methodology, robust data analysis, and explicit acknowledgment of limitations. The speaker is a professor from a reputable institution (TUM), and the study is part of a larger research program. The presentation is balanced and avoids overclaiming.

External References

Contribution & Novelties

The main novelty is the empirical elicitation of lay people’s normative standards for AI-generated representations, showing that users prefer realism over idealistic or corrective depictions. This challenges the common assumption in AI fairness debates that certain ethical standards (like substantive equality) should be imposed. The participatory approach is original in that it brings stakeholders’ concerns directly into the design process.

Pour aller plus loin :

  • alignAI project — EU-funded project on value alignment in AI, directly relevant to the talk’s theme.
  • Responsible AI Forum — The venue where this talk was presented, offering further discussions on responsible AI.
  • Institute for Ethics in AI, TUM — The speaker’s institution, providing background on ongoing research in AI ethics.

116 words

Radar Profile

The radar profile shows balanced scores across dimensions, with slightly higher marks on information quantity and reliability, reflecting the solid empirical grounding and clear presentation. The moderate technical level indicates that the talk is accessible but still requires some familiarity with empirical research.

Reliability 8/10