In today’s world which is marked by a confluence of geopolitical turbulence and stunning advances in digital technologies spearheaded by AI , digital humanism can provide a guide both for future technological research and development as well as the choices we are called to make regarding the applications of technology as individuals and as members of society. What kind of digital world can we envision in which democracy, the rule of law and basic human rights, including the pursuit of happiness, are supported by technology rather than being threatened by it? How can we work backwards from such a vision and identify the steps to be taken by academia, industry and governments required to enable it as well as identify the traps to be avoided along the way? Is the slide towards “electoral authoritarianism” irreversible or can your generation re-invent democracy and create a present that the future will be…nostalgic of. Utopia may be unattainable, but dystopia is not the only other choice!

slides

Reinforcement learning from human feedback has emerged as one of the primary mechanisms for aligning large language models. In this framework, (human) workers report their preferences over possible responses to a query, and their reports are used to fine-tune the model. This process collects feedback from multiple raters, and therefore implicitly aggregates the raters' preferences. In this tutorial, we will discuss voting methods that underpin popular RLHF approaches, and investigate whether tools of social choice can be used to design better fine-tuning procedures.

Location: El9 Hlawka, next building

Excursion to the Austrian Parliament.

Groups of 4-5 students select one of several use cases.
With a special ad-hoc presentation by Ittai Bar Siman Tov at the Austrian Parliament.

parliament slides

EC slides

video

Planned reception at the City Hall of Vienna.

The Digital Humanism manifesto, with its diagnosis from 2019, is still largely valid today. This is surprising given that it was prepared before the advent of recent generative AI systems that have significantly changed how people work, who they ask for advice, and also how they see themselves. This talk will thus revisit Digital Humanism from the perspective of AI. It will investigate what is sometimes construed as a gap between fragile humans and powerful machines from the perspective of ethics. Diving into various ethical theories we investigate the relation of ethics and digital humanism and its use for constructively approaching a digital future worth living.

slides

The collapse of artists’ revenues and the rise of AI are certainly not coincidental: they represent two gears within the same instrument. Streaming’s economic model rewards infinite supply at near-zero unit cost. While human-created music is finite and expensive, AI music is infinite and cheap. For a platform whose margins improve as payout obligations shrink, the logical endgame is obvious: keep the streams, remove the artists. My lecture is grounded on my experience as a professional musician, union activist, and a member of European Commission expert groups focused on Music in the Digital Age. I will attempt to peel part of this very large onion to answer such questions as:

1. How have streaming platforms changed the economics of the music business?
2. How can we protect the rights of creators when AI is trained to compete with human creation?
3. How can we build sustainable and equitable compensation models in the streaming market?
4. What role can/should legislation play to secure ethical AI to promote a fairness-oriented digital market without posing a threat to AI-related innovations?
5. What can we expect from recent EU regulatory mechanisms with a focus on the EU AI Act?
6. How can musicians/other creatives rally to gain a fair market share of streaming revenues?
7. Should unions retreat from an animosity towards AI “AI will become the worker vs. AI supports the work”? If so, what strategies for future success could be promoted?

Time permitting, we will take a look at several recent, groundbreaking cases. Interestingly, most ‘wins’ have been booked in protecting creatives from piracy lags behind.

slides

Political communication is how political actors (such as political parties and governments), media/journalists, and citizens create and exchange messages about politics. AI is rapidly transforming this process through data-driven, emotionally targeted communication, offering new opportunities for engagement while raising risks of manipulation, polarization, and fragmented public discourse. This lecture connects empirical insights from political and communication research with the principles of Digital Humanism, asking how democratic politics can effectively communicate in the age of AI.

Floris Geerts - combined ESSAI Keynote @ TU Vienna: The Expressive Power of Graph Learning
A key component of graph and relational learning methods is the computation of vector representations of input graphs or relations. In this talk, we view this computation through the lens of queries: mappings from discrete relational structures into spaces of real-valued vectors. This perspective provides a unifying language for revisiting the expressive power of graph learning methods and helps bridge ideas from database theory and machine learning.
Location EI 7

A special Ukraine event on short notice:

With short presentations:
Dietmar Pichler, researcher from https://inved.eu, on disinformation, slides
Boryslav Ostretsov, a Ukraine drone developer, no slides
Michael Trusov, PhD student at the Department of Cybersecurity of Information Systems, Networks, and Technologies, slides

Denys Sadilo, Dept. Computer Science, "Finding the hidden connections nobody wrote down"
Topic: Finding the hidden connections nobody wrote down
Investigations run on messy, multilingual, often private text, and the connection that cracks a case is rarely written down. The research detects these hidden links between entities in large dynamic graphs and names what kind of link each is. A two-stage pipeline does this: stage one builds a graph from raw text cheaply and offline, so it can run air-gapped; stage two reasons over the graph with an inductive graph neural network and knowledge-graph embeddings, discovering relation types from the data rather than a fixed taxonomy. With no ground truth for hidden links, evaluation stacks independent tests on public data, and a human stays accountable throughout.
slides

To discuss disinformation, drones, democracy, war, ...
Location FAV01

One of the central challenges of AI and its impact on society is the assessment or audits of developed AI models for their alignment with ethical principles and values. Often, ethical concerns with AI models are uncovered only after their launch, making fixes particularly ad hoc. Fortunately, a legion of auditing methods is being developed. In this tutorial, I discuss the AI alignment problem and review various types of auditing methods, discussing the central obstacles to the practical realization of AI Ethics Audits. Guided by the question “how can we empower (small) developers of AI and lower the threshold for embedding ethical reflection throughout an AI’s life-cycle?,” I will briefly present a workshop method intended to be used by AI developers for ethical self-assessment. The tutorial is based on joint work with Tobias Christoph and Katta Spiel.

My talk will introduce the concept of democratic trustworthiness of AI, shifting the focus from thinking about AI users to protecting citizens; and from output expectations to protected interests within AI systems. Most contemporary debates on Trustworthy AI (TAI) emphasise technical and output-oriented criteria such as accuracy, robustness, safety, and lawfulness. While necessary and valid, this talk argues that these markers are insufficient from a strictly democratic perspective on citizen interests. Drawing on Roger Hardin’s concept of trustworthiness as “encapsulated interest,” I reframe trustworthiness in AI as a relational and political question: whose interests are structurally taken into account by AI systems? Can everyone equally afford to trust AI, independent of its reliability? The central claim is that dominant TAI frameworks tend to adopt a perspective in which the trustworthiness of AI is more about accurate information and less about power and representation. While I argue that AI should only be considered trustworthy if it demonstrably encapsulates the interests of those affected by it, especially those of vulnerable groups.

Slides

Beyond Personalization: Recommender Systems and Digital Humanism Recommender systems influence what billions of people see, buy, hear, and believe. This presentation moves beyond personalization to examine recommender systems from a Digital Humanism perspective, asking what these systems optimize, who benefits, and which values are embedded in their design. It explains how modern recommendation algorithms learn from user behavior and curate digital experiences at scale. Rather than treating recommendation as a purely technical problem of prediction, the talk highlights its societal implications and explores how recommender systems can better balance predictive accuracy and business objectives with human agency, fairness, transparency, and societal well-being.

Slides

The ongoing process of digitalization makes the realization of a fully direct democracy—one in which all political decisions are made directly by the people—appear increasingly feasible, even in large and populous states. This might suggest that majority decisions, determined through general, direct, free, equal, and secret elections and referenda, constitute the sole defining feature of democratic decision-making. However, democracy is more than majority rule. This talk explores what democracy presupposes beyond voting procedures. Democracy entails human rights, the separation of powers, and the role of expert consultation, which must not be undermined by digital participation tools. It also requires cooperation: it cannot function if a minority defeated in a vote obstructs the implementation of the decision made by the majority or if citizens act solely as self-interest-maximizing agents. Moreover, in a democracy, political debates cannot be reduced to the assertion of the interests of various lobby and interest groups, but also involve the articulation of political positions put forward with a claim to truth (while remaining open to the possibility of error). Finally, from a philosophical perspective, the freedom and equality of citizens constitute normative preconditions of democracy.

Slides

The lecture looks at the ongoing convergence of artificial intelligence and modern warfare. On battlefields such as those in Ukraine, Gaza or Iran, the use of quasi autonomous systems and drones is already a reality. A powerful military-digital complex is taking shape, dominated largely by the U.S. and China. This automation carries enormous risks of escalation, opaque decision-making processes, and ethical dual-use issues.

Slides

Machine learning (ML), particularly deep learning, is being used everywhere. However, not always is used well, ethically and scientifically. In this talk we first do a deep dive in the limitations of supervised ML and data, its key component. We cover small data, datification, bias, predictive optimization issues, evaluating success instead of harm, and pseudoscience, among other problems. The second part is about our own limitations using ML, including different types of human incompetence: cognitive biases, unethical applications, no administrative competence, misinformation, and the impact on mental health. In the final part we discuss regulation on the use of AI and responsible AI principles, that can mitigate the problems outlined above.

Slides

A common critique of AI systems is their reduced transparency (“black boxes"), which often raises a critical challenge for ensuring that they function in a principled and fair way while complying with legal regulations. To address this issue we work on representing the internal structure of AI systems in detail. Additionally, we develop methods and tools for Auditing AI systems for several criteria, including some of those listed as requirements by the EU AI Act. An important question remains how input from social science (e.g., philosophy) can be infused in our work as part of the broader area of Digital Humanism

AI systems increasingly affect society, yet questions of democratic control remain largely unresolved. At present, many consequential choices in the development of AI systems — from data collection to deployment — are made by a small number of decision-makers, with little public input or oversight. Current debates on democratic AI governance tend to focus on institutional and regulatory frameworks at the macro level. In this talk, I complement this perspective with an approach grounded in social choice theory, the mathematical study of how conflicting preferences can be aggregated into collective decisions. I argue that opportunities for meaningful collective input arise at many stages of the machine learning development pipeline, yet are rarely recognised as such. Social choice theory provides both a modelling language to make these decision points explicit and principled criteria - drawing on notions such as fair representation, participation, and efficiency - to evaluate mechanisms for incorporating collective input. The talk aims to provide an accessible introduction to social choice theory in the context of AI governance, requiring no prior background in either area.

Project Presentations