Verantwortung der Informatik – Accountability in AI (AIAI 2026)

What is the environmental cost of training and operating Large Language Models (LLMs)? Analyze IT energy consumption, cooling requirements for data centers, and the carbon footprint of inference vs. training. The core question: How does the computational demand of AI intersect with global sustainability goals and what frameworks exist to measure and audit this impact?

Literature:

• Luccioni et al., Estimating the Carbon Footprint of BLOOM

Generative models are trained on massive datasets scraped from the Internet, often containing copyrighted material. This topic examines the legal and ethical dimensions of data scraping, focusing on "fair use" doctrines, the blurring lines of authorship, and the ethics of academic and creative integrity when an LLM regurgitates exact or lightly modified source material without attribution.

Literature:

• Samuelson, P., Generative AI meets copyright law

LLMs are prone to "hallucinations" — generating confident but entirely false or misleading information. You will examine the architectural reasons why this happens (e.g., probabilistic token generation vs. factual retrieval), the ethical implications in high-stakes fields like law and medicine, and how this accelerates global misinformation and truth decay.

Literature:

• Ji et al., Survey of Hallucination in Natural Language Generation

This topic examines how the rise of generative AI impacts professional artists, musicians, and filmmakers. As systems like diffusion models and AI video generators rival human creation, intricate challenges emerge. You will explore the economic displacement of creatives, the diminishing uniqueness of art, and potential solutions (like royalty systems or licensing) to ensure a fair coexistence.

Literature:

• AI Art and its Impact on Artists
• Art and the science of generative AI

Federated Learning promises inherent data privacy by distributed training with local models and only updating a global model's weights, theoretically keeping the training data local. But is it truly private? You will explore the vulnerabilities of this system, specifically focusing on model inversion and data reconstruction attacks that can extract sensitive private data purely from shared model weights.

Literature:

• Kairouz et al., Advances and Open Problems in Federated Learning

In this topic, you will delve into the concept of "Panic as a Service" (PaaS), a term that has emerged in the age of technology-driven anxiety. PaaS refers to the increasing prevalence of platforms, services, and technologies that exploit, amplify, or profit from societal fears, uncertainties, and doubts. You are encouraged to explore the dynamics of this phenomenon, examining the ways in which technology, media, and other entities leverage panic for various purposes. AI plays a significant role in PaaS by amplifying, enabling, and sometimes even perpetuating the mechanisms behind this phenomenon. You'll analyze the different aspects oh how AI affects PaaS. Furthermore you can put a focus on AI as the central subject of panic. As we witness the rapid advancements in AI and its increasing integration into our daily lives, questions and concerns surrounding the impact of AI on society's collective anxiety come to the forefront.

Literature:

• The uselessness of AI ethics
• In defense of ethical guidelines
• A Taxonomy of AI Panic Facilitators

When an autonomous vehicle crashes or causes harm, who is legally and ethically accountable? The manufacturer, the software engineer, the "driver," or the AI itself? You will investigate the current legal and regulatory frameworks surrounding autonomous systems, the dilemma of edge-case decision making, and how liability shifts as we move from Level 2 to Level 5 autonomy.

Literature:

• Legal and regulatory implications of artificial intelligence (AI)
• Global Guide to Autonomous Vehicles

The European Union has pioneered comprehensive AI regulation with the AI Act, which categorizes AI systems by risk (unacceptable, high, limited, minimal). You will analyze the definitions and parameters of this regulation, evaluate its strengths and limitations, and discuss the "Brussels Effect" — how this European legislation is forcing global tech giants to change their worldwide practices.

Literature:

• European Commission, The EU Artificial Intelligence Act

Biases in data can lead to different behaviour resulting from various ethnic and/or socio-demographic subpopulations. Especially for applications in medicine and healthcare, this can have critical effects on underrepresented groups. Which metrics or methods can be used to detect biases in data collection, pre-processing and/or model performance? What can be done to remove biases? What consequences can biases have?

Literature:

• Fairness in Deep Learning: A Computational Perspective
• Ethical Machine Learning in Health Care

To ensure quality for a product, so-called audits are carried out to check if the product meets certain standards and regulations. For AI for health, the requirements for safety and efficacy are particularly strict and difficult to evaluate. These evaluations include the analysis of model performance, privacy, robustness, fairness, bias, and interpretability among other things. Which processes could the evaluation part of an audit include? Which regulations or guidelines are available? How does AI for health differ from other applications regarding the evaluations?

Literature: * ML4H Auditing: From Paper to Practice * Ethical Machine Learning in Health Care * Ethics and governance of artificial intelligence for health * Against explainability requirements for ethical artificial intelligence in health care

Most popular (social media) websites and apps use recommender systems to individually filter content and provide users with suggestions of movies to watch, news articles to read, music to listen to, etc. Suggestions are based on previous user interactions and optimized to match the users' interests, maximizing user engagement. By providing a constant feed of interesting content recommender systems can lead to the excessive use (addiction) of Internet applications. On the other hand, recommender systems also create echo chambers, reinforcing the opinions of users, by only showing content that agrees with a user's preexisting opinion. These virtual echo chambers or "bubbles" make critical discourse much harder, because a ground truth, that both parties can agree on no longer exists. How can recommender systems be built to be less addicting, while still providing relevant content? How to pop the "bubbles"?

Literature:

• Excessive use of online video streaming services: Impact of recommender system use, psychological factors, and motives
• Stray et al., Building Recommender Systems with Human Values in Mind

Big, state-of-the-art neural networks need big datasets for training. How does this fit with the principles of the European data protection laws (GDPR)? How can we ensure data protection and progress in AI at the same time? What restrictions are already imposed on AI research and products by the GDPR? How can requirements such as the principle of data economy (German: Grundsatz der Datensparsamkeit) and the right to be forgotten be implemented? How can we ensure the privacy of uninvolved third parties such as relatives, which share parts of their genome?

Literature:

• Artificial Intelligence Ethics and Challenges in Healthcare Applications: A Comprehensive Review in the Context of the European GDPR Mandate
• GDPR: an impediment to research?
• Mourby et al., Are 'pseudonymised' data always personal data? Implications of the GDPR
• The Effect of the General Data Protection Regulation on Medical Research

Nowadays, machine learning (ML) techniques are gaining more and more popularity because of their performance. With them, we are able to model complex phenomena receiving high accuracy of results. One of the limitations of ML models is the fact the mean value of the output is estimated. As a result, we do not know how our model is confident about the result. Consequently, we – end users – are not sure if the model is trustworthy. To overcome that problem the uncertainty of the prediction can be estimated. Uncertainty tells us with the given probability how far from the mean value the real value is expected to be. Uncertainty estimation allows users to decide if they can trust the model and when there is a need for an additional human decision.

Literature:

• Gawlikowski et al., A Survey of Uncertainty in Deep Neural Networks
• Importance of being uncertain

Artificial Intelligence (AI) is attracting the attention of more and more specialists, not only computer scientists and statisticians but also medical personnel, engineers, and economists. AI enables the modeling of complex phenomena receiving high accuracy of results. One of the limitations is the fact that AI models are "black boxes". It means that it is difficult to explain the relationship between the input features and the output. Consequently, we – end users – are not sure if the model is trustworthy. To overcome this problem different algorithms and approaches have been developed to explain predictions. It also allows estimating which features have the highest impact on the result. As a result, the users can assess the outcome with state-of-the-art knowledge and discover new patterns in an analyzed phenomenon.

Literature:

• Interpretable machine learning. A Guide for Making Black Box Models Explainable
• Explainable AI: A Review of Machine Learning Interpretability Methods
• Explainable Artificial Intelligence: a Systematic Review
• Against explainability requirements for ethical artificial intelligence in health care
• Rudin, C., Stop explaining black box ML models for high stakes decisions and use interpretable models instead

Synthetic Media is a catch-all term for the artificial production, manipulation, and modification of data and media by automated means. Synthetic media as a field has grown rapidly since the creation of generative adversarial networks, primarily through the rise of Deepfakes as well as music synthesis, text generation, human image synthesis, speech synthesis, and more. While this technology may bear positive effects, like cutting cost in production, it's rather obvious that its possibilities are rather concerning. From misrepresenting well-known politicians, and thus misinforming the public, to social hacking. You will analyze the malicious use of this technology, focusing on social hacking, the misrepresentation of politicians, and the subsequent threat to democratic elections and public trust.

Literature:

• Synthesizing Obama: learning lip sync from audio
• The Creation and Detection of Deepfakes: A Survey
• Everybody's Talkin': Let Me Talk as You Want
• Vaccari & Chadwick, Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News

AI is transitioning from passive chatbots that answer questions to "Agents" that take autonomous actions on the Internet (e.g., executing trades, booking flights, or scraping databases). Students will examine the accountability gap here: If an AI agent executes a financially ruinous action or commits cyber-fraud on a user's behalf, who is liable? The developer, the user, or the API provider?

Literature:

• Chan et al., Harms from Increasingly Agentic Algorithmic Systems

The AI community is fractured over how models should be released. Proponents of "closed" AI argue that open-sourcing powerful foundation models is too dangerous (enabling bioterrorism or mass spam). Proponents of "open" AI argue that closed APIs consolidate power among tech oligarchies and prevent independent safety auditing. You will critically analyze both sides of this safety vs. democratization debate.

Literature:

• Widder et al., Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI

Expanding on the concept of "Dual-Use," this topic directly addresses the military application of AI. You will investigate the deployment of AI target-generation systems and Lethal Autonomous Weapons Systems (LAWS). The core debate: Can meaningful human control ever be maintained in warfare operating at algorithm speeds and is a global ban on weaponized AI enforceable?

Literature:

• Amoroso, D., Tamburrini, G. Autonomous Weapons Systems and Meaningful Human Control: Ethical and Legal Issues

The "magic" of AI is often built on the backs of precarious human labor. You will investigate the global supply chain of AI, specifically the low-wage click-workers in the Global South who annotate training data, write RLHF (Reinforcement Learning from Human Feedback) prompts, and traumatize themselves filtering toxic, illegal content. How can AI companies be held accountable for fair labor practices?

Literature:

• Tubaro et al., The trainer, the verifier, the imitator: Three ways in which human platform workers support artificial intelligence

AI has revolutionized computer vision, enabling real-time facial recognition, emotion detection, and predictive policing. You will explore how authoritarian regimes and democratic police forces alike deploy biometric surveillance. The presentation should cover the severe civil liberty implications, algorithmic racism inherent in these systems, and why the EU explicitly banned certain biometric categorizations.

Literature:

• Punia et al., Ethical Considerations and Legal Frameworks for Biometric Surveillance Systems: The Intersection of AI, Soft Biometrics, and Human Surveillance
• BSI, Biometrie als KI-Anwendungsfeld

AI models are highly proficient at writing code that includes malicious payloads. Students will explore the dual-use dilemma in cybersecurity: AI can be used defensively to patch vulnerabilities and detect network intrusions, but it can equally be used offensively to automate hacking, mass-generate convincing phishing campaigns, and write mutating malware. How can we govern the access to cybersecurity-trained models?

Literature:

• Ayzenshteyn et al., The Best Defense is a Good Offense: Countering LLM-Powered Cyberattacks

In response to AI companies scraping the web without consent, artists, and developers have turned to "data poisoning" (e.g., tools like Nightshade or Fawkes). These tools alter the pixels of an image in ways invisible to humans but catastrophic to AI training data, effectively ruining the model's outputs. You will discuss the ethics, legality, and technical mechanisms of using adversarial attacks as grassroots accountability.

Literature:

• Shan et al., Nightshade: Prompt-Specific Poisoning Attacks on Text-to-Image Generative Models

As generative models produce indistinguishable synthetic text, audio, and video, technical accountability relies heavily on proving the origin of digital content. You will explore cryptographic provenance standards (like the C2PA Content Credentials) and in-pixel watermarking techniques (like Google's SynthID). The presentation should assess the technical robustness of these methods against tampering, the "liar's dividend" (where real evidence is dismissed as AI), and whether watermarking should be a mandatory legal requirement for foundation models.

Literature:

• Gowal et al., SynthID-Image: Image watermarking at Internet scale

The EU Cyber Resilience Act (CRA) mandates strict "security by design", rapid vulnerability patching, and the creation of Software Bills of Materials (SBOMs) for all digital products. Students will investigate the massive friction of applying these rules to AI models. How do you create an "ingredients list" (AI-Bills of Materials) for a black-box model trained on billions of opaque data points? If a foundational model is found to have a severe vulnerability (like susceptibility to prompt injection or a data-poisoning backdoor), how do developers practically and legally "patch" a neural network within the CRA's strict deadlines? The presentation should also explore whether the CRA's liability framework threatens the open-source AI ecosystem.

Literature:

• European Parliament (2025). Interplay between the AI Act and the EU digital legislative framework (Section on CRA).

• What is the topic?
• How is it defined?
• Are there multiple, different definitions?
• Why is it important?

• Present a method/paper/tool which addresses the problem
• Check topic description/literature section for some suggestions
• Explain the main idea
• Highlight benefits and potential shortcomings

• Provide 2-3 points/questions to start the interactive discussion