Representing Humans

Title: Representing Humans
Research Area: Responsible AI

Human identity is unique and dynamic. Yet digital representations of individuals used as inputs for computational tasks, such as algorithmic loan decisions or college admissions, often fail to account for this individuality. Instead, a single representation scheme is typically assumed to capture all individuals equally well. This project challenges this conventional approach by developing strategies for grounding human representations in individual realities and methods for validating their digital representations in this regard.

Aims

There are two main goals of this project: One goal is to quantify representation fidelity, i.e. how faithfully an input representation reflects an individual person’s unique reality. The second goal strives to develop feature selection strategies for representations of humans that take into account their unique and dynamic reality.

Problems

The central question this project examines is “What are appropriate strategies to deal with the uniqueness of human reality in algorithmic contexts?” If algorithmic decisions about a human individual are based on an unfaithful representation, decision outcomes for this individual are unjustified. In the worst case, these outcomes are incorrect and have a negative impact on this person’s life. For example, an incorrect college application rejection can result in lesser education level, or a falsely accepted credit approval can result in personal default of the applicant.

Technology

On the one hand, we use data-centric approaches to develop datasets, run experiments, and design auditing methods for algorithmic decision making systems. We use methods from machine learning and information retrieval alongside with human data annotation to develop automatic and manual auditing methods that quantify human representation fidelity on our test data. Our test dataset is synthesized with using large language models, such as and GPT, LLaMA, and others.

Outlook

This project investigates the grounded creation and validation of digital representations of humans in algorithmic decision making. These methods increase transparency of algorithmic decision making by informing affected individuals as well as deploying institutions of human representation fidelity. This yields not only an increased agency for individuals to contest algorithmic decisions but also allows for inter-system comparisons within a specific decision context (i.e. comparison of multiple credit scoring algorithms).

Publications

• Theresa Elstner and Martin Potthast. Representation Fidelity:Auditing Algorithmic Decisions About Humans Using Self-Descriptions. CoRR, 2603.05136, March 2026.
• Theresa Elstner, Bärbel Hanle, Frank Loebe, Maik Fröbe, Nikolay Kolyada, Janis Mohr, Jörg Frochte, Sven Hofmann, Benno Stein, and Martin Potthast. Classification of Shared Tasks Used in Teaching. In Proceedings of the 2024 on Innovation and Technology in Computer Science Education V. 1, pages 164-170, July 2024. ACM.
• Theresa Elstner, Frank Loebe, Yamen Ajjour, Christopher Akiki, Alexander Bondarenko, Maik Fröbe, Lukas Gienapp, Nikolay Kolyada, Janis Mohr, Stephan Sandfuchs, Matti Wiegmann, Jörg Frochte, Nicola Ferro, Sven Hofmann, Benno Stein, Matthias Hagen, and Martin Potthast. Shared Tasks as Tutorials: A Methodical Approach. In 13th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI 23), pages 15807–15815, June 2023. EAAI.

Team

Lead

• Prof. Martin Potthast, Universität Kassel, hessian.AI, ScaDS.AI Dresden/Leipzig

Team Members with Affiliation

• Theresa Elstner, Universität Kassel und hessian.AI