Timo Freiesleben

My work in the philosophy of machine learning examines the interplay between machine learning and the philosophy of science. On the one hand, I aim to provide an epistemological grounding for machine learning practices such as benchmarking, robustness testing, and interpretability. On the other hand, I study the philosophical implications of applying machine learning tools in scientific inquiry, particularly with respect to scientific goals such as prediction and explanation.

• The Benchmarking Epistemology: Validity Theory for Evaluating Machine Learning Models.
  We investigate benchmarking as an epistemic practice in machine learning and develop a validity theory for evaluating machine learning models. We show that benchmark results can support a wide range of scientific inferences, but only if the corresponding validity conditions are met.
  Philosophy of Science.
  Joint work with Sebastian Zezulka.
• Scientific Inference with Interpretable Machine Learning.
  We argue that interpretable machine learning techniques can take over the inferential role that parameters play in traditional statistical models. We provide a methodological recipe for new interpretation techniques that support well-defined scientific inferences.
  Minds and Machines.
  Joint work with Gunnar König, Christoph Molnar, and Álvaro Tejero-Cantero.
• Supervised Machine Learning for Science.
  In this book, we analyze how machine learning is shifting scientific inquiry from an explanation-focused to a prediction-focused enterprise. However, we show that recent work on interpretability, robustness, and causality allows us to gain insights from machine learning models that go beyond mere predictions.
  Self-published.
  Joint work with Christoph Molnar.
• A Theory of Robustness in Machine Learning.
  We develop a unified account of robustness in machine learning. We conceptualize robustness as the relative stability of a target (e.g. predictive performance) under specific interventions on a modifier (e.g. natural distribution shifts).
  Synthese.
  Joint work with Thomas Grote.
• Establishing Construct Validity in LLM Capability Benchmarks Requires Nomological Networks.
  I argue that nomological networks provide the most philosophically adequate framework currently available for conceptualizing LLM capabilities.
  Philosophy of Science Association (PSA) Conference.
• Artificial Neural Nets and the Representation of Human Concepts.
  I argue against the view that individual units in trained neural networks represent unique human concepts.
  Book chapter in Synthese Library.
• Foundation Models in Healthcare Require Rethinking Reliability.
  We argue that standard out-of-sample testing is insufficient for assessing the reliability of foundation models in healthcare contexts.
  Nature Machine Intelligence.
  Joint work with Thomas Grote and Philipp Berens.
• The Intriguing Relation Between Counterfactual Explanations and Adversarial Examples.
  I argue that the main difference between counterfactual explanations and adversarial examples lies in their relationship to the true label.
  Minds and Machines.

My work in AI ethics focuses on the challenges faced by individuals who receive unfavorable outcomes from automated decision-making systems. In particular, I study the normative implications of deploying machine learning systems in social contexts.

• Explainable AI Isn't Enough! Rethinking Algorithmic Contestability. 🏆 Best Paper Award
  We define what it means for an algorithmic decision to be contestable and show that common XAI explanations alone are insufficient to establish contestability. We demonstrate that other forms of evidence, including predictive multiplicity, offer stronger support and provide greater value for enabling contestability.
  AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026).
  Joint work with Kristof Meding and Gunnar König.
• Performative Validity of Recourse Explanations.
  We analyze the performative effects that recourse explanations have on their own validity. We find that recourse recommendations that focus on causes of the target largely remain valid under their own performative effects.
  NeurIPS Conference.
  Joint work with Gunnar König and others.
• Improvement-Focused Causal Recourse.
  We argue that many counterfactual explanations enable users to game systems rather than improve their qualifications. Consequently, we show how causal knowledge can support meaningful recourse.
  AAAI Conference on Artificial Intelligence.
  Joint work with Gunnar König and Moritz Grosse-Wentrup.

My work on explainable AI examines the conceptual foundations of explanation techniques. I provide critiques of common misconceptions as well as philosophically motivated technical contributions.

• Dear XAI Community, We Need to Talk!
  We analyze widespread misconceptions in XAI research. The most crucial one concerns the lack of clarity about the purposes of explanation methods.
  World Conference on Explainable AI.
  Joint work with Gunnar König.
• Relating Partial Dependence and Permutation Feature Importance.
  We connect interpretability methods with the data-generating process. We define confidence intervals for partial dependence and permutation feature importance.
  World Conference on Explainable AI.
  Joint work with Christoph Molnar, Gunnar König, and others.
• General Pitfalls of Model-Agnostic Interpretation Methods.
  We discuss common mistakes in model-agnostic interpretation, particularly those related to feature dependence.
  Book chapter in Lecture Notes in Computer Science.
  Joint work with Christoph Molnar and others.
• CountARFactuals.
  We formulate and implement an efficient algorithm for generating plausible model-agnostic counterfactual explanations.
  World Conference on Explainable AI.
  Joint work with Susanne Dandl, Kristin Blesch, Gunnar König, and others.