Research on scientific practice, human competence, and responsibility

Artificial intelligence makes some old philosophical questions newly practical. What is a system? What counts as evidence that it can do something? When does flexible behavior amount to agency? And where should responsibility fall when consequential actions are distributed across models, people, institutions and infrastructure?

I have only recently started focusing on AI, driven by curiosity and surprise at the recent developments in the field of machine learning. My interest grows directly from my research on scientific method, measurement, and the construction of systems. I do not begin by asking whether machines think in exactly the way human beings do, but I start with questions that can be tied more closely to scientific and engineering practice (i.e. how AI systems are built, tested and calibrated; how their capacities are made visible through benchmarks; how control emerges in systems without a central executive; and how the boundaries of an artificial agent are drawn.) Across these projects, I am interested in understanding the process of learning and evolution which these systems undergo. As a philosopher, my attention gets captured by the transformation of technical success into an epistemic or moral claim, specifically about what additional work is needed to justify that transition.

Measuring AI

AI evaluation is often presented as if performance were simply there to be read off. But benchmark scores, confidence estimates and safety evaluations are measurements. They depend on choices about what is being measured, how a test environment is constructed, which errors matter, and how far a result can travel beyond the conditions in which it was produced. A model can receive an impressive score even when a benchmark is narrow, saturated, contaminated or insensitive to the failures that matter in use.

I am developing a project on the metrology of AI: the study of calibration, uncertainty and evaluation as problems of measurement. What does it mean for a model’s confidence to be calibrated? Does a confidence score represent a frequency, a degree of rational belief, or merely a stable relation between outputs and a test set? What is a benchmark score a measurement of – performance on a task, a more general capability, or conformity to the evaluator’s chosen construct? These questions become especially urgent under distribution shift, when a measure that appeared stable in one setting may cease to support the same inferences elsewhere.

This work brings the philosophy of measurement into conversation with machine learning. It treats evaluation not as the production of a leaderboard, but as the design and continued testing of an instrument. The point is not to deny that AI capacities can be measured. It is to ask what has to be built, maintained and checked before a number can bear the evidential weight placed upon it.

Interpretability and dependence

AI systems can produce answers, explanations, and fluent statements of reasons. But an answer is an output, whereas answerability is a relationship. To be answerable is not simply to reveal how an output was generated. It is to stand in a relation in which someone has the standing to demand an account, commitments persist beyond a single exchange, failures can alter one’s standing, and acknowledgment or repair is possible.

This distinction changes how I approach the familiar responsibility gap. The gap is usually traced to features of an AI system such as autonomy, opacity or unpredictability. I ask a prior question: where do the boundaries of the candidate agent fall? A conversational interface encourages us to see a self-contained entity speaking in the first person. Yet the model is the visible surface of a much larger sociotechnical system made up of training data, annotators, engineers, firms, servers, energy, capital and institutions. If we draw the boundary at the interface, we may create the very gap we then try to solve.

Thomas Hobbes once asked his readers to imagine human beings springing from the earth like mushrooms, fully mature and without prior attachments. Modern mycology gives the image an unexpected turn: the visible mushroom is not the self-contained organism, but the fruiting body of a larger and mostly hidden network. AI interfaces invite a similar error of individuation. The question is not only how extensively a model depends on human beings. Dependence, however dense, does not by itself generate obligation. AI systems depend upon us, and we are becoming increasingly dependent upon them, without either fact settling who can be held answerable.

I am working on a paper titled Answers Without Answerability, argues that responsibility depends heavily on the boundary drawn around the system. Re-individuating the relevant agent can redirect attention toward the people and institutions that build, deploy and benefit from it.

Scientific competence in the age of AI

AI can already generate conjectures, search literatures, plan experiments, write prose and assist with formal proof. These achievements make it important to distinguish the production of a successful result from the formation of a competent knower. Scientific competence includes knowing how a result was obtained, what would count against it, when an instrument should be trusted, how to respond to criticism, and who is responsible for carrying a claim into the shared body of knowledge.

This question connects my work on AI with my research on education and the formation of intellectual agency. Human education is not only the transfer of information or the shaping of reliable behavior. It inducts people into practices of giving reasons, receiving correction, judging the authority of others, and taking responsibility for what they claim. AI may reproduce some products of those practices without having undergone the trajectory through which human beings acquire standing within them.

I am interested in how AI changes the distribution of scientific labor without assuming that automation either destroys inquiry or solves it. An artificial system may be an instrument of extraordinary quality. The harder question is what must remain with a scientific community if the outputs of that instrument are to count as warranted knowledge rather than merely successful performance.

Agency without controllers

A planned project brings neuroscience and AI into comparison. Both fields study systems that can exhibit flexible, goal-directed behavior without containing anything like a single controller, executive or discrete decision stage. In neuroscience, activity associated with control is distributed across networks, and experimental tasks may impose divisions that the underlying dynamics do not contain. In artificial intelligence, reinforcement-learning agents, tool-using language models and embodied systems can behave adaptively even when no component plays the role of a central decision-maker.

I am exploring whether control and decision are sometimes better understood as system-level achievements: patterns of activity stabilized by task demands, reward structures, temporal pressures and environmental constraints. This does not assume that modular explanations are mistaken. It asks what evidence would distinguish a system that is modular by design from one that appears modular because an experiment has temporarily constrained its dynamics into a sequence of stages.

The comparison is useful only if the differences between brains and machines remain visible. I do not use AI as a simplified model of the brain, or neuroscience as a source of decorative analogies for AI. Each field provides a way of testing the concepts used in the other. The aim is to clarify what we mean by control and agency before we decide where they are located.

A connected research program

These projects are connected by a common concern with construction and boundaries. Metrics construct objects of evaluation. Experimental tasks help construct the behaviors identified as control. Interfaces construct an apparent agent. Institutions construct the paths through which outputs acquire authority and consequences.

My approach is deliberately cautious. I do not want to dismiss artificial systems as mere automation when their behavior requires better concepts, but neither do I want to infer agency, understanding, or responsibility from fluency alone. The task is to identify what our methods actually establish, what these systems can reliably do, and what remains ours to answer for.