Indrė Žliobaitė

The growing use of artificial intelligence (AI) in scientific research is reshaping how measurements are produced, interpreted, and trusted. Rather than relying on fixed, physically motivated measurement functions and summary statistics, contemporary science increasingly employs computational inference methods that learn mappings from data to scientifically relevant quantities. These learned measurement instruments support the analysis of complex, large-scale data and open new possibilities for scientific discovery, but they also challenge classical assumptions about measurement, uncertainty, and epistemic responsibility. In this paper, I examine AI-based methods in their role as learned measurement instruments. The behaviour of these instruments is determined not only by design and calibration, but also by training data, modeling choices, and implicit assumptions. I argue that such methods introduce forms of epistemic risk that go beyond stochastic noise and are not fully captured by existing measurement theory or current explainable AI techniques. In particular, I show that when measurement instruments learn, epistemic uncertainty no longer merely constrains inference drawn from measurements, but plays a formative role in determining what counts as a measurement outcome in the first place. At the same time, because these instruments perform inferential tasks that were traditionally part of human scientific reasoning, their use blurs the traditional boundary between instruments and agents, particularly when AI systems are described as generating hypotheses, guiding experimental design, or contributing to scientific reasoning. By analyzing this reconfiguration of measurement and inference, I clarify what changes when measurement instruments learn and show how AI reshapes the traditional boundary between instruments and agents in science without attributing strong agency to AI systems. I conclude by outlining the epistemic risks and responsibilities that arise from delegating measurement and inference to learned instrumentation, and by reflecting on how scientific understanding can be maintained when using AI-based and partially non-transparent instrumentation for scientific inference.

The growing use of artificial intelligence (AI) in scientific research is reshaping how measurements are produced, interpreted, and trusted. Rather than relying on fixed, physically motivated measurement functions and summary statistics, contemporary science increasingly employs computational inference methods that learn mappings from data to scientifically relevant quantities. These learned measurement instruments support the analysis of complex, large-scale data and open new possibilities for scientific discovery, but they also challenge classical assumptions about measurement, uncertainty, and epistemic responsibility. In this paper, I examine AI-based methods in their role as learned measurement instruments. The behaviour of these instruments is determined not only by design and calibration, but also by training data, modeling choices, and implicit assumptions. I argue that such methods introduce forms of epistemic risk that go beyond stochastic noise and are not fully captured by existing measurement theory or current explainable AI techniques. In particular, I show that when measurement instruments learn, epistemic uncertainty no longer merely constrains inference drawn from measurements, but plays a formative role in determining what counts as a measurement outcome in the first place. At the same time, because these instruments perform inferential tasks that were traditionally part of human scientific reasoning, their use blurs the traditional boundary between instruments and agents, particularly when AI systems are described as generating hypotheses, guiding experimental design, or contributing to scientific reasoning. By analyzing this reconfiguration of measurement and inference, I clarify what changes when measurement instruments learn and show how AI reshapes the traditional boundary between instruments and agents in science without attributing strong agency to AI systems. I conclude by outlining the epistemic risks and responsibilities that arise from delegating measurement and inference to learned instrumentation, and by reflecting on how scientific understanding can be maintained when using AI-based and partially non-transparent instrumentation for scientific inference.

Increasing numbers of decisions about everyday life are made using algorithms. By algorithms we mean predictive models (decision rules) captured from historical data using data mining. Such models often decide prices we pay, select ads we see and news we read online, match job descriptions and candidate CVs, decide who gets a loan, who goes through an extra airport security check, or who gets released on parole. Yet growing evidence suggests that decision making by algorithms may discriminate people, even if the computing process is fair and well-intentioned. This happens due to biased or non-representative learning data in combination with inadvertent modeling procedures. From the regulatory perspective there are two tendencies in relation to this issue: (1) to ensure that data-driven decision making is not discriminatory, and (2) to restrict overall collecting and storing of private data to a necessary minimum. This paper shows that from the computing perspective these two goals are contradictory. We demonstrate empirically and theoretically with standard regression models that in order to make sure that decision models are non-discriminatory, for instance, with respect to race, the sensitive racial information needs to be used in the model building process. Of course, after the model is ready, race should not be required as an input variable for decision making. From the regulatory perspective this has an important implication: collecting sensitive personal data is necessary in order to guarantee fairness of algorithms, and law making needs to find sensible ways to allow using such data in the modeling process.