Donal Khosrowi

Recommender Systems (RS) are among the most widely-deployed types of algorithmic systems, shaping the contents and items that billions of users see, engage with, and purchase on a daily basis. A dominant narrative in the literature characterizes RS as estimating user’s preferences and using this information to recommend good items for users. What is striking about this narrative is that it offers a welfare consequentialist justification for RS, in which preference satisfaction is the core welfare criterion that tells us how to rank outcomes in terms of agents' welfare. We argue that this justification is severely undertheorized and defective. More specifically, RS research currently lacks a coherent conceptualization of the relationship between preference, choice, and welfare. As a consequence, most RS research lacks an explicit and coherent theory of welfare-relevant preferences. Instead, it gestures towards outdated economic theories that have since been contravened by decades of research in the behavioral sciences and philosophy. To make our case, we draw out several interconnected challenges that call into question whether RS deals in genuine information about welfare-relevant preferences. To address these challenges, we sketch the contours of an interdisciplinary research program to put RS on coherent welfare-theoretical foundations.

Machine Unlearning (MU) aims to remove the influence of unwanted data from trained AI models, driven by ethical/legal concerns like privacy (e.g., the Right to be Forgotten), bias mitigation, security, and copyright protection. This paper critically examines MU, arguing that it is currently unclear whether its technical methods and ethical goals are suitably aligned. Currently, important questions around what MU does, what it should do, and how its efforts align with stakeholder needs remain unaddressed. Drawing on insights from social epistemology and the ethics of forgetting, the paper makes progress in clarifying what MU is and whether it aligns with the relevant goals. It does so by distinguishing three different senses of unlearning that vary in regard to what stakeholder needs they can cater to. Building upon cases regarding copyright and data privacy, the paper highlights potential alignment gaps between MU’s methods and its wider goals, and emphasizes the need for more concrete guidelines to assess MU’s effectiveness, clearer ethical foundations, and improved stakeholder engagement.

Some predictive systems do not merely predict, but their predictions shape and steer the world towards certain outcomes rather than others; they are performative. When predictive systems are performative, their development and deployment raises urgent ethical challenges and may place novel responsibilities on developers, deployers, regulators and policy-makers. While FAccT and other related communities have focused considerable attention on ethically significant problems regarding bias, fairness, discrimination, and others, little attention has been paid so far to the challenges raised by performative prediction. This paper details this gap in existing work, provides a snapshot of ongoing work across computer science and philosophy to point to areas where fruitful connections may be established, and issues a community-wide call for action and research to investigate and manage performative prediction and the new challenges it harbors.

AI assistants are increasingly used for navigating and analysing the contents of major archives. Applying Retrieval Augmented Generation to existing large language models, these tools draw on indexes of the relevant archives to answer, in natural language, users’ questions. In addition to being powerful finding aids, archival AI assistants are also presented as being capable of providing useful, automated answers to questions about the past. This article argues that such tools and how they are marketed result in major conceptual disruptions and uncertainties, placing pressure on our understanding of a range of roles, forms of information, and outputs involved in the production of historical knowledge. In particular, we argue that these tools may obscure well-established beliefs that ‘sources’ and ‘archives’ are not unmediated, clearly navigable, or necessarily comprehensive, and that the processes by which these are used to write ‘history’ are by no means straightforward or instantaneous. With the aim of mitigating these misunderstandings, the article makes suggestions for how deployers could more carefully frame and describe the intended use of archival AI assistants (especially for public users), to ensure that their benefits for accessibility are exploited while also avoiding misconceptions and safeguarding rigorous historical practice.

Researchers in history and the historical sciences explore the use of generative AI (GenAI) systems for reconstructing destroyed artifacts. This paper poses a novel question: Can such GenAI systems generate evidence that provides new knowledge about the world or can they only produce hypotheses that we might seek evidence for? Exploring responses to this question, the paper argues that (1) GenAI outputs can at least be understood as higher-order evidence (Parker 2022) and (2) may also constitute de novo synthetic evidence.

Algorithmic risk assessment tools, such as COMPAS, are increasingly used in criminal justice systems to predict the risk of defendants to reoffend in the future. This paper argues that these tools may not only predict recidivism, but may themselves causally induce recidivism through self-fulfilling predictions. We argue that such “performative” effects can yield severe harms both to individuals and to society at large, which raise epistemic-ethical responsibilities on the part of developers and users of risk assessment tools. To meet these responsibilities, we present a novel desideratum on algorithmic tools, called explainability-in-context, which requires clarifying how these tools causally interact with the social, technological, and institutional environments they are embedded in. Risk assessment practices are thus subject to high epistemic standards, which haven't been sufficiently appreciated to date. Explainability-in-context, we contend, is a crucial goal to pursue in addressing the ethical challenges surrounding risk assessment tools.

The recent wave of generative AI (GAI) systems like Stable Diffusion that can produce images from human prompts raises controversial issues about creatorship, originality, creativity and copyright. This paper focuses on creatorship: who creates and should be credited with the outputs made with the help of GAI? Existing views on creatorship are mixed: some insist that GAI systems are mere tools, and human prompters are creators proper; others are more open to acknowledging more significant roles for GAI, but most conceive of creatorship in an all-or-nothing fashion. We develop a novel view, called CCC (collective-centered creation), that improves on these existing positions. On CCC, GAI outputs are created by collectives in the first instance. Claims to creatorship come in degrees and depend on the nature and significance of individual contributions made by the various agents and entities involved, including users, GAI systems, developers, producers of training data and others. Importantly, CCC maintains that GAI systems can sometimes be part of a co-creating collective. We detail how CCC can advance existing debates and resolve controversies around creatorship involving GAI.

Public policymakers and institutional decision-makers routinely face questions about whether interventions “work”: does universal basic income improve people’s welfare and stimulate entrepreneurial activity? Do gated alleyways reduce burglaries or merely shift the crime burden to neighbouring communities? What is the most cost-effective way to improve students’ reading abilities? These are empirical questions that seem best answered by looking at the world, rather than trusting speculations about what will be effective. Evidence-based policy (EBP) is a movement that concretizes this intuition. It maintains that policy should be based on evidence of “what works.” Not any evidence will do, however. Following on the heels of its intellectual progenitor, evidence-based medicine (EBM), EBP insists on the use of high-quality evidence produced in accordance with rigorous methodological standards. Though intuitively compelling, EBP has attracted significant criticism from methodologists, political scientists, and philosophers. This chapter provides a critical overview of EBP through a philosophical lens, reviewing and discussing some of the most pressing challenges that EBP faces and outlining some proposals for improving it. I first provide a brief overview of EBP and distinguish a broader and narrower understanding of it. I then review existing lines of criticism of EBP, two of which are considered in detail. The first elaborates how EBP struggles with extrapolation, that is, the use of evidence from study populations to make inferences about the effects of policies on novel target populations. The second maintains that EBP’s methodological tenets are deeply entwined with moral and political values, which can threaten EBP’s promise to promote objectivity in policymaking. Finally, I make some proposals for how EBP could be improved going forward.

Scientific models can be performative: they can causally affect the phenomena they are intended to represent. The existing literature offers two responses. The appraisal view emphasizes that performativity can sometimes be a good-making model attribute, e.g., when predictions steer the public’s behavior in desirable ways. The mitigation view seeks to endogenize agents’ behavioral response to model-issued forecasts to get rid of performativity instead. This paper argues that neither approach is fully compelling: the appraisal view encounters severe concerns about moral values illegitimately encroaching on how modelers construct and use models, while the mitigation view fails to acknowledge that endogenization is itself a choice that involves substantive value-judgments relating to the desirability of certain social outcomes.

Extrapolating causal effects from experiments to novel populations is a common practice in evidence-based-policy, development economics and other social science areas. Drawing on experimental evidence of policy effectiveness, analysts aim to predict the effects of policies in new populations, which might differ importantly from experimental populations. Existing approaches made progress in articulating the sorts of similarities one needs to assume to enable such inferences. It is also recognized, however, that many of these assumptions will remain surrounded by significant uncertainty in practice. Unfortunately, the existing literature says little on how analysts may articulate and manage these uncertainties. This paper aims to make progress on these issues. First, it considers several existing ideas that bear on issues of uncertainty, elaborates the challenges they face, and extracts some useful rationales. Second, it outlines a novel approach, called the support graph approach, that builds on these rationales and allows analysts to articulate and manage uncertainty in extrapolation in a systematic and unified way.

Models not only represent but may also influence their targets in important ways. While models’ abilities to influence outcomes has been studied in the context of economic models, often under the label ‘performativity’, we argue that this phenomenon also pertains to epidemiological models, such as those used for forecasting the trajectory of the Covid-19 pandemic. After identifying three ways in which a model by the Covid-19 Response Team at Imperial College London may have influenced scientific advice, policy, and individual responses, we consider the implications of epidemiological models’ performative capacities. We argue, first, that performativity may impair models’ ability to successfully predict the course of an epidemic; but second, that it may provide an additional sense in which these models can be successful, namely by changing the course of an epidemic.

This paper investigates how intuitions about scientific discovery using artificial intelligence can be used to improve our understanding of scientific discovery more generally. Traditional accounts of discovery have been agent-centred: they place emphasis on identifying a specific agent who is responsible for conducting all, or at least the important part, of a discovery process. We argue that these accounts experience difficulties capturing scientific discovery involving AI and that similar issues arise for human discovery. We propose an alternative, collective-centred view as superior for understanding discovery, with and without AI. This view maintains that discovery is performed by a collective of agents and entities, each making contributions that differ in significance and character, and that attributing credit for discovery depends on various finer-grained properties of the contributions made. Detailing its conceptual resources, we argue that this view is considerably more compelling than its agent-centred alternative. Considering and responding to several theoretical and practical challenges, we point to concrete avenues for further developing the view we propose.

Extrapolating causal effects is becoming an increasingly important kind of inference in Evidence-Based Policy, development economics, and microeconometrics more generally. While several strategies have been proposed to aid with extrapolation, the existing methodological literature has left our understanding of what extrapolation consists of and what constitutes successful extrapolation underdeveloped. This paper addresses this lack in understanding by offering a novel account of successful extrapolation. Building on existing contributions pertaining to the challenges involved in extrapolation, this more nuanced and comprehensive account seeks to provide tools that facilitate the scrutiny of specific extrapolative inferences and general strategies for extrapolation. Offering such resources is important especially in view of the increasing amounts of real-world decision-making in policy, development, and beyond that involve extrapolation.

Acceding to the demand that public policy should be based on “the best available evidence” can come at significant moral cost. Important policy questions cannot be addressed using “the best available evidence” as defined by the evidence-based policy paradigm; the paradigm can change the meaning of questions so that they can be addressed using the preferred kind of evidence; and important evidence that does not meet the standard defined by the paradigm can get ignored. We illustrate these problems in three contexts of evidence-based policy.

I consider recent strategies proposed by econometricians for extrapolating causal effects from experimental to target populations. I argue that these strategies fall prey to the extrapolator’s circle: they require so much knowledge about the target population that the causal effects to be extrapolated can be identified from information about the target alone. I then consider comparative process tracing as a potential remedy. Although specifically designed to evade the extrapolator’s circle, I argue that CPT is unlikely to facilitate extrapolation in typical econometrics and evidence-based policy applications. To argue this, I offer a distinction between two kinds of extrapolation, attributive and predictive, the latter being prevalent in econometrics and evidence-based policy. I argue that CPT is not helpful for predictive extrapolation when using the kinds of evidence that econometricians and evidence-based policy researchers prefer. I suggest that econometricians may need to consider qualitative evidence to overcome this problem.

In Simulation and Similarity, Michael Weisberg offers a similarity-based account of the model–world relation, which is the relation in virtue of which successful models are successful. Weisberg’s main idea is that models are similar to targets in virtue of sharing features. An important concern about Weisberg’s account is that it remains silent on what it means for models and targets to share features, and consequently on how feature-sharing contributes to models’ epistemic success. I consider three potential ways of concretizing the concept of shared features: as identical, quantitatively sufficiently close, and sufficiently similar features. I argue that each of these concretizations faces significant challenges, leaving unclear how Weisberg’s account substantially contributes to elucidating the relation in virtue of which successful models are successful. Against this background, I outline a pluralistic revision and argue that this revision may not only help Weisberg's account evade several of the problems that I raise, but also offers a novel perspective on the model–world relation more generally. 1Introduction 2Weisberg’s Feature-Sharing Account 3What Is a Shared Feature? 3.1Identity 3.2Sufficient closeness 3.3Sufficient similarity 4Turning Weisberg’s Account ‘Upside Down’ 5Conclusion

Proponents of evidence-based policy (EBP) call for public policy to be informed by high-quality evidence from randomized controlled trials. This methodological preference aims to promote several epistemic values, e.g. rigor, unbiasedness, precision, and the ability to obtain causal conclusions. I argue that there is a trade-off between these epistemic values and several non-epistemic, moral and political values. This is because the evidence afforded by preferred EBP methods is differentially useful for pursuing different moral and political values. I expand on how this challenges ideals of value-freedom and -neutrality in EBP, and offer suggestions for how EBP methodology might be revised.