Johannes Himmelreich

This chapter examines how AI adoption in government threatens democracy by undermining the administrative state. Public servants do more than execute tasks—they uphold norms of responsible public service that include maintaining stability, gathering evidence, pursuing the public interest, and resisting democratic pandering. Through cases including Stanislav Petrov’s decision not to escalate a false nuclear alarm, the EPA’s initiative on radon regulation, and General Milley’s assurances during a contested transfer of power, the chapter illustrates the difference between performing a task well and doing a good job. AI excels at the former but struggles with the latter. The chapter develops norms of responsible public service as a framework for evaluating government AI and proposes practical governance measures including evidence standards, model deployment protections, and limits to automation.

co-edited with Bullock, Justin B., Chen, Yu-Che, Hudson, Valerie M., Korinek, Anton, Young, Matthew M., Zhang, Baobao

This handbook examines the bidirectional relationship between artificial intelligence and governance systems. Spanning 49 chapters across nine sections, it covers value foundations, regulatory ecosystems, governance frameworks, assessment and implementation, economic dimensions, domestic policy applications, and international politics. Contributors come from diverse disciplines including political science, computer science, economics, philosophy, and ethics, drawing on perspectives from academia, industry, think tanks, and civil society. The handbook addresses both the challenges and opportunities that AI presents for collective governance and individual wellbeing.

with Lim, Désirée

This chapter argues for a structural injustice approach to the governance of AI. Structural injustice has an analytical and evaluative component. The analytical component consists of structural explanations that are well known in the social sciences. The evaluative component is a theory of justice. Structural injustice is a powerful conceptual tool that allows researchers and practitioners to identify, articulate, and perhaps even anticipate, AI biases. The chapter begins with an example of racial bias in AI that arises from structural injustice. The chapter then presents the concept of structural injustice as introduced by the philosopher Iris Marion Young. The chapter moreover argues that structural injustice is well suited as an approach to the governance of AI and compares this approach to alternative approaches that start from analyses of harms and benefits or from value statements. The chapter suggests that structural injustice provides methodological and normative foundations for the values and concerns of diversity, equity, and inclusion (DEI). The chapter closes with a look into the idea of “structure” and responsibility. The idea of structure is central to justice. An open theoretical research question is to what extent AI is itself part of the structure of society. Finally, the practice of responsibility is central to structural injustice. Even if they cannot be held responsible for the existence of structural injustice, every individual and every organization has some responsibility to address structural injustice going forward.

This paper argues against the call to democratize artificial intelligence (AI). Several authors demand to reap purported benefits that rest in direct and broad participation: In the governance of AI, more people should be more involved in more decisions about AI—from development and design to deployment. This paper opposes this call. The paper presents five objections against broadening and deepening public participation in the governance of AI. The paper begins by reviewing the literature and carving out a set of claims that are associated with the call to “democratize AI”. It then argues that such a democratization of AI (1) rests on weak grounds, because it does not answer to a demand of legitimization, (2) is redundant in that it overlaps with existing governance structures, (3) is resource intensive, which leads to injustices, (4) is morally myopic and thereby creates popular oversights and moral problems of its own, and finally, (5) is neither theoretically nor practically the right kind of response to the injustices that animate the call. The paper concludes by suggesting that AI should be democratized not by broadening and deepening participation but by increasing the democratic quality of the administrative and executive elements of collective decision making. In a slogan: The question is not so much whether AI should be democratized but how.

This chapter reviews and evaluates different ways in which digital technologies may affect democracy. Specifically, the chapter develops a framework to evaluate democratic practices that is rooted in the tradition of deliberative democracy. The chapter then applies this framework to evaluate proposals of how technology may improve democracy. The chapter distinguishes three families of proposals depending on the depth of the change that they affect. Mere changes, such as automatic fact checking on social media, augment existing practices. Moderate reforms, such as apps that enable and reward participation in local government, facilitate new practices. Radical revisions, such as using artificial intelligence to replace parliaments, are constitutive of new practices often replacing existing ones. This chapter then concentrates on three radical revisions—Wiki democracy, avatar democracy, and data democracy—and identifies meaningful benefits in the first and deep problems in the latter two proposals.

with Young, Matthew M, Bullock, Justin B, Kim, Kyoung-Cheol

Artificial intelligence (AI) offers challenges and benefits to the public sector. This article presents an ethical framework to analyze the effects of AI in public organizations, guide empirical and theoretical research in public administration, and inform practitioner deliberation and decision making on AI adoption. The article puts forward six propositions on how the use of AI by public organizations may facilitate or prevent unnecessary harm. The framework builds on the theory of administrative evil and contributes to it in two ways. First, the article interprets the theory of administrative evil through the lens of agency theory, examining how the mechanisms stipulated by the former relate to the underlying mechanisms of the latter. Specifically, it highlights how mechanisms of administrative evil can be analyzed as information problems in the form of adverse selection and moral hazard. Second, the article describes possible causal pathways of the theory of administrative evil and associates each with a level of analysis: individual (micro), organizational (meso), and cultural (macro). The article then develops both descriptive and normative propositions on AI’s potential to increase or decrease the risk of administrative evil, contributing an institutional and public administration lens to the growing literature on AI safety and value alignment.

The ongoing debate on the ethics of self-driving cars typically focuses on two approaches to answering such questions: moral philosophy and social science. This article argues that these two approaches are both lacking. Neither deducing answers from individual moral theories nor expecting social science to give complete answers suffices. To supplement these approaches, political philosophy is needed. The issues at stake are collective decisions made together rather than individual decisions made in light of what individuals each have reason to value. Political philosophy adds three basic concerns to the conceptual toolkit: reasonable pluralism, human agency, and legitimacy. These three concerns have so far been largely overlooked in the debate on the ethics of self-driving cars.

This essay examines how U.S. export controls on artificial intelligence infrastructure constitute a form of technological gatekeeping with significant equity implications. The author identifies three primary “chokepoints” (specialized hardware, foundation models, and API access) that concentrate AI capabilities among wealthy nations and their allies. Drawing on political philosophy, the work argues that such restrictions violate principles of fair trade participation, constitute coercive economic practices, and ignore resource interdependencies in an interconnected global economy. While acknowledging legitimate security concerns, the analysis contends that these controls may prove counterproductive and that regional cooperation and restrained export policies better serve long-term international stability than technological dominance through artificial scarcity.

This chapter evaluates the global digital economy from the perspective of political, socio-economic, and intergenerational justice. The digital economy is defined as infrastructure provided or accessed online, which brings into clearer view the relevance of justice given its global reach and role in facilitating production across all aspects of life. The chapter argues that the digital economy raises no distinctive concerns from the perspective of socio-economic or intergenerational justice. Instead, the crucial problem is one of political justice in the form of illegitimate power. The digital economy poses four problems: an abridgment of state power, a degradation of economic opportunities and political relations, support of authoritarian politics, and leverage of international dominance.

with Hsu, Arbie, Veomett, Ellen, Lum, Kristian

A yet unmet challenge in algorithmic fairness is the problem of intersectionality, that is, achieving fairness across the intersection of multiple groups—and verifying that such fairness has been attained. Because intersectional groups tend to be small, verifying whether a model is fair raises statistical as well as moral-methodological challenges. This paper (1) elucidates the problem of intersectionality in algorithmic fairness, (2) develops desiderata to clarify the challenges underlying the problem and guide the search for potential solutions, (3) illustrates the desiderata and potential solutions by sketching a proposal using simple hypothesis testing, and (4) evaluates, partly empirically, this proposal against the proposed desiderata.

Much of the debate on the ethics of self-driving cars has revolved around trolley scenarios. This paper instead takes up the political or institutional question of who should decide how a self-driving car drives. Specifically, this paper is on the question of whether and why passengers should be able to control how their car drives. The paper reviews existing arguments—those for passenger ethics settings and for mandatory ethics settings respectively—and argues that they fail. Although the arguments are not successful, they serve as the basis to formulate desiderata that any approach to regulating the driving behavior of self-driving cars ought to fulfill. The paper then proposes one way of designing passenger ethics settings that meets these desiderata.

with Young, Matthew M., Honcharov, Danylo, Soundarajan, Sucheta

Administrative errors in unemployment insurance (UI) decisions give rise to a public values conflict between efficiency and efficacy. This article analyzes whether artificial intelligence (AI)—in particular, methods in machine learning (ML)—can be used to detect administrative errors in UI claims decisions, both in terms of accuracy and normative tradeoffs. The article uses 16 years of US Department of Labor audit and policy data on UI claims to analyze the accuracy of 7 different random forest and deep learning models. The article further tests weighting schemas and synthetic data approaches to correcting imbalances in the training data. A random forest model using gradient descent boosting is more accurate, along several measures, and preferable in terms of public values, than every deep learning model tested. Adjusting model weights produces significant recall improvements for low-n outcomes, at the expense of precision. Synthetic data produces attenuated improvements and drawbacks relative to weights.

with Köhler, Sebastian

The advent of intelligent artificial systems has sparked a dispute about the question of who is responsible when such a system causes a harmful outcome. This paper champions the idea that this dispute should be approached as a conceptual engineering problem. Towards this claim, the paper first argues that the dispute about the responsibility gap problem is in part a conceptual dispute about the content of responsibility and related concepts. The paper then argues that the way forward is to evaluate the conceptual choices we have, in the light of a systematic understanding of why the concept is important in the first place—in short, the way forward is to engage in conceptual engineering. The paper then illustrates what approaching the responsibility gap problem as a conceptual engineering problem looks like. It outlines argumentative pathways out of the responsibility gap problem and relates these to existing contributions to the dispute.

with Young, Matthew M, Honcharov, Danylo, Soundarajan, Sucheta

This article explores the extent to which machine learning can be used to detect administrative errors. It concentrates on administrative errors in unemployment insurance (UI) decisions, which give rise to a public values conflict between efficiency and effectiveness. This conflict is first described and then highlighted in the history of the US UI regime. Machine learning may not only mitigate this conflict but it may also help to combat fraud and reduce the backlog of claims associated with economic crises such as the COVID-19 pandemic. The article uses data about improper UI payments throughout the US from 2002 through 2018 to analyze the accuracy of random forests and deep learning models. The analysis finds that a random forest model using gradient descent boosting is more accurate, along several measures, than every deep learning model tested. This finding could be explained by the goodness-of-fit between the machine learning method and the available data. Alternatively, deep learning performance could be attenuated by necessary limits to publicly-accessible claims data.

Future weapons will make life-or-death decisions without a human in the loop. When such weapons inflict unwarranted harm, no one appears to be responsible. There seems to be a responsibility gap. This paper first reconstructs the argument for such responsibility gaps and then argues that this argument is not sound. The argument assumes that commanders have no control over whether autonomous weapons inflict harm. The paper argues against this assumption. Although this investigation concerns a specific case of autonomous weapons systems, the paper takes steps towards vindicating the more general idea that superiors can be morally responsible in virtue of being in command.

The asylum system faces problems on two fronts. States undermine it with populist politics, and migrants use it to satisfy their migration preferences. To address these problems, asylum services should be commodified. States should be able to pay other states to provide determination and protection-elsewhere. This article aims to identify a way of implementing this idea that is both feasible and desirable. First, the article sketches a policy proposal for a commodification of asylum services. Then, it argues that this policy proposal is not only compatible with the right to asylum, but also supported by moral considerations. Despite some undesirable moral features, a market in asylum facilitates the provision of asylum to those who need it.

Trolley cases are widely considered central to the ethics of autonomous vehicles. This paper cautions against this by identifying four problems. (1) Trolley cases, given technical limitations, rest on assumptions that are in tension with one another. Furthermore, (2) trolley cases illuminate only a limited range of ethical issues insofar as they cohere with a certain design framework. Furthermore, (3) trolley cases seem to demand a moral answer when a political answer is called for. Finally, (4) trolley cases might be epistemically problematic in several ways. To put forward a positive proposal, the paper illustrates how ethical challenges arise from mundane driving situations. It argues that mundane situations are relevant because of the specificity they require and the scale they exhibit. The paper then illustrates some of the ethical challenges arising from optimizing for safety, balancing safety with other values such as mobility, and adjusting to incentives of legal frameworks.

Related: The everyday ethical challenges of self-driving cars (The Conversation)

The disappearing agent problem is an argument in the metaphysics of agency. Proponents of the agent-causal approach argue that the rival event-causal approach fails to account for the fact that an agent is active. This paper examines an analogy between this disappearing agent problem and the exclusion problem in the metaphysics of mind. The paper develops the analogy between these two problems and surveys existing solutions. It suggests that some solutions that have received significant attention in response to the exclusion problem have seen considerably less attention in response to the disappearing agent problem. For example, one solution to the exclusion problem is to reject the exclusion assumption. Analogously, one solution to the disappearing agent problem could be to deny the claim that the agent-causal approach and the event-causal approach are mutually exclusive. Similarly, proportionality theories of causation, a solution to the exclusion problem, can be transferred to the disappearing agent problem. After establishing the plausibility of the analogy between the two problems, the paper examines how this latter solution in particular can be transferred from the one problem to the other.

Hierarchical groups shape social, political, and personal life. This paper concerns the question of how individuals within such groups can be responsible. The paper explores how individual responsibility can be partially grounded in difference-making. The paper concentrates on the control condition of responsibility and takes into view three distinct phenomena of responsibility in hierarchical groups. First, a superior can be responsible for outcomes that her subordinates bring about. Second, a subordinate can be responsible although she is unable to prevent the outcome she brings about. Third, a superior can sometimes be responsible to a greater degree than her subordinates. It is argued that difference-making, as an interpretation of the control condition that partially grounds responsibility, accounts for all three of these phenomena within a limited but significant range of circumstances and can hence partially ground individual moral responsibility in hierarchical groups. The paper provides an element of a theory of individual responsibility to complement theories of corporate responsibility.

This paper is on the problem of profligate omissions. The problem is that counterfactual definitions of causation identify as a cause anything that could have prevented an effect but that did not actually occur, which is a highly counterintuitive result. Many solutions of this problem appeal to normative, epistemic, pragmatic, or metaphysical considerations. These existing solutions are in some sense substantive. In contrast, this paper concentrates on the semantics of counterfactuals. The paper proposes to replace Strong Centering with Weak Centering. This allows that the actual world is not the only world that is closest to the actual world. As a result, some counterfactuals that would otherwise have been true, turn out to be false. When these counterfactuals concern causation, fewer causal claims are true. In addition to describing steps towards solving the problem of profligate omissions, the proposal captures an abstraction that is shared by many of the existing solutions: depending on how the distance ordering underlying the Weak Centering condition is constructed and interpreted, some of these existing solutions can be recovered.

A central dispute in social ontology concerns the existence of group minds and actions. This paper argues that some authors in this dispute rely on rival views of existence without sufficiently acknowledging this divergence. The paper proceeds in three steps. First, it defines the phenomenon as an implicit higher-order disagreement by drawing on an analysis of verbal disputes. Second, it distinguishes two theories of existence—the theory-commitments view and the truthmaker view—in both their eliminativist and their constructivist variants. Third, it examines individual contributions to the dispute about the existence of group minds and actions to argue that these contributions have an implicit higher-order disagreement. This paper serves two purposes. First, it is a study to apply recent advances in meta-ontology. Second, it contributes to the debate on social ontology by illustrating how meta-ontology matters for social ontology.

with Lawford-Smith, Holly

Punishing groups raises a difficult question, namely, how their punishment can be justified at all. Some have argued that punishing groups is morally problematic because of the effects that the punishment entails for their members. This paper argues against this view. The paper distinguishes the question of internal justice—how punishment-effects are distributed—from the question of external justice—whether the punishment is justified. The paper argues that issues of internal justice do not in general undermine the permissibility of punishment. The paper also defends the permissibility of what some call “random punishment,” arguing that, for some kinds of collectives, there is no general obligation to internally distribute the punishment-effects equally or in proportion to individual contribution.

This paper develops a taxonomy of kinds of actions that can be seen in group agency, human–machine interactions, and virtual realities. These kinds of actions are special in that they are not embodied in the ordinary sense. The paper begins by analysing the notion of embodiment into three separate assumptions that together comprise the Embodiment View. Although this view may find support in paradigmatic cases of agency, the paper suggests that each of its assumptions can be relaxed. With each assumption that is given up, a different kind of disembodied action becomes available. The taxonomy gives a systematic overview and suggests that disembodied actions have the same theoretical relevance as the actions of any ordinarily embodied human.

This paper is about the status of collective actions. According to one view, collective actions metaphysically reduce to individual actions because sentences about collective actions are merely a shorthand for sentences about individual actions. The paper reconstructs an argument for this view and shows via counterexamples that it is not sound. The argument relies on a paraphrase procedure to unpack alleged shorthand sentences about collective actions into sentences about individual actions. The paper argues that the best paraphrase procedure that has been put forward so far fails to produce adequate results.

Related: This paper prompted a discussion note

This paper replicates a prior study claiming that OpenAI’s reasoning model o1-mini produces more utilitarian responses than GPT-4o on moral dilemmas. The trolley problem finding does not hold: GPT-4o’s lower utilitarian response rate stems from safety refusals rather than philosophical commitments. When prompts ask “Is it morally permissible...?” instead of “Should I...?”, GPT-4o reaches 99% utilitarian responses. All models converge on utilitarian answers when prompt confounds are removed. The footbridge dilemma finding partially survives, though reasoning models frequently refuse to answer or provide non-utilitarian responses. The study concludes that single-prompt evaluations of LLM moral reasoning are unreliable and advocates for multi-prompt robustness testing as standard practice when assessing language model behavior.

with Cohen, Joshua

This article describes a teaching plan for a discussion-driven introduction to moral reasoning and explains its philosophical and pedagogical rationale. The teaching plan consists of a sequence of thought experiments that build on one another, and ends with participants addressing some morally complex, real-life issues. The plan rests on extensive experience teaching moral reasoning in several different professional learning environments. The main contribution of this article is practical. The goal is to equip educators with a pedagogical approach and ready-to-use teaching materials. To this end, the article offers the methodological background, identifies learning objectives as well as pitfalls of teaching the trolley problem, and describes the pedagogy of the session.

with Alexander, Jason McKenzie, Thompson, Christopher

This article examines two questions about scientists’ search for knowledge. First, which search strategies generate discoveries effectively? Second, is it advantageous to diversify search strategies? The article argues, pace Weisberg and Muldoon’s “Epistemic Landscapes and the Division of Cognitive Labor” (this journal, 2009), that a search strategy deliberately seeking novel research approaches need not be optimal. On the second question, the article argues that Weisberg and Muldoon have not shown epistemic reasons exist for the division of cognitive labor, identifying the errors that led to their conclusions. Furthermore, the article generalizes the epistemic landscape model, showing that one should be skeptical about the benefits of social learning in epistemically complex environments.