Markus Kneer

The ubiquitous integration of AI-powered systems in morally consequential decision-making procedures raises a thorny question: when such systems generate harm, who should be held responsible? A prominent regulatory response proposes that suitably designed control architectures can ensure that blame is appropriately allocated. To live up to this promise, the proposed architectures must be both normatively adequate and regarded as such, since otherwise they are at best practically useless, and at worst useless because morally problematic. We examine two of the most widely discussed proposals. Experiment 1 (n=260) investigates whether laypeople attribute responsibility differently across architectures that place human agents in-, on-, and out-of-the-loop. Experiment 2 (n=520) extends this inquiry by differentiating between control structures that either violate or satisfy Santoni de Sio’s and van den Hoven’s (2018) ‘track and trace’ requirements. Our results reveal that neither proposal fully succeeds in directing laypeople’s responsibility judgements along the pathways they prescribe. Implications are discussed.

Philosophers disagree about what makes news “fake”: some believe it is falsity, whereas others emphasize intentional deception about the content or the source. We report two preregistered studies (total N=1200) testing whether falsity and source deception predict fake-news classification. Participants evaluated scenarios in which a claim was either true or false, and the article appeared on either the official New York Times website or that of a near-identical impersonator. In both studies, falsity and source deception independently increased fake-news classifications, indicating that many participants did not treat “fake news” as a synonym for “false news”. Whereas with a predictive claim (Study 1), the factors amplified each other, with a past-tense claim (Study 2), falsity was near-determinative and the interaction not reliable. Across both studies, conservatism predicted increased fake-news classifications, but the association disappeared once perceived source reliability was included, suggesting politics operates via biased priors rather than differences in meaning.

Legal judgment is threatened by a plethora of biases. Among those, the effect of outcome bias – the tendency to let outcome severity distort assessments of mental states and culpability – is particularly well documented and its impact is hypothesized to be pervasive. AI advisory systems could in principle help alleviate human bias, but Almeida et al. (2024) found substantial outcome effects in several commercial LLMs, suggesting AI judgment may be comparably biased. We reassess these findings using eight current open-source models (Cogito, Deepseek- R1, Gemma 3, Llama 3.1, Mistral, Phi 4) across multiple parameter sizes and temperature settings, replicating the Flood Paradigm from Kneer and Skoczeń (2023). Using a multilevel model with 100 responses per condition per temperature, we find that current open-source models exhibit little to no outcome bias—in stark contrast to both human participants and the older LLMs tested previously. Only Cogito and Deepseek-R1 show small effects, and temperature variation contributes negligibly. However, models differ substantially in baseline judgment severity—Mistral rates culpability considerably higher than Gemma 3 or Phi 4—demonstrating that absence of outcome bias does not guarantee neutrality. Our findings support cautious optimism about research into open-source LLMs as debiasing tools in legal contexts, while emphasizing that bias is multidimensional and responsible deployment requires evaluation across bias types and doctrinal domains.

Two topics at the center of Ethics of AI and HRI regard trust in AI agents as well as the adjudication of moral responsibility in situations where AI causes harm. In this paper we aim to advance the state of the art concerning these topics in several regards: First, we propose and evaluate a new empirical paradigm for measuring appropriate or calibrated trust in AI, that is, attitudes which are neither too trusting nor too cautious. The best way to measure calibrated trust, we argue, is by contrasting trust vested in AI agents when their relevant capacities equal those of a human expert in the domain. A second shortcoming of extant work regards generalizability: Trust in, and reliance on, AI are standardly explored with respect to a single context or domain. To investigate context-sensitivity, we ran experiments (total N=1276) across five key areas of AI application. Finally, we explored perceived moral responsibility for harm caused in human-AI interaction, with a particular focus on recent philosophical debates on the topic. Our findings suggest that approximately half of the participants vest equal trust in AI and human agents when their capacities are the same. However, there is considerable variation in trust calibration across domains, suggesting that context-sensitivity needs more attention. Human agents are attributed more moral responsibility than AI agents, whereas their supervisors are blamed less than those of AI agents. This suggests that, at least according to folk morality, there are no perceived "responsibility gaps" (Matthias 2004; Sparrow 2007) and that "retribution gaps" are a genuine possibility (Danaher 2016).

This article reports three empirical studies regarding partially aggregative moral theories in distributive ethics (total N = 417). We begin by documenting the widespread occurrence of the intuitions that motivate partial aggregation views. Thereafter, we advance the literature along two dimensions: First, we extend experimental work by ascertaining which amongst existing versions of partial aggregation (localised vs. global) chimes more fully with moral common sense. Specifically, we document how judgments about ‘irrelevant goods’ (Kamm Citation1993) in tie-breaking cases are just as robust as the original aggregative/non-aggregative pair of judgments that constitute partial aggregation. Second, we conduct an experiment that pairs laypeople’s moral judgments in standard cases with their intuitions about the limits of permissible self-prioritisation. We investigate whether one prominent explanation of why irrelevant claims may not be aggregated (Voorhoeve’s ‘personal prerogative’ argument) can be said to underpin people’s intuitions about the (ir)relevance relation of claims in conflict cases. We close with a discussion of our findings’ import and highlight avenues for future research.

Complex moral trade-offs are a basic feature of human life: for example, confronted with scarce medical resources, doctors must frequently choose who amongst equally deserving candidates receives medical treatment. But choosing what to do in moral trade-offs is no longer a ‘humans-only’ task, but often falls to AI agents. In this article, we report findings from a series of experiments (N=1029) intended to establish whether agent-type (Human vs. AI) matters for what should be done in moral trade-offs. We find that, relative to a human decision-maker, participants more often judge that AI agents should opt for fairness at the expense of maximizing utility. In our discussion, we explain how the reported differences (we call them agent-type ‘value forks’) matters for the study of moral value alignment, and we hypothesize what could explain these value forks. We close by reflecting on limits of our results and indicate avenues of further research.

What is the folk concept of art? Does it track any of the major definitions of art philosophers have proposed? In two preregistered experiments (N = 888) focusing on two types of artworks (paintings and musical works), we manipulate three potential features of artworks: intentional creation, the possession of aesthetic value, and institutional recognition. This allows us to investigate whether the folk concept of art fits an essentialist definition drawing on one or more of the manipulated factors or whether it might be a disjunctive or cluster concept. The results suggest that none of the three manipulated properties alone suffices for an object to be considered art. The folk concept of art might thus well be a cluster concept instead of an essentialist concept.

Danaher (2016) has argued that increasing robotization can lead to retribution gaps: Situations in which the normative fact that nobody can be justly held responsible for a harmful outcome stands in conflict with our retributivist moral dispositions. In this paper, we report a cross-cultural empirical study based on Sparrow’s (2007) famous example of an autonomous weapon system committing a war crime, which was conducted with participants from the US, Japan and Germany. We find that (1) people manifest a considerable willingness to hold autonomous systems morally responsible, (2) partially exculpate human agents when interacting with such systems, and that more generally (3) the possibility of normative responsibility gaps is indeed at odds with people’s pronounced retributivist inclinations. We discuss what these results mean for potential implications of the retribution gap and other positions in the responsibility gap literature.

The concept of a good life is usually assumed by philosophers to be equivalent to that of well-being, or perhaps of a morally good life, and hence has received little attention as a potentially distinct subject matter. In a series of experiments participants were presented with vignettes involving socially sanctioned wrongdoing toward outgroup members. Findings indicated that, for a large majority, judgments of bad character strongly reduce ascriptions of the good life, while having no impact at all on ascriptions of happiness or well-being. Taken together with earlier findings these results suggest that the lay concept of a good life is clearly distinct from those of happiness, well-being, or morality, likely encompassing both morality and well-being, and perhaps other values as well: whatever matters in a person’s life. Importantly, morality appears not to play a fundamental role in either happiness or well-being among the folk.

In two experiments (total N=693) we explored whether people are willing to consider paintings made by AI-driven robots as art, and robots as artists. Across the two experiments, we manipulated three factors: (i) agent type (AI-driven robot v. human agent), (ii) behavior type (intentional creation of a painting v. accidental creation), and (iii) object type (abstract v. representational painting). We found that people judge robot paintings and human paintings as art to roughly the same extent. However, people are much less willing to consider robots as artists than humans, which is partially explained by the fact that they are less disposed to attribute artistic intentions to robots.

Responding to recent concerns about the reliability of the published literature in psychology and other disciplines, we formed the X-Phi Replicability Project (XRP) to estimate the reproducibility of experimental philosophy (osf.io/dvkpr). Drawing on a representative sample of 40 x-phi studies published between 2003 and 2015, we enlisted 20 research teams across 8 countries to conduct a high-quality replication of each study in order to compare the results to the original published findings. We found that x-phi studies – as represented in our sample – successfully replicated about 70% of the time. We discuss possible reasons for this relatively high replication rate in the field of experimental philosophy and offer suggestions for best research practices going forward.

According to Anscombe, acting intentionally entails knowledge in action. This thesis has been near-universally rejected due to a well-known counterexample by Davidson: a man intending to make ten legible carbon copies might not believe with confidence, and hence not know, that he will succeed. If he does, however, his action surely counts as intentional. Damaging as it seems, an even more powerful objection can be levelled against Anscombe: while acting, there is as yet no fact of the matter as to whether the agent will succeed. Since his belief that he will is not yet true while his action is in progress, he cannot possibly know that he is indeed bringing about the intended goal. Knowledge in action is not only unnecessary for intentional action, it seems, but-at least as regards success-bound types of action-impossible to attain in the first place.

Despite a voluminous literature on happiness and well‐being, debates have been stunted by persistent dissensus on what exactly the subject matter is. Commentators frequently appeal to intuitions about the nature of happiness or well‐being, raising the question of how representative those intuitions are. In a series of studies, we examined lay intuitions involving happiness‐ and well‐being‐related terms to assess their sensitivity to internal (psychological) versus external conditions. We found that all terms, including ‘happy’, ‘doing well’ and ‘good life’, were far more sensitive to internal than external conditions, suggesting that for laypersons, mental states are the most important part of happiness and well‐being. But several terms, including ‘doing well’, ‘good life’ and ‘enviable life’ were substantially more sensitive to external conditions than others, such as ‘happy’, consistent with dominant philosophical views of well‐being. Interestingly, the expression ‘happy’ was completely insensitive to external conditions for about two thirds of our participants, suggesting a purely psychological concept among most individuals. Overall, our findings suggest that lay thinking in this domain divides between two concepts, or families thereof: a purely psychological notion of being happy, and one or more concepts equivalent to, or encompassing, the philosophical concept of well‐being. In addition, being happy is dominantly regarded as just one element of well‐being. These findings have considerable import for philosophical debates, empirical research and public policy.

This chapter reviews empirical research on the rules governing assertion and retraction, with a focus on the normative role of truth. It examines whether truth is required for an assertion to be considered permissible, and whether there is an expectation that speakers retract statements that turn out to be false. Contrary to factive norms (such as the influential “knowledge norm”), empirical data suggests that there is no expectation that speakers only make true assertions. Additionally, contrary to truth-relativist accounts, there is no requirement for speakers to retract statements that are false at the context of assessment. We conclude by suggesting that truth still plays a crucial role in the evaluation of assertions: as a standard for evaluating their success, rather than permissibility.

Extant research has shown that ordinary causal judgments are sensitive to normative factors. For instance, agents who violate a norm are standardly deemed more causal than norm-conforming agents in identical situations. In this paper, we explore two competing explanations for the Norm Effect: the Responsibility View and the Bias View. According to the former, the Norm Effect arises because ordinary causal judgment is intimately intertwined with moral responsibility. According to the alternative view, the Norm Effect is the result of a blame-driven bias. In a series of five preregistered experiments (N = 2688), we present evidence that predominantly favours the Bias View. In particular, and against predictions made by the Responsibility View, we show that participants deem agents who violate nonpertinent or silly norms – norms that do not relate to the outcome at hand – as more causal, and that this effect cannot be explained in terms of plausible mediators such as the agent’s foreknowledge and desire, or the foreseeability of harm. We close with a discussion regarding the implications of these findings, in as regards the just assessment of proximate cause in the law.

In two experiments (N=888), we explore to what extent the folk concept of art is compatible with the leading philosophical definitions of art, and whether it is an essentialist or a non-essentialist concept. We manipulate three factors: whether an object is created intentionally, whether it has aesthetic value, and whether it is institutionally recognized. In addition, we also manipulate the artistic domain (visual art or music). The results suggest that none of the three properties is seen by the folk as necessary for an object to be considered art, which suggests that the folk concept of art might be a cluster concept.

A cross-cultural survey experiment revealed a dominant tendency to rely on a rule’s letter over its spirit when deciding which behaviors violate the rule. This tendency varied markedly across (k = 15) countries, owing to variation in the impact of moral appraisals on judgments of rule violation. Compared with laypeople, legal experts were more inclined to disregard their moral evaluations of the acts altogether and consequently exhibited stronger textualist tendencies. Finally, we evaluated a plausible mechanism for the emergence of textualism: in a two-player coordination game, incentives to coordinate in the absence of communication reinforced participants’ adherence to rules’ literal meaning. Together, these studies (total n = 5,794) help clarify the origins and allure of textualism, especially in the law. Within heterogeneous communities in which members diverge in their moral appraisals involving a rule’s purpose, the rule’s literal meaning provides a clear focal point—an identifiable point of agreement enabling coordinated interpretation among citizens, lawmakers, and judges.

In an important and widely discussed series of studies, Jonathan Phillips and colleagues have suggested that the ordinary concept of happiness has a substantial moral component. For in- stance, two persons who enjoy the same extent of positive emotions and are equally satisfied with their lives are judged as happy to different degrees if one is less moral than the other. Considering that the relation between morality and happiness or self-interest has been one of the central questions of moral philosophy since at least Plato, such a result would be of considerable philosophical interest. On closer examination of the original research and new studies, we suggest that the data point to a different conclusion: in the dominant folk understanding of happiness, morality has no fundamental role. Findings seeming to indicate a moralized concept are better explained, we suggest, by folk theories on which extreme moral turpitude indicates that an individual suffers from psychological dysfunction.

Predictability is a prerequisite for effective human control of artificial intelligence (AI). The inability to predict malfunctioning of AI, for example, impedes timely human intervention. In this paper, we empirically investigate how AI’s predictability compares to the predictability of humans in a real-effort task. We show that humans are worse at predicting AI performance than at predicting human performance. Importantly, participants are not aware of the differences in relative predictability of AI and overestimate their prediction skills. These results raise doubts about the human ability to effectively exercise control of AI — at least in certain contexts.

The reasonable person standard is of great importance to US criminal and tort law. According to the law, whether or not an agent acted reasonably does not depend on features of the outcome which are not under her control. Mock juror attributions of reasonableness, however, are shown to be outcome-dependent. A series of experiments reveals that this outcome-dependence does not constitute a bias, since the very folk concept of reasonableness is outcome-sensitive. Consequently, the law makes a mistaken assumption as to what kind of concept of reasonableness lay jurors will apply in court. This conceptual misalignment, it is argued, could lead to serious injustice in US trials. This paper presents a series of studies (total N=579) which demonstrate that folk judgments concerning the reasonableness of decisions and actions depend strongly on whether they engender positive or negative consequences. A particular decision is deemed more reasonable in retrospect when it produces beneficial consequences than when it produces harmful consequences, even if the situation in which the decision was taken and the epistemic circumstances of the agent are held fixed across conditions. This finding is worrisome for the law, where the reasonable person standard plays a prominent role. The legal concept of reasonableness is outcome-insensitive: whether the defendant acted in a reasonable fashion or not depends exclusively on her context of action, no matter how things play out. Folk judgments of reasonableness are thus inconsistent with the legal concept of reasonableness. Problematically, in common law jurisdictions, the decision whether a defendant’s behavior was reasonable or not is frequently (though not necessarily) delegated to a lay jury.

While philosophers hold that it is patently absurd to blame robots or hold them morally responsible [1], a series of recent empirical studies suggest that people do ascribe blame to AI systems and robots in certain contexts [2]. This is disconcerting: Blame might be shifted from the owners, users or designers of AI systems to the systems themselves, leading to the diminished accountability of the responsible human agents [3]. In this paper, we explore one of the potential underlying reasons for robot blame, namely the folk's willingness to ascribe inculpating mental states or "mens rea" to robots. In a vignette-based experiment (N=513), we presented participants with a situation in which an agent knowingly runs the risk of bringing about substantial harm. We manipulated agent type (human v. group agent v. AI-driven robot) and outcome (neutral v. bad), and measured both moral judgment (wrongness of the action and blameworthiness of the agent) and mental states attributed to the agent (recklessness and the desire to inflict harm). We found that (i) judgments of wrongness and blame were relatively similar across agent types, possibly because (ii) attributions of mental states were, as suspected, similar across agent types. This raised the question - also explored in the experiment - whether people attribute knowledge and desire to robots in a merely metaphorical way (e.g., the robot "knew" rather than really knew). However, (iii), according to our data people were unwilling to downgrade to mens rea in a merely metaphorical sense when given the chance. Finally, (iv), we report a surprising and novel finding, which we call the inverse outcome effect on robot blame: People were less willing to blame artificial agents for bad outcomes than for neutral outcomes. This suggests that they are implicitly aware of the dangers of overattributing blame to robots when harm comes to pass, such as inappropriately letting the responsible human agent off the moral hook.

The reasonable person standard is key to both Criminal Law and Torts. What does and does not count as reasonable behavior and decision-making is frequently deter- mined by lay jurors. Hence, laypeople’s understanding of the term must be considered, especially whether they use it predominately in an evaluative fashion. In this corpus study based on supervised machine learning models, we investigate whether laypeople use the expression ‘reasonable’ mainly as a descriptive, an evaluative, or merely a value-associated term. We find that ‘reasonable’ is predicted to be an evaluative term in the majority of cases. This supports prescriptive accounts, and challenges descriptive and hybrid accounts of the term—at least given the way we operationalize the latter. Interestingly, other expressions often used interchangeably in jury instructions (e.g. ‘careful,’ ‘ordinary,’ ‘prudent,’ etc), however, are predicted to be descriptive. This indicates a discrepancy between the intended use of the term ‘reasonable’ and the understanding lay jurors might bring into the court room.

A growing body of literature has revealed ordinary causal judgement to be sensitive to normative factors, such that a norm-violating agent is regarded more causal than their non-norm-violating counterpart. In this paper, we explore two competing explanations for this phenomenon: the Responsibility View and the Bias View. The Bias View, but not the Responsibility View, predicts features peripheral to the agent’s responsibility to impact causal attributions. In a series of three preregistered experiments (N = 1162), we present new evidence that the Norm Effect arises from such peripheral features, namely from nonpertinent or entirely silly norm violations. Furthermore, we show that this effect cannot be explained by recourse to the agent’s foreknowledge or desire of the outcome, nor by its foreseeability: the Norm Effect arises even when participants judge the norm-violating agent’s doing as equally foreseeable. This, we argue, provides evidence in favour of the Bias View.

This article reports three empirical studies regarding partially aggregative moral theories in distributive ethics (total N = 417). We begin by documenting the widespread occurrence of the intuitions that motivate partial aggregation views. Thereafter, we advance the literature along two dimensions: First, we extend experimental work by ascertaining which amongst existing versions of partial aggregation (localised vs. global) chimes more fully with moral common sense. Specifically, we document how judgments about ‘irrelevant goods’ (Kamm 1993) in tie-breaking cases are just as robust as the original aggregative/non-aggregative pair of judgments that constitute partial aggregation. Second, we conduct an experiment that pairs laypeople’s moral judgments in standard cases with their intuitions about the limits of permissible self-prioritisation. We investigate whether one prominent explanation of why irrelevant claims may not be aggregated (Voorhoeve’s ‘personal prerogative’ argument) can be said to underpin people’s intuitions about the (ir)relevance relation of claims in conflict cases. We close with a discussion of our findings’ import and highlight avenues for future research.

The concept of a good life is usually assumed by philosophers to be equivalent to that of well-being, or perhaps of a morally good life, and hence has received little attention as a potentially distinct subject matter. In a series of experiments participants were presented with vignettes involving socially sanctioned wrongdoing toward outgroup members. Findings indicated that, for a large majority, judgments of bad character strongly reduce ascriptions of the good life, while having no impact at all on ascriptions of happiness or well-being. Taken together with earlier findings these results suggest that the lay concept of a good life is clearly distinct from those of happiness, well-being, or morality , likely encompassing both morality and well-being, and perhaps other values as well: whatever matters in a person’s life. Importantly, morality appears not to play a fundamental role in either happiness or well-being among the folk.

In a series of ten preregistered experiments (N=2043), we investigate the effect of outcome valence on judgments of probability, negligence, and culpability – a phenomenon sometimes labelled moral (and legal) luck. We found that harmful outcomes, when contrasted with neutral outcomes, lead to increased perceived probability of harm ex post, and consequently to increased attribution of negligence and culpability. Rather than simply postulating a hindsight bias (as is common), we employ a variety of empirical means to demonstrate that the outcome-driven asymmetry across perceived probabilities constitutes a systematic cognitive distortion. We then explore three distinct strategies to alleviate the hindsight bias and its downstream effects on mens rea and culpability ascriptions. Not all are successful, but at least some prove promising. They should, we argue, be taken into consideration in criminal jurisprudence, where distortions due to the hindsight bias are likely considerable and deeply disconcerting.