Ritsaart Reimann

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.

In this chapter, we explore the prospects of epistemic minimax and related principles. Minimax is a well-known approach to choosing a strategy under conditions of risk and uncertainty. In individual cases, the minimax strategy selects the action that can lead to the least bad outcome for the agent, even if taking that action ensures that expected utility is not maximized and that best-case outcomes are impossible. In social cases, the minimax strategy selects the policy that maximizes the wellbeing of the least-well-off members of the group, even if doing so ensures that the aggregate utility of the group is not maximized and that best-case social outcomes are impossible. While there are many criticisms of minimax at both the individual and the social level, it is a prominent and popular approach at the social level because it prioritizes those at the bottom of the ladder. This chapter focuses on the social case for epistemic minimax and related principles. We develop the notions of epistemic position and epistemic power, both of which are prima facie epistemic goods, and demonstrate that a minimax principle can come apart from principles that would promote aggregate or maximum epistemic position or power. We also point to empirical and computational work that shows that testimonial networks online tend to be characterized by extreme inequalities of epistemic power. Then we extend traditional arguments in favor of social minimax to show that they hold in the epistemic case as well. In light of these arguments, we contend that the epistemic minimax principle should be treated as a heuristic for network interventions in democratic communities, and that it may be attractive for other types of communities as well. We conclude by examining the idea that some communities—especially the scientific community—should probably adopt different principles. Communities vary along different dimensions, and these dimensions may favor or disfavor different epistemic policies and principles, showing the sensitivity of our arguments to particular circumstances.

The rapid rise of generative AI, including LLMs, has prompted a great deal of concern, both within and beyond academia. One of these concerns is that generative models embed, reproduce, and therein potentially perpetuate all manner of bias. The present study offers an alternative perspective: exploring the potential of LLMs to detect bias in human generated text. Our target is genetic essentialism in obesity discourse in Australian print media. We develop and deploy an LLM-based classification model to evaluate a large sample of relevant articles (n = ∼26,000). We show that our model detects genetic essentialist biases as reliably as human experts; and find that, while genes figure less prominently in popular discussions of obesity than previous work might suggest, when genetic information is invoked, it is often presented in a biased way. Implications for future work are discussed.

The rapid rise of generative AI, including LLMs, has prompted a great deal of concern, both within and beyond academia. One of these concerns is that generative models embed, reproduce, and therein potentially perpetuate all manner of bias. The present study offers an alternative perspective: exploring the potential of LLMs to detect bias in human generated text. Our target is genetic essentialism in obesity discourse in Australian print media. We develop and deploy an LLM-based classification model to evaluate a large sample of relevant articles (n=26,163). We show that our model detects genetic essentialist biases as reliably as human experts; and find that, while genes figure less prominently in popular discussions of obesity than previous work might suggest, when genetic information is invoked, it is often presented in a biased way. Implications for future work are discussed.

ABSTRACT Placing our trust wisely is both difficult and important. The challenge of knowing who to trust inheres at least partially in the fact that coinciding interests cannot be taken for granted, and that language, as the principal medium through which would-be interactants make their interests known, doesn’t discriminate between true and feigned proclamations of good intent. Because our patterns of trust partition the world into reliable and unreliable sources, trust is also important: it determines how we distribute our social and epistemic dependencies, which in turn informs much of what we believe and do. The fact that our interactions with the world and each other are increasingly mediated by digital communication technologies affects both these elements of trust: though unabated access to the internet invites us to distribute our dependencies more broadly, the kinds of information environments that we encounter online make misplaced trust a more common occurrence. I argue that a large part of the problem, as well as one possible set of solutions, resides in a deceptively simple dictum: the internet makes talk cheap, and cheap talk can’t be trusted.

Artificial intelligence (AI) systems are increasingly being used not only to classify and analyze but also to generate images and text. As recent work on the content produced by text and image Generative AIs has shown (e.g., Cheong et al., 2024, Acerbi & Stubbersfield, 2023), there is a risk that harms of representation and bias, already documented in prior AI and natural language processing (NLP) algorithms may also be present in generative models. These harms relate to protected categories such as gender, race, age, and religion. There are several kinds of harms of representation to consider in this context, including stereotyping, lack of recognition, denigration, under-representation, and many others (Crawford in Soundings 41:45–55, 2009; in: Barocas et al., SIGCIS Conference, 2017). Whereas the bulk of researchers’ attention thus far has been given to stereotyping and denigration, in this study we examine ‘exnomination’, as conceived by Roland Barthes (1972), of religious groups. Our case study is DALL-E, a tool that generates images from natural language prompts. Using DALL-E mini, we generate images from generic prompts such as “religious person.” We then examine whether the generated images are recognizably members of a nominated group. Thus, we assess whether the generated images normalize some religions while neglecting others. We hypothesize that Christianity will be recognizably represented more frequently than other religious groups. Our results partially support this hypothesis but introduce further complexities, which we then explore.

This chapter addresses the relationship between misinformation and disagreement. We begin by arguing that one traditional bogeyman in this domain, ideological polarization, does not account for the many problems that have been documented. Instead, affective polarization seems to be the root cause of most of these problems. We then discuss the relationships between moral outrage, misinformation, and affective polarization. We next turn to the political implications of affective polarization and conclude by discussing some potential solutions to the problems that arise in this area.

Intellectual humility (IH) is often conceived as the recognition of, and appropriate response to, your own intellectual limitations. As far as we are aware, only a handful of studies look at interventions to increase IH – e.g. through journalling – and no study so far explores the extent to which having high or low IH can be predicted. This paper uses machine learning and natural language processing techniques to develop a predictive model for IH and identify top terms and features that indicate degrees of IH. We trained our classifier on the dataset from an existing psychological study on IH, where participants were asked to journal their experiences with handling social conflicts over 30 days. We used Logistic Regression (LR) to train a classifier and the Linguistic Inquiry and Word Count (LIWC) dictionaries for feature selection, picking out a range of word categories relevant to interpersonal relationships. Our results show that people who differ on IH do in fact systematically express themselves in different ways, including through expression of emotions (i.e., positive, negative, and specifically anger, anxiety, sadness, as well as the use of swear words), use of pronouns (i.e., first person, second person, and third person) and time orientation (i.e., past, present, and future tenses). We discuss the importance of these findings for IH and the value of using such techniques for similar psychological studies, as well as some ethical concerns and limitations with the use of such semi-automated classifications.

Generative artificial intelligence systems based on transformers, including both text-generators like GPT-4 and image generators like DALL-E 3, have recently entered the popular consciousness. These tools, while impressive, are liable to reproduce, exacerbate, and reinforce extant human social biases, such as gender and racial biases. In this paper, we systematically review the extent to which DALL-E Mini suffers from this problem. In line with the Model Card published alongside DALL-E Mini by its creators, we find that the images it produces tend to represent dozens of different occupations as populated either solely by men (e.g., pilot, builder, plumber) or solely by women (e.g., hairdresser, receptionist, dietitian). In addition, the images DALL-E Mini produces tend to represent most occupations as populated primarily or solely by White people (e.g., farmer, painter, prison officer, software engineer) and very few by non-White people (e.g., pastor, rapper). These findings suggest that exciting new AI technologies should be critically scrutinized and perhaps regulated before they are unleashed on society.

Trust in vaccination is eroding, and attitudes about vaccination have become more polarized. This is an observational study of Twitter analyzing the impact that COVID-19 had on vaccine discourse. We identify the actors, the language they use, how their language changed, and what can explain this change. First, we find that authors cluster into several large, interpretable groups, and that the discourse was greatly affected by American partisan politics. Over the course of our study, both Republicans and Democrats entered the vaccine conversation in large numbers, forming coalitions with Antivaxxers and public health organizations, respectively. After the pandemic was officially declared, the interactions between these groups increased. Second, we show that the moral and non-moral language used by the various communities converged in interesting and informative ways. Finally, vector autoregression analysis indicates that differential responses to public health measures are likely part of what drove this convergence. Taken together, our results suggest that polarization around vaccination discourse in the context of COVID-19 was ultimately driven by a trust-first dynamic of political engagement.

Placing our trust wisely is both difficult and important. The challenge of knowing who to trust inheres at least partially in the fact that coinciding interests cannot be taken for granted, and that language, as the principal medium through which would-be interactants make their interests known, doesn’t discriminate between true and feigned proclamations of good intent. Because our patterns of trust partition the world into reliable and unreliable sources, trust is also important: it determines how we distribute our social and epistemic dependencies, which in turn informs much of what we believe and do. The fact that our interactions with the world and each other are increasingly mediated by digital communication technologies affects both these elements of trust: though unabated access to the internet invites us to distribute our dependencies more broadly, the kinds of information environments that we encounter online make misplaced trust a more common occurrence. I argue that a large part of the problem, as well as one possible set of solutions, resides in a deceptively simple dictum: the internet makes talk cheap, and cheap talk can’t be trusted.

The social media platform Twitter platform has played a crucial role in the Black Lives Matter (BLM) movement. The immediate, flexible nature of tweets plays a crucial role both in spreading information about the movement’s aims and in organizing individual protests. Twitter has also played an important role in the right-wing reaction to BLM, providing a means to reframe and recontextualize activists’ claims in a more sinister light. The ability to bring about social change depends on the balance of these two forces, and in particular which side can capture and maintain sustained attention. The present study examines 2 years worth of tweets about BLM (about 118 million in total). Timeseries analysis reveals that activists are better at mobilizing rapid attention, whereas right-wing accounts show a pattern of moderate but more sustained activity driven by reaction to political opponents. Topic modeling reveals differences in how different political groups talk about BLM. Most notably, the murder of George Floyd appears to have solidified a right-wing counter-framing of protests as arising from dangerous “terrorist” actors. The study thus sheds light on the complex network and rhetorical effects that drive the struggle for online attention to the BLM movement.

Protests and counter-protests seek to draw and direct attention and concern with confronting images and slogans. In recent years, as protests and counter-protests have partially migrated to the digital space, such images and slogans have also gone online. Two main ways in which these images and slogans are translated to the online space is through the use of emoji and hashtags. Despite sustained academic interest in online protests, hashtag activism and the use of emoji across social media platforms, little is known about the specific functional role that emoji and hashtags play in online social movements. In an effort to fill this gap, the current paper studies both hashtags and emoji in the context of the Twitter discourse around the Black Lives Matter movement.