TikTok is a popular platform that enables users to see tailored content feeds, particularly short videos with novel content. In recent years, TikTok has been criticized at times for presenting users with overly homogenous feeds, thereby reducing the diversity of content with which each user engages. In this paper, we consider whether TikTok has an ethical obligation to employ a novelty bias in its content recommendation engine. We explicate the principal morally relevant values and interests of key stakeholders, and observe that key empirical questions must be answered before a precise recommendation can be provided. We argue that TikTok's own values and interests mean that its actions should be largely driven by the values and interests of its users and creators. Unlike some other content platforms, TikTok's ethical obligations are not at odds with the values of its users, and so whether it is obligated to include a novelty bias depends on what will actually advance its users' interests.

There is little debate about the importance of ethics in health care, and clearly defined rules, regulations, and oaths help ensure patients' trust in the care they receive. However, standards are not as well established for the data professions within health care, even though the responsibility to treat patients in an ethical way extends to the data collected about them. Increasingly, data scientists, analysts, and engineers are becoming fiduciarily responsible for patient safety, treatment, and outcomes, and will require training and tools to meet this responsibility. We developed a data ethics checklist that enables users to consider the possible ethical issues that arise from the development and use of data products. The combination of ethics training for data professionals, a data ethics checklist as part of project management, and a data ethics committee holds potential for providing a framework to initiate dialogues about data ethics and can serve as an ethical touchstone for rapid use within typical analytic workflows, and we recommend the use of this or equivalent tools in deploying new data products in hospitals.

Machine learning algorithms have gained widespread usage across a variety of domains, both in providing predictions to expert users and recommending decisions to everyday users. However, these AI systems are often black boxes, and end-users are rarely provided with an explanation. The critical need for explanation by AI systems has led to calls for algorithmic transparency, including the "right to explanation'' in the EU General Data Protection Regulation (GDPR). These initiatives presuppose that we know what constitutes a meaningful or good explanation, but there has actually been surprisingly little research on this question in the context of AI systems. In this paper, we (1) develop a generalizable framework grounded in philosophy, psychology, and interpretable machine learning to investigate and define characteristics of good explanation, and (2) conduct a large-scale lab experiment to measure the impact of different factors on people's perceptions of understanding, usage intention, and trust of AI systems. The framework and study together provide a concrete guide for managers on how to present algorithmic prediction rationales to end-users to foster trust and adoption, and elements of explanation and transparency to be considered by AI researchers and engineers in designing, developing, and deploying transparent or explainable algorithms.

Many arguments have concluded that our autonomous technologies must be intelligible, interpretable, or explainable, even if that property comes at a performance cost. In this paper, we consider the reasons why some property like these might be valuable, we conclude that there is not simply one kind of 'intelligibility', but rather different types for different individuals and uses. In particular, different interests and goals require different types of intelligibility (or explanations, or other related notion). We thus provide a typography of 'intelligibility' that distinguishes various notions, and draw methodological conclusions about how autonomous technologies should be designed and deployed in different ways, depending on whose intelligibility is required.

Trust is one of the most critical relations in our human lives, whether trust in one another, trust in the artifacts that we use everyday, or trust of an AI system. Even a cursory examination of the literatures in human-computer interaction, human-robot interaction, and numerous other disciplines reveals a deep, persistent concern with the nature of trust in AI, and the conditions under which it can be generated, reduced, repaired, or influenced. At a high level, we often understand trust as a relation in which the trustor makes oneself vulnerable based on positive expectations about the behavior or intentions of the trustee [1]. For example, when I trust my car to start in the morning, I make myself vulnerable (e.g., I risk that I will be late to work if it does not start) because I have the positive expectation that it actually will start. This high-level characterization is relatively unhelpful, however, particularly given the wide range of disciplines that have examined the relation of trust, ranging from organizational behavior to game theory to ethics to cognitive science. The picture that emerges from, for example, social psychology (i.e., two distinct kinds of trust depending on whether one knows the trustee's behaviors or intentions/ values) appears to be quite different from the one that emerges from moral philosophy (i.e., a single, highly-moralized notion), even though both are consistent with this high-level characterization. This talk first introduces that diversity of types of 'trust', but then argues that we can make progress towards a unified characterization by focusing on the function of trust. That is, we should ask why care whether we can trust our artifacts, AI, or fellow humans, as that can help to illuminate features of trust that are shared across domains, trustors, and trustees. I contend that one reason to desire trust is an "almost-necessary" condition on ethical action: namely, that the user has a reasonable belief that the system (whether human or machine) will behave approximately as intended. This condition is obviously not sufficient for ethical use, nor is it strictly necessary since the best available option might nonetheless be one for which the user lacks appropriate reasonable beliefs. Nonetheless, it provides a reasonable starting point for an analysis of 'trust'. More precisely, I propose that this condition indicates a role for trust as providing precisely those reasonable beliefs, at least when we have appropriately grounded trust. That is, we can understand 'appropriate trust' as obtaining when the trustor has justified beliefs that the trustee has suitable dispositions. As there is variation in the trustor's goals and values, and also the openness of the context of use, then different specific versions of 'appropriate trust' result as those variations lead to different types of focal dispositions, specific dispositions, or observability of dispositions, respectively. For example, in an open context (i.e., one where the possibilities cannot be exhaustively enumerated), the trustee's full dispositions will not be directly observable, but rather must be inferred from observations. This framework provides a unification of the different theories of 'trust' developed in different disciplines. Moreover, it provides clarity about one key function of trust, and thereby helps us to understand the value of (appropriate) trust. We need to trust our AI systems because that is a precondition for the ethical, responsible use of them.

Autonomous vehicles are regularly touted as holding the potential to provide significant benefits for diverse populations. There are significant technological barriers to be overcome, but as those are solved, autonomous vehicles are expected to reduce fatalities; decrease emissions and pollutants; provide new options to mobility-challenged individuals; enable people to use their time more productively; and so much more. In this paper, we argue that these high expectations for autonomous vehicles almost certainly cannot be fully realized. More specifically, the proposed benefits divide into two high-level groups, centered around efficiency and safety improvements, and increases in people's agency and autonomy. The first group of benefits is almost always framed in terms of rates: fatality rates, traffic flow per mile, and so forth. However, we arguably care about the absolute numbers for these measures, not the rates; number of fatalities is the key metric, not fatality rate per vehicle mile traveled. Hence, these potential benefits will be reduced, perhaps to non-existence, if autonomous vehicles lead to increases in vehicular usage. But that is exactly the result that we should expect if the second group of benefits is realized: if people's agency and autonomy is increased, then they will use vehicles more. There is an inevitable tension between the benefits that are proposed for autonomous vehicles, such that we cannot fully have all of them at once. We close by pointing towards other types of AI technologies where we should expect to find similar types of necessary and inevitable tradeoffs between classes of benefits.

Over recent decades, technological development has been accompanied by the proposal of new rights by various groups and individuals: the right to public anonymity, the right to be forgotten, and the right to disconnect, for example. Although there is widespread acknowledgment of the motivation behind these proposed rights, there is little agreement about their actual normative status. One potential challenge is that the claims only arise in contingent social-technical contexts, which may affect how we conceive of them ethically (albeit, not necessarily in terms of policy). What sort of morally legitimate rights claims depend on such contingencies? Our paper investigates the grounds on which such proposals might be considered "actual" rights. The full paper can be found at http://www.andrew.cmu.edu/user/cgparker/Parker_Danks_RevealedRights.pdf. We propose the notion of a revealed right, a right that only imposes duties -- and thus is only meaningfully revealed -- in certain technological contexts. Our framework is based on an interest theory approach to rights, which understands rights in terms of a justificatory role: morally important aspects of a person's well-being (interests) ground rights, which then justify holding someone to a duty that promotes or protects that interest. Our framework uses this approach to interpret the conflicts that lead to revealed rights in terms of how technological developments cause shifts in the balance of power to promote particular interests. Different parties can have competing or conflicting interests. It is also generally accepted that some interests are more normatively important than others (even if only within a particular framework). We can refer to this difference in importance by saying that the former interest has less "moral weight" than the latter interest (in that context). The moral weight of an interest is connected to its contribution to the interest-holder's overall well-being, and thereby determines the strength of the reason that a corresponding right provides to justify a duty. Improved technology can offer resources that grant one party increased causal power to realize its interests to the detriment of another's capacity to do so, even while the relative moral weight of their interests remain the same. Such changes in circumstance can make the importance of protecting a particular interest newly salient. If that interest's moral weight justifies establishing a duty to protect it, thereby limiting the threat posed by the new socio-technical context, then a right is revealed. Revealed rights justify realignment between the moral weight and causal power orderings so that people with weightier interests have greater power to protect those interests. In the extended paper, we show how this account can be applied to the interpretation of two recently proposed "rights": the right to be forgotten, and the right to disconnect. Since we are focused on making sense of revealed rights, not any particular substantive theory of interests or well-being, the characterization of 'weights' is a free parameter in this account. Our framework alone cannot provide means to resolve the question of whether specific rights exist, but it can be used to identify empirical questions that need to be answered to decide the existence or non-existence of such rights. The emergence of a revealed right depends on a number of factors, including: whether the plausible uses of the technology could potentially impede another's well-being or interests; whether the technology is sufficiently common to have a wider, social impact; and whether the technology has actually changed the balance of power sufficiently to yield a frequent possibility for misalignment between causal power and moral weight. This approach confronts the question of how, in principle, such rights could be justified, without requiring specific commitments on the ontology of rights. Our account explains why the rhetoric of "new rights" is both accurate (since the rights were not previously recognized) and inaccurate (since the rights were present all along, but without corresponding duties). Further, it explains the rights without grounding their normative status in considerations related to right-holders' capacities to rationally waive or assert claims. This is especially important given that many of the relevant disruptive technological developments pose challenges to understanding by affected parties for the same reasons they pose threats to those parties' well-being. In the course of our discussion, we confront a number of potential objections to the account. We argue that our framework's ability to accommodate highly specific or derivative-seeming rights is un-problematic. We also head off worries that our use of interest theory makes the account likely to recognize absurd rights claims.

Artificial Intelligence and robotics are rapidly moving into healthcare, playing key roles in specific medical functions, including diagnosis and clinical treatment. Much of the focus in the technology development has been on human-machine interactions, leading to a host of related technology-centric questions. In this paper, we focus instead on the impact of these technologies on human-human interactions and relationships within the healthcare domain. In particular, we argue that trust plays a central role for relationships in the healthcare domain, and the introduction of healthcare AI can potentially have significant impacts on those relations of trust. We contend that healthcare AI systems ought to be treated as assistive technologies that go beyond the usual functions of medical devices. As a result, we need to rethink regulation of healthcare AI systems to ensure they advance relevant values. We propose three distinct guidelines that can be universalized across federal regulatory boards to ensure that patient-doctor trust is not detrimentally affected by the deployment and widespread adoption of healthcare AI technologies.

The widespread deployment and testing of autonomous vehicles in real-world environments raises key questions about how such systems should be regulated. Much of the current debate presupposes that the regulatory system we currently use for regular vehicles is also appropriate for semi- and fully-autonomous ones. In opposition, we first argue that there are serious challenges to regulating autonomous vehicles using current approaches, due to the nature of both autonomous capabilities (and their connections to operational domains), and also the systems' tasks and surrounding uncertainties. Instead, we argue that vehicles with autonomous capabilities are similar in key respects to drugs and other medical inter-ventions. Thus, we propose (on a "first principles" basis) a dynamic regulatory system with staged approvals and monitoring, analogous to the system used by the U.S. Food & Drug Administration. We provide details about the operation of such a potential system, and conclude by characterizing its benefits, costs, and plausibility.

Algorithms play a key role in the functioning of autonomous systems, and so concerns have periodically been raised about the possibility of algorithmic bias. However, debates in this area have been hampered by different meanings and uses of the term, "bias." It is sometimes used as a purely descriptive term, sometimes as a pejorative term, and such variations can promote confusion and hamper discussions about when and how to respond to algorithmic bias. In this paper, we first provide a taxonomy of different types and sources of algorithmic bias, with a focus on their different impacts on the proper functioning of autonomous systems. We then use this taxonomy to distinguish between algorithmic biases that are neutral or unobjectionable, and those that are problematic in some way and require a response. In some cases, there are technological or algorithmic adjustments that developers can use to compensate for problematic bias. In other cases, however, responses require adjustments by the agent, whether human or autonomous system, who uses the results of the algorithm. There is no "one size fits all" solution to algorithmic bias.

Causal structure learning from time series data is a major scientific challenge. Existing algorithms assume that measurements occur sufficiently quickly; more precisely, they assume that the system and measurement timescales are approximately equal. In many scientific domains, however, measurements occur at a significantly slower rate than the underlying system changes. Moreover, the size of the mismatch between timescales is often unknown. This paper provides three distinct causal structure learning algorithms, all of which discover all dynamic graphs that could explain the observed measurement data as arising from undersampling at some rate. That is, these algorithms all learn causal structure without assuming any particular relation between the measurement and system timescales; they are thus rate-agnostic. We apply these algorithms to data from simulations. The results provide insight into the challenge of undersampling.

Structure learning algorithms for graphical models have focused almost exclusively on stable environments in which the underlying generative process does not change; that is, they assume that the generating model is globally stationary. In real-world environments, however, such changes often occur without warning or signal. Real-world data often come from generating models that are only locally stationary. In this paper, we present LoSST, a novel, heuristic structure learning algorithm that tracks changes in graphical model structure or parameters in a dynamic, real-time manner. We show by simulation that the algorithm performs comparably to batch-mode learning when the generating graphical structure is globally stationary, and significantly better when it is only locally stationary.

In many domains, data are distributed among datasets that share only some variables; other recorded variables may occur in only one dataset. There are several asymptotically correct, informative algorithms that search for causal information given a single dataset, even with missing values and hidden variables. There are, however, no such reliable procedures for distributed data with overlapping variables, and only a single heuristic procedure (Structural EM). This paper describes an asymptotically correct procedure, ION, that provides all the information about structure obtainable from the marginal independence relations. Using simulated and real data, the accuracy of ION is compared with that of Structural EM, and with inference on complete, unified data.

theories of human causal focus primarily on long-run predictions: learning and Current psychological two by judgment estimating parameters of a causal Bayes nets (though for different parameterizations), and a third through structural learning. This short-run behavior by examining paper dynamical versions of these three theories, and comparing their predictions to a real-world dataset. focuses on people's