Vincent Conitzer

In metaphysics, there are a number of distinct but related questions about the existence of “further facts”—facts that are contingent relative to the physical structure of the universe. These include further facts about qualia, personal identity, and time. In this article I provide a sequence of examples involving computer simulations, ranging from one in which the protagonist can clearly conclude such further facts exist to one that describes our own condition. This raises the question of where along the sequence the protagonist stops being able to soundly conclude that further facts exist.

A-theorists and B-theorists debate whether the “Now” is metaphysically distinguished from other time slices. Analogously, one may ask whether the “I” is metaphysically distinguished from other perspectives. Few philosophers would answer the second question in the affirmative. An exception is Caspar Hare, who has devoted two papers and a book to arguing for such a positive answer. In this paper, I argue that those who answer the first question in the affirmative—A-theorists— should also answer the second question in the affirmative. This is because key arguments in favor of the A-theory are more effective as arguments in favor of the resulting combined position, and key arguments against the A-theory are ineffective against the combined position.

How should we update de dicto beliefs in the face of de se evidence? The Sleeping Beauty problem divides philosophers into two camps, halfers and thirders. But there is some disagreement among halfers about how their position should generalize to other examples. A full generalization is not always given; one notable exception is the Halfer Rule, under which the agent updates her uncentered beliefs based on only the uncentered part of her evidence. In this brief article, I provide a simple example for which the Halfer Rule prescribes credences that, I argue, cannot be reasonably held by anyone. In particular, these credences constitute an egregious violation of the Reflection Principle. I then discuss the consequences for halfing in general.

When two people need a kidney transplant, but only one kidney is available, we need to decide who gets it. If one of the potential recipients needs the kidney because of their own voluntary behavior, but the other is not at all responsible for needing a kidney, then we need to decide whether this fault should be a consideration in favor of the other patient getting the kidney. While there has been considerable philosophical debate on this issue, there is far less research into the views of the public. To explore opinions on these issues, we first asked survey participants to ascribe or deny responsibility for the risky behavior for kidney disease, and for being deprived of a kidney in cases of drinking alcohol, drug abuse, smoking and unhealthy eating when the patient did or did not stop the risky behavior after being diagnosed with kidney disease. Next, we asked participants who should get the kidney when the patient who engaged in risky behavior did or did not know, or have easy access to, the information that the behavior was risky. We found that participants generally ascribed responsibility on the basis of knowledge but allocated the kidney on the basis of access to information. These findings have important implications for moral theories as well as medical policies.

Background In the early stages of the COVID-19 pandemic, many health systems, including those in the UK, developed triage guidelines to manage severe shortages of ventilators. At present, there is an insufficient understanding of how the public views these guidelines, and little evidence on which features of a patient the public believe should and should not be considered in ventilator triage. Methods Two surveys were conducted with representative UK samples. In the first survey, 525 participants were asked in an open-ended format to provide features they thought should and should not be considered in allocating ventilators for COVID-19 patients when not enough ventilators are available. In the second survey, 505 participants were presented with 30 features identified from the first study, and were asked if these features should count in favour of a patient with the feature getting a ventilator, count against the patient, or neither. Statistical tests were conducted to determine if a feature was generally considered by participants as morally relevant and whether its mean was non-neutral. Results In Survey 1, the features of a patient most frequently cited as being morally relevant to determining who would receive access to ventilators were age, general health, prospect of recovery, having dependents, and the severity of COVID symptoms. The features most frequently cited as being morally irrelevant to determining who would receive access to ventilators are race, gender, economic status, religion, social status, age, sexual orientation, and career. In Survey 2, the top three features that participants thought should count in favour of receiving a ventilator were pregnancy, having a chance of dying soon, and having waited for a long time. The top three features that participants thought should count against a patient receiving a ventilator were having committed violent crimes in the past, having unnecessarily engaged in activities with a high risk of COVID-19 infection, and a low chance of survival. Conclusions The public generally agreed with existing UK guidelines that allocate ventilators according to medical benefits and that aim to avoid discrimination based on demographic features such as race and gender. However, many participants expressed potentially non-utilitarian concerns, such as inclining to deprioritise ventilator allocation to those who had a criminal history or who contracted the virus by needlessly engaging in high-risk activities.

Significant controversy remains about what constitute correct self-locating beliefs in scenarios such as the Sleeping Beauty problem, with proponents on both the “halfer” and “thirder” sides. To attempt to settle the issue, one natural approach consists in creating decision variants of the problem, determining what actions the various candidate beliefs prescribe, and assessing whether these actions are reasonable when we step back. Dutch book arguments are a special case of this approach, but other Sleeping Beauty games have also been constructed to make similar points. Building on a recent article :491–508, 2013), I show that in general we should be wary of such arguments, because unintuitive actions may result for reasons that are unrelated to the beliefs. On the other hand, I show that, when we restrict our attention to additive games, then a thirder will necessarily maximize her ex ante expected payout, but a halfer in some cases will not. I conclude that this does not necessarily settle the issue and speculate about what might.

In the context of the Sleeping Beauty problem, it has been argued that so-called “halfers” can avoid Dutch book arguments by adopting evidential decision theory. I introduce a Dutch book for a variant of the Sleeping Beauty problem and argue that evidential decision theorists fall prey to it, whether they are halfers or thirders. The argument crucially requires that an action can provide evidence for what the agent would do not only at other decision points where she has exactly the same information, but also at decision points where she has different but “symmetric” information.

In metaphysics, there are a number of distinct but related questions about the existence of “further facts”—facts that are contingent relative to the physical structure of the universe. These include further facts about qualia, personal identity, and time. In this article I provide a sequence of examples involving computer simulations, ranging from one in which the protagonist can clearly conclude such further facts exist to one that describes our own condition. This raises the question of where along the sequence the protagonist stops being able to soundly conclude that further facts exist.

Foundation models such as GPT-4 are fine-tuned to avoid unsafe or otherwise problematic behavior, such as helping to commit crimes or producing racist text. One approach to fine-tuning, called reinforcement learning from human feedback, learns from humans' expressed preferences over multiple outputs. Another approach is constitutional AI, in which the input from humans is a list of high-level principles. But how do we deal with potentially diverging input from humans? How can we aggregate the input into consistent data about "collective" preferences or otherwise use it to make collective choices about model behavior? In this paper, we argue that the field of social choice is well positioned to address these questions, and we discuss ways forward for this agenda, drawing on discussions in a recent workshop on Social Choice for AI Ethics and Safety held in Berkeley, CA, USA in December 2023.

It is sometimes the case that one solution concept in game theory is equivalent to applying another solution concept to a modified version of the game. In such cases, does it make sense to study the former separately, or should we entirely subordinate it to the latter? The answer probably depends on the particular circumstances, and indeed the literature takes different approaches in different cases. In this article, I consider the specific example of Stackelberg mixed strategies. I argue that, even though a Stackelberg mixed strategy can also be seen as a subgame perfect Nash equilibrium of a corresponding extensive-form game, there remains significant value in studying it separately. The analysis of this special case may have implications for other solution concepts.

It is our great pleasure to welcome you to the proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society - AIES 2019. The second edition of this conference was co-located with AAAI-19 on January 27-28, 2019 in Honolulu, Hawaii, USA. Concerns about the impact of AI on society have continued to grow in the year since AAAI and ACM joined to create the first Conference on AI, Ethics and Society. In the vision of this joint effort, it is only through multidisciplinary engagement and scholarship that we can hope to develop good responses to the challenge of ensuring that AI develops in a way that is safe and beneficial for everyone. This year's conference contributed to building a community of shared concepts and concerns. With submissions ranging across a diverse array of fields - computer science, philosophy, economics, sociology, psychology, and law - we enjoyed two days of engaging contributions that provided new paths for research. A panel and four invited talks set the stage. The opening panel on Responsible Artificial Intelligence: Aligning Technology, Engineering and Ethics (with Yolanda Gil, Huw Price, Francesca Rossi, and Dekai Wu, moderated by Peter Hershock) was held at the University of Hawaii - Manoa in the evening preceding the main conference. The main program led off with Ryan Calo's invited talk titled How We Talk About AI (And Why It Matters). He highlighted the dangers of how we talk about AI: terms like "killer robot" and "arms races" may increase rather than decrease international cooperation, and Calo argued that focusing on "ethics" may reduce our attention to formal law and governance. In her talk titled Guiding and Implementing AI, Susan Athey (Stanford) drew on the theories of causal inference and incentive design to generate insights and research paths for the design of AI systems and organizations that integrate AI and human decision makers. Anca Dragan (Berkeley) urged a reconceptualization of the challenge of building robots that achieve human objectives within constraints as a human-robot collaboration problem with the aim of both coordinating behaviour and jointly discovering the goals of collaboration, in her talk titled Specifying AI Objectives as a Human-AI Collaboration Problem. Finally, in his talk titled The Value of Trustworthy AI, David Danks (Carnegie Mellon) explored the philosophical and psychological conditions for 'trustworthy' AI systems, and the importance of this complex notion of 'trustworthiness' for informed AI policy and public action. The program of the conference included peer-reviewed paper presentations, with 35 papers presented as posters with an associated brief spotlight presentation, and another 34 as longer oral presentations. (The latter also had associated posters, if the authors so chose.) The program also included student posters. Conference session themes included algorithmic fairness, norms and explanations, artificial agency, autonomy and lethality, rights and principles, social science models for AI, measurement and justice, AI for social good, and human-machine interaction.

Newcomb’s problem has spawned a debate about which variant of expected utility maximisation should guide rational choice. In this paper, we provide a new argument against what is probably the most popular variant: causal decision theory. In particular, we provide two scenarios in which CDT voluntarily loses money. In the first, an agent faces a single choice and following CDT’s recommendation yields a loss of money in expectation. The second scenario extends the first to a diachronic Dutch book against CDT.