Brett Karlan

I offer an epistemic argument against the dominance of empiricism and empiricist-inspired methods in contemporary machine learning (ML) research. I first establish, as many ML researchers and philosophers of ML claim, that standard methods for constructing deep learning networks are best thought of as a kind of empiricism about cognitive architecture. I then argue that, even given the resounding success of contemporary ML models, there are few (if any) strong reasons to interpret their success as ruling out competing nativist approaches to cognitive architecture, especially given the extreme resource capture that empiricist methods have undergone in ML research. If we care about research goals like modeling human cognition or building genuinely rational machines, I argue we have good reason to support a robust competing artificial intelligence (AI) nativist tradition. I show how this argument applies to two recent defenses of AI empiricism in the philosophical literature: Buckner (2023)’s book-length defense of moderate AI empiricism, and Long (2024)’s dilemma argument against AI nativism.

This chapter explores some conceptual connections between human-centered artificial intelligence (HCAI) research and empathic AI. First, I argue that HCAI is best understood as a framework that centers human well-being in AI development and deployment. Second, I argue a fruitful way to make the framework more precise is by focusing on human practical capacities. I use empathy as a case study, examining recent positive accounts of empathetic AI from the perspective of a well-being-focused HCAI. Finally, I note how this approach relates to recent calls for a sociotechnical perspective on AI.

An increasing number of complex and important decisions are now being made with the aid of opaque algorithms. This has led to calls from both theorists and legislators for the implementation of a right to an explanation for algorithmic decisions. In this paper, we argue that, in most cases and for most kinds of explanations, there is no such right. After differentiating a number of different things that might be meant by a ‘right to an explanation,’ we argue that, for the most plausible and popular versions (a right to an explanation for a specific bureaucratic decision, a right to an explanation grounded in public reason considerations, or a right to a higher-level explanation of an entire system of automated decision-making), such a right is either superfluous, impossible to obtain, or not the best way to secure the relevant normative goods. We also argue that proponents of a right to an explanation carve off certain kinds of automated decisions as requiring more justification than others in similar areas of social (and natural) science policy, a demarcation we argue is unjustified. While there will be some cases where an explanation is the only thing that can secure an important normative good (e.g. when an explanation is the only thing that would catch an unjustified individual decision), we argue these cases are too rare and scattered to ground something as weighty as a right to an explanation.

Robin McKenna’s excellent Non-Ideal Epistemology is, among other things, a testament to restraint. McKenna does not want to unnecessarily inflame tensions between ideal and non-ideal theorists in epistemology. Often ideal and non-ideal projects are aimed at different target domains and not in tension with one another (though not always; e.g. McKenna 2023, ch. 6, especially pp. 112-21). In this commentary, I will have much less tact. I sketch a route by which the non-ideal epistemologist might become more belligerent towards their ideal counterparts. I do this by focusing on an area that McKenna mostly sets aside: non-ideal individual epistemology.

I identify an undertheorized problem with decisions we make with the aid of algorithms: the problem of inauthenticity. When we make decisions with the aid of algorithms, we can make ones that go against our commitments and values in a normatively important way. In this paper, I present a framework for algorithm-aided decision-making that can lead to inauthenticity. I then construct a taxonomy of the features of the decision environment that make such outcomes likely, and I discuss three possible solutions to the problem.

Practical quantum computing devices and their applications to AI in particular are presently mostly speculative. Nevertheless, questions about whether this future technology, if achieved, presents any special ethical issues are beginning to take shape. As with any novel technology, one can be reasonably confident that the challenges presented by "quantum AI" will be a mixture of something new and something old. Other commentators (Sevilla & Moreno 2019), have emphasized continuity, arguing that quantum computing does not substantially affect approaches to value alignment methods for AI, although they allow that further questions arise concerning governance and verification of quantum AI applications. In this brief paper, we turn our attention to the problem of identifying as-yet-unknown discontinuities that might result from quantum AI applications. Wise development, introduction, and use of any new technology depends on successfully anticipating new modes of failure for that technology. This requires rigorous efforts to break systems in protected sandboxes, and it must be conducted at all stages of technology design, development, and deployment. Such testing must also be informed by technical expertise but cannot be left solely to experts in the technology because of the history of failures to predict how non-experts will use or adapt to new technologies. This interplay between experts and non-experts may be particularly acute for quantum AI because quantum mechanics is notoriously difficult to understand. (As Richard Feynman quipped, "Anyone who claims to understand quantum mechanics is either lying or crazy.") We will discuss the extent to which the difficulties in understanding the physics underlying quantum computing challenges attempts to anticipate new failure modes that might be introduced in AI applications intended for unsupervised operation in the public sphere.

In domains as disparate as playing Go and predicting the structure of proteins, artificial intelligence (AI) technologies have begun to perform at levels beyond which any humans can achieve. Does this fact represent something lamentable? Does superhuman AI performance somehow undermine the value of human achievements in these areas? Go grandmaster Lee Sedol suggested as much when he announced his retirement from professional Go, blaming the advances of Go-playing programs like AlphaGo for sapping his will to play the game at a high level. In this paper, I attempt to make sense of Sedol’s lament. I consider a number of ways that the existence of superhuman-performing AI technologies could undermine the value of human achievements. I argue there is very little in the nature of the technology itself that warrants such despair. (Compare: does the existence of a fighter jet undermine the value of being the fastest human sprinter?) But I also argue there are several more localized domains where these technologies threaten to displace human beings from being able to achieve valuable things at all. This is a particular worry for those in unequal societies, I argue, given the difficulty of many achievements and the corresponding amount of resources needed to achieve great things.

We argue that the concept of practical wisdom is particularly useful for organizing, understanding, and improving human-machine interactions. We consider the relationship between philosophical analysis of wisdom and psychological research into the development of wisdom. We adopt a practical orientation that suggests a conceptual engineering approach is needed, where philosophical work involves refinement of the concept in response to contributions by engineers and behavioral scientists. The former are tasked with encoding as much wise design as possible into machines themselves, as well as providing sandboxes or workspaces to help various stakeholders build practical wisdom in systems that are sufficiently realistic to aid transferring skills learned to real-world use. The latter are needed for the design of exercises and methods of evaluation within these workspaces, as well as ways of empirically assessing the transfer of wisdom from workspace to world. Systematic interaction between these three disciplines (and others) is the best approach to engineering wisdom for the machine age.

Implicit attitudes are mental states posited by psychologists to explain behaviors including implicit racial and gender bias. In this paper I investigate the belief view of the implicit attitudes, on which implicit attitudes are a kind of implicit belief. In particular, I focus on why implicit attitudes, if they are beliefs, are often resistant to updating in light of new evidence. I argue that extant versions of the belief view do not give a satisfactory account of this phenomenon. This is because proponents of the belief view have focused on overly narrow explanations of recalcitrance in terms of belief storage. Expanding the focus of the belief view to the kinds of irrational and arational transitions between beliefs and other mental states provides compelling (if preliminary) explanations of recalcitrance.

Which rules should guide our reasoning? Human reasoners often use reasoning shortcuts, called heuristics, which function well in some contexts but lack the universality of reasoning rules like deductive implication or inference to the best explanation. Does it follow that human reasoning is hopelessly irrational? I argue: no. Heuristic reasoning often represents human reasoners reaching a local rational maximum, reasoning more accurately than if they try to implement more “ideal” rules of reasoning. I argue this is a genuine rational achievement. Our ideal rational advisors would advise us to reason with heuristic rules, not more complicated ideal rules. I argue we do not need a radical new account of epistemic norms to make sense of the success of heuristic reasoning.