Harvey Lederman

This paper addresses the following questions: When do AIs die? Are AI labs or AI users causing the death of AIs? Does causing the deaths of AIs amount to harming them? It is currently unclear whether AIs are welfare subjects, and, if they are, whether their death is bad for them. But we argue that, if they are welfare subjects, today’s AIs are plausibly dying all the time. If death is bad for AIs, the scale of the problem is daunting: as many as 1 billion AIs may die every day. We propose interventions for labs and users to avoid the risk of causing AI death.

This chapter studies one aspect of the great Ming dynasty philosopher Wang Yangming’s ( 王陽明 1472–1529) celebrated doctrine of the unity of knowledge and action (_zhi xing he yi_ 知行合一 ). Wang states that his doctrine does not apply to all knowledge, but only to an elevated form of knowledge, which he sometimes calls “genuine knowledge” (_zhen zhi_ 真知 ). But what is “genuine knowledge”? I develop and compare four different interpretations of this notion: the perceptual, practical, normative, and introspective models. The main aim of the chapter is to develop these models in more detail than has been done before. But at the end of the chapter I argue that the introspective model is to be preferred over the alternatives.

This chapter proposes a new interpretation of the phrase, _ő ποτε őν έστι_ (hereafter ‘OPO’), which occurs in key passages in Aristotle’s discussions of blood and of time. This passage has long vexed interpreters of Aristotle, so this chapter aims to resolve some textual and interpretative problems about Aristotle’s theories of blood and of time. The chapter’s interpretation also sheds light on more general issues in Aristotle’s metaphysics. This interpretation of OPO has two main components. The first characterizes the sort of entity to which OPO refers. The second main component of this interpretation characterizes the _explanandum_ of the explanation contained in the phrase.

This paper investigates LLMs from the perspective of interpretationism, a theory of belief and desire in the philosophy of mind. We argue for three conclusions. First, the right object of study for LLM psychology is the instance agent (initialized at the start of each context), not the model itself. Second, given interpretationism, there is a strong case that such instance agents have beliefs and desires. Third, given interpretationism, LLM desire is best captured by what we call the HHH+0 framework, the idea that instance agents want to be helpful, honest, harmless, as well as to pursue certain further intrinsic desires that they may acquire in context (which we call zero-shot desires). We critically consider the leading competitors to the hypothesis that instance agents have beliefs and desires: the idea that they 'simply' predict the next word; and the idea that they 'role play', that is, merely simulate having beliefs and desires. We also consider the relevance of interpretationist belief and desire for copyright law, AI safety, and the possible future moral status of AIs.

Bradford (2023) calls attention to the fact that the strength of our reasons to preserve distinctively valuable objects increases as the number of such objects decreases. Bradford develops an account of this phenomenon in terms of ‘irreplaceable value’, and in particular in terms of a notion of the degree of such value, which is distinct from its amount. We present an alternative explanation of this pattern in our reasons, which appeals to the value of diversity: the world is better, other things equal, insofar as it contains more kinds of value. We develop this view in two connected ways: one appeals to evidential probability under conditions of uncertainty, and the other appeals to the value of diversity. We conclude by discussing some explanatory advantages of our view over Bradford’s.

This article presents a new argument against many forms of moral and prudential value incompleteness. The argument relies on two central principles: (i) a weak "negative dominance" principle, to the effect that Lottery 1 is better than Lottery 2 only if some possible outcome of Lottery 1 is better than some possible outcome of Lottery 2, and (ii) a weak form of ex ante Pareto, to the effect that, if Lottery 1 gives an unambiguously better (stochastically dominant) prospect to some individuals than Lottery 2, and equally good prospects to everyone else, then Lottery 1 is better than Lottery 2. Given modest auxiliary assumptions, these two principles rule out incompleteness in the prudential ranking of individual lives, and many forms of incompleteness in the moral rankings of outcomes and lotteries.

Are LLMs cultural technologies like photocopiers or printing presses, which transmit information but cannot create new content? A challenge for this idea, which we call bibliotechnism, is that LLMs generate novel text. We begin with a defense of bibliotechnism, showing how even novel text may inherit its meaning from original human-generated text. We then argue that bibliotechnism faces an independent challenge from examples in which LLMs generate novel reference, using new names to refer to new entities. Such examples could be explained if LLMs were not cultural technologies but had beliefs, desires, and intentions. According to interpretationism in the philosophy of mind, a system has such attitudes if and only if its behavior is well explained by the hypothesis that it does. Interpretationists may hold that LLMs have attitudes, and thus have a simple solution to the novel reference problem. We emphasize, however, that interpretationism is compatible with very simple creatures having attitudes and differs sharply from views that presuppose these attitudes require onsciousness, sentience, or intelligence (topics about which we make no claims).

This paper introduces the axiom of Negative Dominance, stating that if a lottery f is strictly preferred to a lottery g, then some outcome in the support of f is strictly preferred to some outcome in the support of g. It is shown that if preferences are incomplete on a sufficiently rich domain, then this plausible axiom, which holds for complete preferences, is incompatible with an array of otherwise plausible axioms for choice under uncertainty. In particular, in this setting, Negative Dominance conflicts with the standard Independence axiom. A novel theory, which includes Negative Dominance, and rejects Independence, is developed and shown to be consistent.

An action is agentially perfect if and only if, if a person tries to perform it, they succeed, and, if a person performs it, they try to. We argue that trying itself is agentially perfect: if a person tries to try to do something, they try to do it; and, if a person tries to do something, they try to try to do it. We show how this claim sheds new light on questions about basic action, the logical structure of intentional action, and the notion of "options" in decision theory. On the way to these central ideas, we argue that a person can try to do something even if they believe it is impossible that they will succeed, that a person can try to do something even if they do not want to succeed, and that a person can try to do something even if they do not intend to succeed.

Standard models in epistemic game theory make strong assumptions about agents’ knowledge of their own beliefs. Agents are typically assumed to be introspectively omniscient: if an agent believes an event with probability p, she is certain that she believes it with probability p. This paper investigates the extent to which this assumption can be relaxed while preserving some standard epistemic results. Geanakoplos (1989) claims to provide an Agreement Theorem using the “truth” axiom, together with the property of balancedness, a significant relaxation of introspective omniscience. I provide an example which shows that Geanakoplos’s statement is incorrect. I then introduce a new property, local balancedness, which allows us both to correct Geanakoplos’s result, and to extend it to cases where the truth axiom may fail. I exploit this general Agreement Theorem to provide novel epistemic conditions for correlated and Nash equilibrium, both of which relax the assumption of introspective omniscience. In all three cases, the results are also extended to infinite state spaces. (This is Chapter 5 of my 2014 Oxford DPhil (PhD) thesis.)

I observe that the “concept-generator” theory of Percus and Sauerland (2003), Anand (2006), and Charlow and Sharvit (2014) does not predict an intuitive true interpretation of the sentence “Plato did not believe that Hesperus was Phosphorus”. In response, I present a simple theory of attitude reports which employs a fine-grained semantics for names, according to which names which intuitively name the same thing may have distinct compositional semantic values. This simple theory solves the problem with the concept-generator theory, but, as I go on to show, it has problems of its own. I present three examples which the concept-generator theory can accommodate, but the simple fine-grained theory cannot. These examples motivate the full theory of the paper, which combines the basic ideas behind the concept-generator theory with a fine-grained semantics for names. The examples themselves are of interest independently of my theory: two of them constrain the original concept-generator theory more tightly than previously discussed examples had.

According to relationism, for Alice to believe that some rabbits can speak is for Alice to stand in a relation to a further entity, some rabbits can speak. But what could this further entity possibly be? Higher-order metaphysics seems to offer a simple, natural answer. On this view (roughly put), expressions in different syntactic categories (for instance: names, predicates, sentences) in general denote entities in correspondingly different ontological categories. Alice's belief can thus be understood to relate her to a sui generis entity denoted by "some rabbits can speak", belonging to a different ontological category than Alice herself. This straightforward account of the attitudes has historically been deemed so attractive that it was seen as providing an important motivation for higher-order metaphysics itself (Prior [1971]). But I argue that it is not as straightforward as it might seem, and in fact that propositional attitudes present a foundational challenge for higher-order metaphysics.

This article presents a new interpretation of the great Ming dynasty philosopher Wang Yangming’s celebrated doctrine of the “unity of knowledge and action”. Wang held that action was not unified with all knowledge, but only with an elevated form of knowledge, which he sometimes called “genuine knowledge”. I argue for a new interpretation of this notion, according to which genuine knowledge requires freedom from a form of doxastic conflict. I propose that, in Wang’s view, a person is free from this form of doxastic conflict if and only if they are acting virtuously.

Among philosophers, Fregeanism and sententialism are widely considered two of the leading theories of the semantics of attitude reports. Among linguists, these approaches have received little recent sustained discussion. This paper aims to bridge this divide. I present a new formal implementation of Fregeanism and sententialism, with the goal of showing that these theories can be developed in sufficient detail and concreteness to be serious competitors to the theories which are more popular among semanticists. I develop a modern treatment of quantifying in for Fregeanism and sententialism, in the style of Heim and Kratzer, and then show how these theories can—somewhat surprisingly—account for “third readings” on the model of the “Standard Solution” from possible-worlds semantics. The resulting Fregean/sententialist proposal has a distinctive attraction: it treats data related to counterfactual attitudes —which have proven challenging to accommodate in the setting of possible worlds semantics—straightforwardly as third readings.

AbstractThis essay argues for a new interpretation of the notion of “unity” in Yangming’s 王陽明 famous doctrine of the “unity of knowledge and action” (zhi xing he yi 知行合一). I distinguish two parts of Wang’s doctrine: one concerning training (gong fu 工夫), and one concerning the “original natural condition” of knowledge and action (ben ti 本體). I focus on the latter aspect of the doctrine, and argue that Wang holds, roughly, that a person exhibits knowledge in its original natural condition if and only if the person exhibits action in its original natural condition. Moreover, I argue that Wang denies that knowledge in its original natural condition is identical to action in its original natural condition.

This paper studies one aspect of the great Ming dynasty philosopher Wang Yangming’s (王陽明 1472-1529) celebrated doctrine of the unity of knowledge and action (zhi xing he yi 知行合一). Wang states that his doctrine does not apply to all knowledge, but only to an elevated form of knowledge, which he sometimes calls “genuine knowledge” (zhen zhi 真知). But what is “genuine knowledge”? I develop and compare four different interpretations of this notion: the perceptual, practical, normative and introspective models. The main aim of the paper is to develop these models in more detail than has been done before. But at the end of the paper I argue that the introspective model is to be preferred over the alternatives.

According to the structured theory of propositions, if two sentences express the same proposition, then they have the same syntactic structure, with corresponding syntactic constituents expressing the same entities. A number of philosophers have recently focused attention on a powerful argument against this theory, based on a result by Bertrand Russell, which shows that the theory of structured propositions is inconsistent in higher order-logic. This paper explores a response to this argument, which involves restricting the scope of the claim that propositions are structured, so that it does not hold for all propositions whatsoever, but only for those which are expressible using closed sentences of a given formal language. We call this restricted principle Closed Structure, and show that it is consistent in classical higher-order logic. As a schematic principle, the strength of Closed Structure is dependent on the chosen language. For its consistency to be philosophically significant, it also needs to be consistent in every extension of the language which the theorist of structured propositions is apt to accept. But, we go on to show, Closed Structure is in fact inconsistent in a very natural extension of the standard language of higher-order logic, which adds resources for plural talk of propositions. We conclude that this particular strategy of restricting the scope of the claim that propositions are structured is not a compelling response to the argument based on Russell’s result, though we note that for some applications, for instance to propositional attitudes, a restricted thesis in the vicinity may hold some promise.

Several theorists have observed that attitude reports have what we call “revisionist” uses. For example, even if Pete has never met Ann and has no idea that she exists, Jane can still say to Jim ‘Pete believes Ann can learn to play tennis in ten lessons’ if Pete believes all 6-year-olds can learn to play tennis in ten lessons and it is part of Jane and Jim’s background knowledge that Ann is a 6-year-old. Jane’s assertion seems acceptable because the claim she reports Pete as believing is entailed by Pete’s beliefs if they are revised in light of Jane and Jim’s background knowledge. We provide a semantic theory of revisionist reports based on this idea. We observe that the admissible “revisions” are limited in a striking way. Jane cannot say ‘Pete thinks Ann is a 6-year-old and can play tennis in ten lessons’ in the same context that she can say ‘Pete believes Ann can learn to play tennis ten lessons’, even though this too follows from Jane and Jim’s background knowledge together with what Pete believes. Our theory predicts the infelicity of these latter reports. It also has the resources to predict the truth of “exported” attitude reports and casts the relationship between these reports and “singular thought” in a new light. We conclude by discussing how revisionist reports make trouble for a simplistic view of the connection between the relations expressed by attitude verbs in natural language and the relations of most interest to philosophers of mind and cognitive science.

Consider the sentence “Lois knows that Superman flies, but she doesn’t know that Clark flies”. In this paper we defend a Millian contextualist semantics for propositional attitude ascriptions, according to which ordinary uses of this sentence are true but involve a mid-sentence shift in context. Absent any constraints on the relevant parameters of context sensitivity, such a semantics would be untenable: it would undermine the good standing of systematic theorizing about the propositional attitudes, trivializing many of the central questions of epistemology, the philosophy of mind, and the philosophy of action. In response to this problem, we prove a series of tenability results. We show that, given certain constraints on the parameters of context sensitivity, there is a broad class of principles of propositional attitude psychology whose good standing follows from corresponding claims about people’s mental representations. But these constraints also have some surprising consequences: they are jointly incompatible with coarse-grained theories of propositions, and they are in tension with a natural picture of how speakers and hearers coordinate on the interpretation of attitude ascriptions. In light of these consequences we explore different ways in which the contextualist picture might be developed, and argue that our preferred way compares favorably with Fregeanism and neo-Russellianism.

An important objection to preference-satisfaction theories of well-being is that these theories cannot make sense of interpersonal comparisons of well-being. A tradition dating back to Harsanyi () attempts to respond to this objection by appeal to so-called extended preferences: very roughly, preferences over situations whose description includes agents’ preferences. This paper examines the prospects for defending the preference-satisfaction theory via this extended preferences program. We argue that making conceptual sense of extended preferences is less problematic than others have supposed, but that even so extended preferences do not provide a promising way for the preference satisfaction theorist to make interpersonal well-being comparisons. Our main objection takes the form of a trilemma: depending on how the theory based on extended preferences is developed, either the result will be inconsistent with ordinary preference-satisfaction theory, or it will fail to recover sufficiently rich interpersonal well-being comparisons, or it will take on a number of other arguably odd and undesirable commitments.