Simon Goldstein

Since the release of ChatGPT, there has been a lot of debate about whether AI systems pose an existential risk to humanity. This paper develops a general framework for thinking about the existential risk of AI systems. We analyze a two-premise argument that AI systems pose a threat to humanity. Premise one: AI systems will become extremely powerful. Premise two: if AI systems become extremely powerful, they will destroy humanity. We use these two premises to construct a taxonomy of ‘survival stories’, in which humanity survives into the far future. In each survival story, one of the two premises fails. Either scientific barriers prevent AI systems from becoming extremely powerful; or humanity bans research into AI systems, thereby preventing them from becoming extremely powerful; or extremely powerful AI systems do not destroy humanity, because their goals prevent them from doing so; or extremely powerful AI systems do not destroy humanity, because we can reliably detect and disable systems that have the goal of doing so. We argue that different survival stories face different challenges. We also argue that different survival stories motivate different responses to the threats from AI. Finally, we use our taxonomy to produce rough estimates of ‘P(doom)’, the probability that humanity will be destroyed by AI.

This paper offers the first careful analysis of the possibility that AI and humanity will go to war. The paper focuses on the case of artificial general intelligence, AI with broadly human capabilities. The paper uses a bargaining model of war to apply standard causes of war to the special case of AI/human conflict. The paper argues that information failures and commitment problems are especially likely in AI/human conflict. Information failures would be driven by the difficulty of measuring AI capabilities, by the uninterpretability of AI systems, and by differences in how AIs and humans analyze information. Commitment problems would make it difficult for AIs and humans to strike credible bargains. Commitment problems could arise from power shifts, rapid and discontinuous increases in AI capabilities. Commitment problems could also arise from missing focal points, where AIs and humans fail to effectively coordinate on policies to limit war. In the face of this heightened chance of war, the paper proposes several interventions. War can be made less likely by improving the measurement of AI capabilities, capping improvements in AI capabilities, designing AI systems to be similar to humans, and by allowing AI systems to participate in democratic political institutions.

LLMs have dramatically improved in capabilities in recent years. This raises the question of whether LLMs could become genuine agents with beliefs and desires. This paper demonstrates an in principle limit to LLM agency, based on their architecture. LLMs are next word predictors: given a string of text, they calculate the probability that various words can come next. LLMs produce outputs that reflect these probabilities. I show that next word predictors are exploitable. If LLMs are prompted to make probabilistic predictions about the world, these predictions are guaranteed to be incoherent, and so Dutch bookable. If LLMs are prompted to make choices over actions, their preferences are guaranteed to be intransitive, and so money pumpable. In short, the problem is that selecting an action based on its potential value is structurally different then selecting the description of an action that is most likely given a prompt: probability cannot be forced into the shape of expected value. The in principle exploitability of LLMs raises doubts about how agential they can become. This exploitability also offers an opportunity for humanity to safely control such AI systems.

AI companies are racing to create artificial general intelligence, or “AGI.” If they succeed, the result will be human-level AI systems that can independently pursue high-level goals by formulating and executing long-term plans in the real world. Leading AI researchers agree that some of these systems will likely be “misaligned”–pursuing goals that humans do not desire. This goal mismatch will put misaligned AIs and humans into strategic competition with one another. As with present-day strategic competition between nations with incompatible goals, the result could be violent and catastrophic conflict. Existing legal institutions are unprepared for the AGI world. New foundations for AGI governance are needed, and the time to begin laying them is now, before the critical moment arrives. This Article begins to lay those new legal foundations. It is the first to think systematically about the dynamics of strategic competition between humans and misaligned AGI. The Article begins by showing, using formal game-theoretic models, that, by default, humans and AIs will be trapped in a prisoner’s dilemma. Both parties’ dominant strategy will be to permanently disempower or destroy the other, even though the costs of such conflict would be high. The Article then argues that a surprising legal intervention could transform the game theoretic equilibrium and avoid conflict: AI rights. Not just any AI rights would promote human safety. Granting AIs the right not to be needlessly harmed–as humans have granted to certain non-human animals–would, for example, have little effect. Instead, to promote human safety, AIs should be given those basic private law rights–to make contracts, hold property, and bring tort claims–that law already extends to non-human corporations. Granting AIs these economic rights would enable long-run, small-scale, mutually-beneficial transactions between humans and AIs. This would, we show, facilitate a peaceful strategic equilibrium between humans and AIs for the same reasons economic interdependence tends to promote peace in international relations. Namely, the gains from trade far exceed those from war. Throughout, we argue that human safety, rather than AI welfare, provides the right framework for developing AI rights. This Article explores both the promise and the limits of AI rights as a legal tool for promoting human safety in an AGI world.

It is generally assumed that existing artificial systems are not phenomenally conscious, and that the construction of phenomenally conscious artificial systems would require significant technological progress if it is possible at all. We challenge this assumption by arguing that if Global Workspace Theory (GWT) — a leading scientific theory of phenomenal consciousness — is correct, then instances of one widely implemented AI architecture, the artificial language agent, might easily be made phenomenally conscious if they are not already. Along the way, we articulate an explicit methodology for thinking about how to apply scientific theories of consciousness to artificial systems and employ this methodology to arrive at a set of necessary and sufficient conditions for phenomenal consciousness according to GWT.

This paper examines the question of whether Large Language Models (LLMs) like ChatGPT possess minds, focusing specifically on whether they have a genuine folk psychology encompassing beliefs, desires, and intentions. We approach this question by investigating two key aspects: internal representations and dispositions to act. First, we survey various philosophical theories of representation, including informational, causal, structural, and teleosemantic accounts, arguing that LLMs satisfy key conditions proposed by each. We draw on recent interpretability research in machine learning to support these claims. Second, we explore whether LLMs exhibit robust dispositions to perform actions, a necessary component of folk psychology. We consider two prominent philosophical traditions, interpretationism and representationalism, to assess LLM action dispositions. While we find evidence suggesting LLMs may satisfy some criteria for having a mind, particularly in game-theoretic environments, we conclude that the data remains inconclusive. Additionally, we reply to several skeptical challenges to LLM folk psychology, including issues of sensory grounding, the "stochastic parrots" argument, and concerns about memorization. Our paper has three main upshots. First, LLMs do have robust internal representations. Second, there is an open question to answer about whether LLMs have robust action dispositions. Third, existing skeptical challenges to LLM representation do not survive philosophical scrutiny.

We propose developing AIs whose only final goal is being shut down. We argue that this approach to AI safety has three benefits: (i) it could potentially be implemented in reinforcement learning, (ii) it avoids some dangerous instrumental convergence dynamics, and (iii) it creates trip wires for monitoring dangerous capabilities. We also argue that the proposal can overcome a key challenge raised by Soares et al. (2015), that shutdown-seeking AIs will manipulate humans into shutting them down. We conclude by comparing our approach with Soares et al.'s corrigibility framework.

This paper argues that a range of current AI systems have learned how to deceive humans. We define deception as the systematic inducement of false beliefs in the pursuit of some outcome other than the truth. We first survey empirical examples of AI deception, discussing both special-use AI systems (including Meta's CICERO) built for specific competitive situations, and general-purpose AI systems (such as large language models). Next, we detail several risks from AI deception, such as fraud, election tampering, and losing control of AI systems. Finally, we outline several potential solutions to the problems posed by AI deception: first, regulatory frameworks should subject AI systems that are capable of deception to robust risk-assessment requirements; second, policymakers should implement bot-or-not laws; and finally, policymakers should prioritize the funding of relevant research, including tools to detect AI deception and to make AI systems less deceptive. Policymakers, researchers, and the broader public should work proactively to prevent AI deception from destabilizing the shared foundations of our society.

A mental state is luminous if, whenever an agent is in that state, they are in a position to know that they are. Following Timothy Williamson's Knowledge and Its Limits, a wave of recent work has explored whether there are any non‐trivial luminous mental states. A version of Williamson's anti‐luminosity appeals to a safety‐theoretic principle connecting knowledge and confidence: if an agent knows p, then p is true in any nearby scenario where she has a similar level of confidence in p. However, the relevant notion of confidence is relatively underexplored. This paper develops a precise theory of confidence: an agent's degree of confidence in p is the objective chance they will rely on p in practical reasoning. This theory of confidence is then used to critically evaluate the anti‐luminosity argument, leading to the surprising conclusion that although there are strong reasons for thinking that luminosity does not obtain, they are quite different from those the existing literature has considered. In particular, we show that once the notion of confidence is properly understood, the failure of luminosity follows from the assumption that knowledge requires high confidence, and does not require any kind of safety principle as a premise.

Recent advances in natural language processing have given rise to a new kind of AI architecture: the language agent. By repeatedly calling an LLM to perform a variety of cognitive tasks, language agents are able to function autonomously to pursue goals specified in natural language and stored in a human-readable format. Because of their architecture, language agents exhibit behavior that is predictable according to the laws of folk psychology: they function as though they have desires and beliefs, and then make and update plans to pursue their desires given their beliefs. We argue that the rise of language agents significantly reduces the probability of an existential catastrophe due to loss of control over an AGI. This is because the probability of such an existential catastrophe is proportional to the difficulty of aligning AGI systems, and language agents significantly reduce that difficulty. In particular, language agents help to resolve three important issues related to aligning AIs: reward misspecification, goal misgeneralization, and uninterpretability.

Under what conditions would an artificially intelligent system have wellbeing? Despite its clear bearing on the ethics of human interactions with artificial systems, this question has received little direct attention. Because all major theories of wellbeing hold that an individual’s welfare level is partially determined by their mental life, we begin by considering whether artificial systems have mental states. We show that a wide range of theories of mental states, when combined with leading theories of wellbeing, predict that certain existing artificial systems have wellbeing. Along the way, we argue that there are good reasons to believe that artificial systems can have wellbeing even if they are not phenomenally conscious. While we do not claim to demonstrate conclusively that AI systems have wellbeing, we argue that there is a significant probability that some AI systems have or will soon have wellbeing, and that this should lead us to reassess our relationship with the intelligent systems we create.

There is a large literature exploring how accuracy constrains rational degrees of belief. This paper turns to the unexplored question of how accuracy constrains knowledge. We begin by introducing a simple hypothesis: increases in the accuracy of an agent’s evidence never lead to decreases in what the agent knows. We explore various precise formulations of this principle, consider arguments in its favour, and explain how it interacts with different conceptions of evidence and accuracy. As we show, the principle has some noteworthy consequences for the wider theory of knowledge. First, it implies that an agent cannot be justified in believing a set of mutually inconsistent claims. Second, it implies the existence of a kind of epistemic blindspot: it is not possible to know that one’s evidence is misleading.

You omega know something when you know it, and know that you know it, and know that you know that you know it, and so on. This paper first argues that omega knowledge matters, in the sense that it is required for rational assertion, action, inquiry, and belief. The paper argues that existing accounts of omega knowledge face major challenges. One account is skeptical, claiming that we have no omega knowledge of any ordinary claims about the world. Another account embraces the KK thesis, and identifies knowledge with omega knowledge. This position faces counterexamples, and struggles to make sense of inexact knowledge. The paper then develops a new account of knowledge, by proposing the principle of Reflective Luminosity: if you know that you know something, then you omega know it. I argue that Reflective Luminosity allows for omega knowledge while avoiding the problems for KK.

You omega know p when you possess every iteration of knowledge of p. This book argues that omega knowledge plays a central role in philosophy. In particular, the book argues that omega knowledge is necessary for permissible assertion, action, inquiry, and belief. Although omega knowledge plays this important role, existing theories of omega knowledge are unsatisfying. One theory, KK, identifies knowledge with omega knowledge. This theory struggles to accommodate cases of inexact knowledge. The other main theory is skeptical, claiming that we do not omega know any ordinary claims about the world. This book develops and critically compares three new theories of omega knowledge.

Recent research has identified a tension between the Safety principle that knowledge is belief without risk of error, and the Closure principle that knowledge is preserved by competent deduction. Timothy Williamson reconciles Safety and Closure by proposing that when an agent deduces a conclusion from some premises, the agent’s method for believing the conclusion includes their method for believing each premise. We argue that this theory is untenable because it implies problematically easy epistemic access to one’s methods. Several possible solutions are explored and rejected.

Schlenker (Semant Pragmat 2(3):1–78, 2009; Philos Stud 151(1):115–142, 2010a; Mind 119(474):377–391, 2010b) provides an algorithm for deriving the presupposition projection properties of an expression from that expression’s classical semantics. In this paper, we consider the predictions of Schlenker’s algorithm as applied to attitude verbs. More specifically, we compare Schlenker’s theory with a prominent view which maintains that attitudes exhibit belief projection, so that presupposition triggers in their scope imply that the attitude holder believes the presupposition (Karttunen in Theor Linguist 34(1):181, 1974; Heim in J Semant 9(3):183–221, 1992; Sudo in The art and craft of semantics: a festschrift for Irene Heim, MIT Press, 2014). We show that Schlenker’s theory does not predict belief projection, and discuss several consequences of this result.

This paper develops a question-sensitive theory of intention. We show that this theory explains some puzzling closure properties of intention. In particular, it can be used to explain why one is rationally required to intend the means to one’s ends, even though one is not rationally required to intend all the foreseen consequences of one’s intended actions. It also explains why rational intention is not always closed under logical implication, and why one can only intend outcomes that one believes to be under one’s control.

Contextology is the science of the dynamics of the conversational context. Contextology formulates laws governing how the shared information states of interlocutors evolve in response to assertion. More precisely, the contextologist attempts to construct a function which, when provided with just a conversation’s pre-update context and the content of an assertion, delivers that conversation’s post-update context. Most contextologists have assumed that the function governing the evolution of the context is simple: the post-update context is just the pre-update context intersected with the content of the assertion. We argue that this assumption is wrong: not only is it false, it is also incoherent given standard contextological assumptions. Moreover, it is impossible in principle to revise it to correctly describe the dynamics of context. We conclude that there can be no science of Contextology. The laws governing the evolution of the context in response to assertion must make essential reference to the private information states of interlocutors.

In ‘Sly Pete’ or ‘standoff’ cases, reasonable speakers accept incompatible conditionals, and communicate them successfully to a trusting hearer. This paper uses the framework of dynamic semantics to offer a new model of the conversational dynamics at play in standoffs, and to articulate several puzzles posed by such cases. The paper resolves these puzzles by embracing a dynamic semantics for conditionals, according to which indicative conditionals require that their antecedents are possible in their local context, and update this body of information by eliminating the possibilities where the antecedent is true and the consequent is false. In this way, the dynamic analysis draws on insights from the material conditional and contextualist analyses, while explaining how standoffs are genuine disagreements.

This paper models knowledge in cases where an agent has multiple experiences over time. Using this model, we introduce a series of observations that undermine the pretheoretic idea that the evidential significance of experience depends on the extent to which that experience matches the world. On the basis of these observations, we model knowledge in terms of what is likely given the agent’s experience. An agent knows p when p is implied by her epistemic possibilities. A world is epistemically possible when its probability given the agent’s experiences is not significantly lower than the probability of the actual world given that experience.

This paper explores the principle that knowledge is fragile, in that whenever S knows that S doesn’t know that S knows that p, S thereby fails to know p. Fragility is motivated by the infelicity of dubious assertions, utterances which assert p while acknowledging higher-order ignorance whether p. Fragility is interestingly weaker than KK, the principle that if S knows p, then S knows that S knows p. Existing theories of knowledge which deny KK by accepting a Margin for Error principle can be conservatively extended with Fragility.

This paper develops an information-sensitive theory of the semantics and probability of conditionals and statements involving epistemic modals. The theory validates a number of principles linking probability and modality, including the principle that the probability of a conditional If A, then C equals the probability of C, updated with A. The theory avoids so-called triviality results, which are standardly taken to show that principles of this sort cannot be validated. To achieve this, we deny that rational agents update their credences via conditionalization. We offer a new rule of update, Hyperconditionalization, which agrees with Conditionalization whenever nonmodal statements are at stake but differs for modal and conditional sentences.

Many defend the thesis that when someone knows p, they couldn’t easily have been wrong about p. But the notion of easy possibility in play is relatively undertheorized. One structural idea in the literature, the principle of Counterfactual Closure (CC), connects easy possibility with counterfactuals: if it easily could have happened that p, and if p were the case, then q would be the case, it follows that it easily could have happened that q. We first argue that while CC is false, there is a true restriction of it to cases involving counterfactual dependence on a coin flip. The failure of CC falsifies a model where the easy possibilities are counterfactually similar to actuality. Next, we show that extant normality models, where the easy possibilities are the sufficiently normal ones, are incompatible with the restricted CC thesis involving coin flips. Next, we develop a new kind of normality theory that can accommodate the restricted version of CC. This new theory introduces a principle of Counterfactual Contamination, which says roughly that any world is fairly abnormal if at that world very abnormal events counterfactually depend on a coin flip. Finally, we explain why coin flips and other related events have a special status. A central take home lesson is that the correct principle in the vicinity of Safety is importantly normality-theoretic rather than (as it is usually conceived) similarity-theoretic.

Formal models of appearance and reality have proved fruitful for investigating structural properties of perceptual knowledge. This paper applies the same approach to epistemic justification. Our central goal is to give a simple account of The Preface, in which justified belief fails to agglomerate. Following recent work by a number of authors, we understand knowledge in terms of normality. An agent knows p iff p is true throughout all relevant normal worlds. To model The Preface, we appeal to the normality of error. Sometimes, it is more normal for reality and appearance to diverge than to match. We show that this simple idea has dramatic consequences for the theory of knowledge and justification. Among other things, we argue that a proper treatment of The Preface requires a departure from the internalist idea that epistemic justification supervenes on the appearances and the widespread idea that one knows most when free from error.

We often claim to know what might be—or probably is—the case. Modal knowledge along these lines creates a puzzle for information-sensitive semantics for epistemic modals. This paper develops a solution. We start with the idea that knowledge requires safe belief: a belief amounts to knowledge only if it could not easily have been held falsely. We then develop an interpretation of the modal operator in safety that allows it to non-trivially embed information-sensitive contents. The resulting theory avoids various paradoxes that arise from other accounts of modal knowledge. It also delivers plausible predictions about modal Gettier cases.

A mental state is luminous if, whenever an agent is in that state, they are in a position to know that they are. Following Timothy Williamson’s Knowledge and Its Limits, a wave of recent work has explored whether there are any non-trivial luminous mental states. A version of Williamson’s anti-luminosity appeals to a safety- theoretic principle connecting knowledge and confidence: if an agent knows p, then p is true in any nearby scenario where she has a similar level of confidence in p. However, the relevant notion of confidence is relatively underexplored. This paper develops a precise theory of confidence: an agent’s degree of confidence in p is the objective chance they will rely on p in practical reasoning. This theory of confidence is then used to critically evaluate the anti-luminosity argument, leading to the surprising conclusion that although there are strong reasons for thinking that luminosity does not obtain, they are quite different from those the existing literature has considered. In particular, we show that once the notion of confidence is properly understood, the failure of luminosity follows from the assumption that knowledge requires high confidence, and does not require any kind of safety principle as a premise

Triviality results threaten plausible principles governing our credence in epistemic modal claims. This paper develops a new account of modal credence which avoids triviality. On the resulting theory, probabilities are assigned not to sets of worlds, but rather to sets of information state-world pairs. The theory avoids triviality by giving up the principle that rational credence is closed under conditionalization. A rational agent can become irrational by conditionalizing on new evidence. In place of conditionalization, the paper develops a new account of updating: conditionalization with normalization.

This paper develops a semantic solution to the puzzle of Free Choice permission. The paper begins with a battery of impossibility results showing that Free Choice is in tension with a variety of classical principles, including Disjunction Introduction and the Law of Excluded Middle. Most interestingly, Free Choice appears incompatible with a principle concerning the behavior of Free Choice under negation, Double Prohibition, which says that Mary can’t have soup or salad implies Mary can’t have soup and Mary can’t have salad. Alonso-Ovalle 2006 and others have appealed to Double Prohibition to motivate pragmatic accounts of Free Choice. Aher 2012, Aloni 2018, and others have developed semantic accounts of Free Choice that also explain Double Prohibition. This paper offers a new semantic analysis of Free Choice designed to handle the full range of impossibility results involved in Free Choice. The paper develops the hypothesis that Free Choice is a homogeneity effect. The claim possibly A or B is defined only when A and B are homogenous with respect to their modal status, either both possible or both impossible. Paired with a notion of entailment that is sensitive to definedness conditions, this theory validates Free Choice while retaining a wide variety of classical principles except for the transitivity of entailment. The homogeneity hypothesis is implemented in two different ways, homogeneous alternative semantics and homogeneous dynamic semantics, with interestingly different consequences.

Many have accepted that ordinary counterfactuals and might counterfactuals are duals. In this paper, I show that this thesis leads to paradoxical results when combined with a few different unorthodox yet increasingly popular theses, including the thesis that counterfactuals are strict conditionals. Given Duality and several other theses, we can quickly infer the validity of another paradoxical principle, ‘The Counterfactual Direct Argument’, which says that ‘A> ’ entails ‘A> ’. First, I provide a collapse theorem for the ‘counterfactual direct argument’. The counterfactual direct argument entails the logical equivalence of the subjunctive and material conditional, given a variety of assumptions. Second, I provide a semantics that validates the counterfactual direct argument without collapse. This theory further develops extant dynamic accounts of conditionals. I give a new semantics for disjunction, on which A or B is only true in a context when A and B are both unsettled. The resulting framework validates CDA while invalidating other commonly accepted principles concerning the conditional and disjunction.