We discuss potential biorisks from large language models (LLMs). AI assistants based on LLMs such as ChatGPT have been shown to significantly reduce barriers to entry for actors wishing to synthesize dangerous, potentially novel pathogens and chemical weapons. The harms from deploying such bioagents could be further magnified by AI-assisted misinformation. We endorse several policy responses to these dangers, including prerelease evaluations of biomedical AIs by subject-matter experts, enhanced surveillance and lab screening procedures, restrictions on AI training data, and access (...) controls on biomedical AI. We conclude with a suggestion about future research directions in bioethics. (shrink)

What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? (...) And what sorts of evidence bear upon this question? -/- In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons. (shrink)

The fact that someone is generous is a reason to admire them. The fact that someone will pay you to admire them is also a reason to admire them. But there is a difference in kind between these two reasons: the former seems to be the ‘right’ kind of reason to admire, whereas the latter seems to be the ‘wrong’ kind of reason to admire. The Wrong Kind of Reasons Problem is the problem of explaining the difference between the ‘right’ (...) and the ‘wrong’ kind of reasons wherever it appears. In this article I argue that two recent proposals for solving the Wrong Kind of Reasons Problem do not work. I then offer an alternative solution that provides a unified, systematic explanation of the difference between the two kinds of reasons. (shrink)

Epistemic instrumentalists face a puzzle. In brief, the puzzle is that if the reason there is to believe in accord with the evidence depends, as the instrumentalist says it does, on agents’ idiosyncratic interests, then there is no reason to expect that this reason is universal. Here, I identify and explain two strategies instrumentalists have used to try and solve this puzzle. I then argue that we should find these strategies wanting. Faced with the failure of these strategies, I articulate (...) a heretofore neglected solution on behalf of instrumentalism. (shrink)

The Morality First strategy for developing AI systems that can represent and respond to human values aims to first develop systems that can represent and respond to moral values. I argue that Morality First and other X-First views are unmotivated. Moreover, according to some widely accepted philosophical views about value, these strategies are positively distorting. The natural alternative, according to which no domain of value comes “first” introduces a new set of challenges and highlights an important but otherwise obscured problem (...) for e-AI developers. (shrink)

Mark Schroeder has recently offered a solution to the problem of distinguishing between the so-called " right " and " wrong " kinds of reasons for attitudes like belief and admiration. Schroeder tries out two different strategies for making his solution work: the alethic strategy and the background-facts strategy. In this paper I argue that neither of Schroeder's two strategies will do the trick. We are still left with the problem of distinguishing the right from the wrong kinds of reasons.

L.A. Paul has argued that an ordinary, natural way of making a decision -- by reflecting on the phenomenal character of the experiences one will have as a result of that decision -- cannot yield rational decision in certain cases. Paul's argument turns on the (in principle) epistemically inaccessible phenomenal character of certain experiences. In this paper I argue that, even granting Paul a range of assumptions, her argument doesn't work to establish its conclusion. This is because, as I argue, (...) the phenomenal character of an experience supervenes on epistemically accessible facts about its non-phenomenal character plus what the deciding agent is like. Because there are principles that link the non-phenomenal character of experiences (together with what a particular agent is like) to the phenomenal character of experiences, agents can reasonably form expectations about the valence of the phenomenal character of the experiences that they are deciding whether to undergo. These reasonable expectations are, I argue, enough to make the ordinary, natural way of making a decision yield rational decision. (shrink)

Humean promotionalists about reasons think that whether there is a reason for an agent to ϕ depends on whether her ϕ-ing promotes the satisfaction of at least one of her desires. Several authors have recently defended probabilistic accounts of promotion, according to which an agent’s ϕ-ing promotes the satisfaction of one of her desires just in case her ϕ-ing makes the satisfaction of that desire more probable relative to some baseline. In this paper I do three things. First, I formalize (...) an argument, due to Jeff Behrends and Joshua DiPaolo, to the effect that Mark Schroeder’s and Stephen Finlay’s probabilistic accounts of promotion cannot be correct. Next, I extend this argument to a recent alternative offered by D. Justin Coates and show how Coates’ attempt to avoid the argument by introducing a distinction between ‘intrinsic’ and ‘extrinsic’ probability doesn’t help. Finally, I suggest an alternative way of understanding promotion in terms of increase in degree of fit between the causal upshot of an action and the content of a desire. I show how this view, disjunctively paired with probabilism about promotion, solves the problems with previous accounts. (shrink)

The use of machine learning, or “artificial intelligence” (AI) in medicine is widespread and growing. In this paper, I focus on a specific proposed clinical application of AI: using models to predict incapacitated patients’ treatment preferences. Drawing on results from machine learning, I argue this proposal faces a special moral problem. Machine learning researchers owe us assurance on this front before experimental research can proceed. In my conclusion I connect this concern to broader issues in AI safety.

What is required for an action to promote the satisfaction of a desire? We reject extant answers and propose an alternative. Our account differs from competing answers in two ways: first, it is contrastive, in that actions promote the satisfaction of desires only as contrasted with other possible actions. Second, it employs a notion of expected fit between desire and world, defined as the weighted sum of the fit between the desire and the world in all possible outcomes, where each (...) weight is given by the probability of the agent’s obtaining the relevant outcome. According to our proposal, then, an action promotes a desire when the expected fit for the desire given that the agent performs the action is greater than the expected fit of the desire given that the agent performs the contrasting action. We highlight this account’s attractive features and explain how it improves on its competitors. (shrink)

Suppose there are no in-principle restrictions on the contents of arbitrarily intelligent agents’ goals. According to “instrumental convergence” arguments, potentially scary things follow. I do two things in this paper. First, focusing on the influential version of the instrumental convergence argument due to Nick Bostrom, I explain why such arguments require an account of “promotion,” i.e., an account of what it is to “promote” a goal. Then, I consider whether extant accounts of promotion in the literature -- in particular, probabilistic (...) and fit-based views of promotion -- can be used to support dangerous instrumental convergence. I argue that neither account of promotion can do the work. The opposite is true: accepting either account of promotion undermines support for instrumental convergence arguments’ existentially worrying conclusions. The conclusion is that we needn’t be scared -- at least not because of arguments concerning instrumental convergence. (shrink)

The moral of Buridan's Ass is that it can sometimes be rational to perform one action rather than another even though one lacks stronger reason to do so. Yet it is also commonly believed that it cannot ever be rational to believe one proposition rather than another if one lacks stronger reason to do so. This asymmetry has been taken to indicate a deep difference between epistemic and practical rationality. According to the view articulated here, the asymmetry should instead be (...) explained by the difference between rational intentions and rational actions. Thus, it turns out, Buridan's Ass-style cases do not indicate an asymmetry between epistemic and practical rationality as such. (shrink)

It’s possible to accept or to reject a promise. According to a new proposal by Abraham Roth, accepting a promise involves intending that the promisee perform the promised action. According to Roth, this view is supported by rational symmetries between promissory acceptance and intention. Here, I show how these symmetries actually generate two problems for the view.

John Broome has proposed a theory of fairness according to which fairness requires that agents’ claims to goods be satisfied in proportion to the relative strength of those claims. In the case of competing claims for a single indivisible good, Broome argues that what fairness requires is the use of a weighted lottery as a surrogate to satisfying the competing claims: the relative chance of each claimant's winning the lottery should be set to the relative strength of each claimant's claim. (...) In this journal, James Kirkpatrick and Nick Eastwood have objected that the use of weighted lotteries in the case of indivisible goods is unacceptable. In this article, I explain why Kirkpatrick and Eastwood's objection misses its mark. (shrink)

Except in extraordinary circumstances, patients' clinical care should reflect their preferences. Incapacitated patients cannot report their preferences. This is a problem. Extant solutions to the problem are inadequate: surrogates are unreliable, and advance directives are uncommon. In response, some authors have suggested developing algorithmic "patient preference predictors" (PPPs) to inform care for incapacitated patients. In a recent paper, Earp et al. propose a new twist on PPPs. Earp et al. suggest we personalize PPPs using modern machine learning (ML) techniques. In (...) this paper, I argue that, contrary to Earp et al.'s claim, personalized patient preference predictors are neither technically feasible nor ethically desirable. (shrink)

Patient preference predictors aim to solve the moral problem of making treatment decisions on behalf of incapacitated patients. This commentary on a case of an unrepresented patient at the end of life considers 3 related problems of such predictors: the problem of restricting the scope of inputs to the models (the “scope” problem), the problem of weighing inputs against one another (the “weight” problem), and the problem of multiple reasonable solutions to the scope and weight problems (the “multiple reasonable models” (...) problem). Each of these problems poses challenges to reliably implementing patient preference predictors in important, high-stakes health care decision making. This commentary also suggests a way forward. (shrink)

Epistemic consequentialism has been charged with ignoring the epistemic separateness of propositions and with (thereby) allowing trade-offs between propositions. Here, I do two things. First, I investigate the metaphor of the epistemic separateness of propositions. I argue that either (i) the metaphor is meaningfully unpacked in a way that is modeled on the moral separateness of persons, in which case it doesn’t support a ban on trade-offs or (ii) it isn’t meaningfully unpacked, in which case it really doesn’t support a (...) ban on trade-offs. Second, I consider the strategy of arguing against the trade-off permitting conception of epistemic rationality that is central to epistemic consequentialism on the basis of our intuitive verdicts about the permissibility of trade-offs in cases. I argue the execution of this strategy suffers a methodological mistake that, once corrected, vitiates the probative value of those intuitive verdicts. Hence the conclusion: the separateness of propositions provides no support for a ban on trade-offs, and an influential independent argument for such a ban is flawed. (shrink)

Jokes are sometimes morally objectionable, and sometimes they are not. What’s the relationship between a joke’s being morally objectionable and its being funny? Philosophers’ answers to this question run the gamut. In this paper I present a new argument for the view that the negative moral value of a joke can affect its comedic value both positively and negatively.

Many philosophers think the distinctive function of deontic evaluation is to guide action. This idea is used in arguments for a range of substantive claims. In this paper, we entirely do one completely destructive thing and partly do one not entirely constructive thing. The first thing: we argue that there is an unrecognized gap between the claim that the function of deontic evaluation is to guide action and attempts to put that claim to use. We consider and reject four arguments (...) intended to bridge this gap. The interim conclusion is thus that arguments starting with the claim that the function of deontic evaluation is to guide action have a lacuna. The second thing: we consider a different tack for making arguments of this sort work. We sketch a methodology one could accept that would do the trick. Unfortunately, as we’ll explain, although this methodology would bridge the gap in arguments that put claims about the function of deontic evaluation to work, it would do so in a way that vitiates any interest we might have in such arguments. As an aside, we’ll also point out how epistemologists, who have recently become interested in the function of epistemic evaluation, appear to already recognize this fact. The conclusion is hence a dilemma: either arguments from deontic function to substance have a lacuna or such arguments lack teeth. (shrink)