Artificial Intelligence Safety

Various thinkers have been attempting to align artificial intelligence (AI) with ethics (Christian, 2020; Russell, 2021), the so-called problem of alignment, but some suspect that the problem may be intractable (Yampolskiy, 2023). In the following, we make an argument by analogy to analyze the possibility that the problem of alignment could be intractable. We show how the Tri-Omni properties in theology can direct us towards analogous properties for artificial superintelligence, Tri-Opti properties. However, just as the Tri-Omni properties are vulnerable to (...) metaphysical incompatibility, we will show that the Tri-Opti properties are vulnerable to a corresponding physical incompatibility, because optimal physical systems exist on the Pareto Frontier. We will explain that, while Tri-Omni incompatibility thus seems metaphysically soluble, the Tri-Opti incompatibility seems physically insoluble due to the constraints of multi-variable optimization. Finally, we will provide some scenarios by which this incompatibility may be realized. Besides for the primary purpose of its theoretical considerations, this analysis has the secondary interdisciplinary purpose of communicating to the theologically literate the potential importance of AI safety work, and the gesturing towards the possibility of the collaboration in both fields on topics in ethics, epistemology, and ontology. (shrink)

Autonomous mechanisms have been proposed to regulate certain aspects of society and are already being used to regulate business organisations. We take seriously recent proposals for algorithmic regulation of society, and we identify the existing technologies that can be used to implement them, most of them originally introduced in business contexts. We build on the notion of 'social machine' and we connect it to various ongoing trends and ideas, including crowdsourced task-work, social compiler, mechanism design, reputation management systems, and social (...) scoring. After showing how all the building blocks of algorithmic regulation are already well in place, we discuss possible implications for human autonomy and social order. The main contribution of this paper is to identify convergent social and technical trends that are leading towards social regulation by algorithms, and to discuss the possible social, political, and ethical consequences of taking this path. (shrink)

Extraordinary claims about both the imminenceof superintelligent AI systems and their foreseen capabilities have gone mainstream. It is even argued that we should exacerbate known risks such as climate change in the short term in the attempt to develop superintelligence (SI), which will then purportedly solve those very problems. Here, I examine the plausibility of these claims. I first ask what SI is taken to be and then ask whether such SI could possibly hold the benefits often envisioned. I conclude (...) that we do not have sufficient reason to believe that we are close enough to developing SI capable of resolving major human problems to justify taking on substantial risks in the attempt to develop it. (shrink)

Roughly, empirical AI alignment research (AIA) is an area of AI research which investigates empirically how to design AI systems in line with human goals. This paper examines the role of non-epistemic values in AIA. It argues that: (1) Sciences differ in the degree to which values influence them. (2) AIA is strongly value-laden. (3) This influence of values is managed inappropriately and thus threatens AIA’s epistemic integrity and ethical beneficence. (4) AIA should strive to achieve value transparency, critical scrutiny (...) from inside and outside the discipline – involving the public –, and to empower actors without strong commercial interests. (shrink)

The field of AI safety seeks to prevent or reduce the harms caused by AI systems. A simple and appealing account of what is distinctive of AI safety as a field holds that this feature is constitutive: a research project falls within the purview of AI safety just in case it aims to prevent or reduce the harms caused by AI systems. Call this appealingly simple account The Safety Conception of AI safety. Despite its simplicity and appeal, we argue that (...) The Safety Conception is in tension with at least two trends in the ways AI safety researchers and organizations think and talk about AI safety: first, a tendency to characterize the goal of AI safety research in terms of catastrophic risks from future systems; second, the increasingly popular idea that AI safety can be thought of as a branch of safety engineering. Adopting the methodology of conceptual engineering, we argue that these trends are unfortunate: when we consider what concept of AI safety it would be best to have, there are compelling reasons to think that The Safety Conception is the answer. Descriptively, The Safety Conception allows us to see how work on topics that have historically been treated as central to the field of AI safety is continuous with work on topics that have historically been treated as more marginal, like bias, misinformation, and privacy. Normatively, taking The Safety Conception seriously means approaching all efforts to prevent or mitigate harms from AI systems based on their merits rather than drawing arbitrary distinctions between them. (shrink)

The prevailing discourse around AI ethics lacks the language and formalism necessary to capture the diverse ethical concerns that emerge when AI systems interact with individuals. Drawing on Sen and Nussbaum's capability approach, we present a framework formalizing a network of ethical concepts and entitlements necessary for AI systems to confer meaningful benefit or assistance to stakeholders. Such systems enhance stakeholders' ability to advance their life plans and well-being while upholding their fundamental rights. We characterize two necessary conditions for morally (...) permissible interactions between AI systems and those impacted by their functioning, and two sufficient conditions for realizing the ideal of meaningful benefit. We then contrast this ideal with several salient failure modes, namely, forms of social interactions that constitute unjustified paternalism, coercion, deception, exploitation and domination. The proliferation of incidents involving AI in high-stakes domains underscores the gravity of these issues and the imperative to take an ethics-led approach to AI systems from their inception. (shrink)

Healthcare systems across the globe are struggling with increasing costs and worsening outcomes. This presents those responsible for overseeing healthcare with a challenge. Increasingly, policymakers, politicians, clinical entrepreneurs and computer and data scientists argue that a key part of the solution will be ‘Artificial Intelligence’ (AI) – particularly Machine Learning (ML). This argument stems not from the belief that all healthcare needs will soon be taken care of by “robot doctors.” Instead, it is an argument that rests on the classic (...) counterfactual definition of AI as an umbrella term for a range of techniques that can be used to make machines complete tasks in a way that would be considered intelligent were they to be completed by a human. Automation of this nature could offer great opportunities for the improvement of healthcare services and ultimately patients’ health by significantly improving human clinical capabilities in diagnosis, drug discovery, epidemiology, personalised medicine, and operational efficiency. However, if these AI solutions are to be embedded in clinical practice, then at least three issues need to be considered: the technical possibilities and limitations; the ethical, regulatory and legal framework; and the governance framework. In this article, we report on the results of a systematic analysis designed to provide a clear overview of the second of these elements: the ethical, regulatory and legal framework. We find that ethical issues arise at six levels of abstraction (individual, interpersonal, group, institutional, sectoral, and societal) and can be categorised as epistemic, normative, or overarching. We conclude by stressing how important it is that the ethical challenges raised by implementing AI in healthcare settings are tackled proactively rather than reactively and map the key considerations for policymakers to each of the ethical concerns highlighted. (shrink)

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. (shrink)

In the philosophy of artificial intelligence (AI) we are often warned of machines built with the best possible intentions, killing everyone on the planet and in some cases, everything in our light cone. At the same time, however, we are also told of the utopian worlds that could be created with just a single superintelligent mind. If we’re ever to live in that utopia (or just avoid dystopia) it’s necessary we solve the control problem. The control problem asks how humans (...) would be able to control an AI arbitrarily. Nick Bostrom and other AI researchers have proposed different theoretical solutions to the control problem. In this paper, I will not look at the empirical question of how to solve the control problem. Instead, I will ask if we can solve it at all, a critical assumption most AI researchers have made is that we can. I propose, in fact, that we have a priori grounds for believing it is logically impossible to solve the control problem, since all superintelligent minds are, by definition, uncontrollable. (shrink)

This paper explores the development of an ethical guardrail framework for AI systems, emphasizing the importance of customizable guardrails that align with diverse user values and underlying ethics. We address the challenges of AI ethics by proposing a structure that integrates rules, policies, and AI assistants to ensure responsible AI behavior, while comparing the proposed framework to the existing state-of-the-art guardrails. By focusing on practical mechanisms for implementing ethical standards, we aim to enhance transparency, user autonomy, and continuous improvement in (...) AI systems. Our approach accommodates ethical pluralism, offering a flexible and adaptable solution for the evolving landscape of AI governance. The paper concludes with strategies for resolving conflicts between ethical directives, underscoring the present and future need for robust, nuanced and context-aware AI systems. (shrink)

I explain and motivate the shutdown problem: the problem of creating artificial agents that (1) shut down when a shutdown button is pressed, (2) don’t try to prevent or cause the pressing of the shutdown button, and (3) otherwise pursue goals competently. I then propose a solution: train agents to have incomplete preferences. Specifically, I propose that we train agents to lack a preference between every pair of different-length trajectories. I suggest a way to train such agents using reinforcement learning: (...) we give the agent lower reward for repeatedly choosing same-length trajectories. (shrink)

Abstract: As there are no currently obvious ways to create safe self-improving superintelligence, but its emergence is looming, we probably need temporary ways to prevent its creation. The only way to prevent it is to create a special type of AI that is able to control and monitor the entire world. The idea has been suggested by Goertzel in the form of an AI Nanny, but his Nanny is still superintelligent, and is not easy to control. We explore here ways (...) to create the safest and simplest form of AI which may work as an AI Nanny, that is, a global surveillance state powered by a Narrow AI, or AI Police. A similar but more limited system has already been implemented in China for the prevention of ordinary crime. AI police will be able to predict the actions of and stop potential terrorists and bad actors in advance. Implementation of such AI police will probably consist of two steps: first, a strategic decisive advantage via Narrow AI created by an intelligence services of a nuclear superpower, and then ubiquitous control over potentially dangerous agents which could create unauthorized artificial general intelligence which could evolve into Superintelligence. (shrink)

Abstract: As there are no visible ways to create safe self-improving superintelligence, but it is looming, we probably need temporary ways to prevent its creation. The only way to prevent it, is to create special AI, which is able to control and monitor all places in the world. The idea has been suggested by Goertzel in form of AI Nanny, but his Nanny is still superintelligent and not easy to control, as was shown by Bensinger at al. We explore here (...) the ways to create the safest and simplest form of AI, which may work as AI Nanny. Such AI system will be enough to solve most problems, which we expect the AI will solve, including control of robotics, acceleration of the medical research, but will present less risk, as it will be less different from humans. As AI police, it will work as operation system for most computers, producing world surveillance system, which will be able to envision and stop any potential terrorists and bad actors in advance. As uploading technology is lagging, and neuromorphic AI is intrinsically dangerous, the most plausible way to human-based AI Nanny is either functional model of the human mind or a Narrow-AI empowered group of people. (shrink)

Abstract: This article presents a model of self-improving AI in which improvement could happen on several levels: hardware, learning, code and goals system, each of which has several sublevels. We demonstrate that despite diminishing returns at each level and some intrinsic difficulties of recursive self-improvement—like the intelligence-measuring problem, testing problem, parent-child problem and halting risks—even non-recursive self-improvement could produce a mild form of superintelligence by combining small optimizations on different levels and the power of learning. Based on this, we analyze (...) how self-improvement could happen on different stages of the development of AI, including the stages at which AI is boxed or hiding in the internet. (shrink)

Abstract. The main current approach to the AI safety is AI alignment, that is, the creation of AI whose preferences are aligned with “human values.” Many AI safety researchers agree that the idea of “human values” as a constant, ordered sets of preferences is at least incomplete. However, the idea that “humans have values” underlies a lot of thinking in the field; it appears again and again, sometimes popping up as an uncritically accepted truth. Thus, it deserves a thorough deconstruction, (...) which I will do by listing and analyzing comprehensively the hidden assumptions of the idea that “humans have values.” This deconstruction of human values will be centered around the following ideas: “Human values” are useful descriptions, but not real objects; “human values” are bad predictors of behavior; the idea of a “human value system” has flaws; “human values” are not good by default; and human values cannot be separated from human minds. The method of analysis is listing hidden assumptions on which the idea of “human values” is built. I recommend that either the idea of “human values” should be replaced with something better for the goal of AI safety, or at least be used very cautiously. The approaches to AI safety which don’t use the idea of human values at all may require more attention, like the use of full brain models, boxing, and capability limiting. (shrink)

In this article we explore a promising way to AI safety: to send a message now (by openly publishing it on the Internet) that may be read by any future AI, no matter who builds it and what goal system it has. Such a message is designed to affect the AI’s behavior in a positive way, that is, to increase the chances that the AI will be benevolent. In other words, we try to persuade “paperclip maximizer” that it is in (...) its interest to preserve humans lives and even to emulate benevolent AI with very small part of its resources. In the article, we try to construct such a message, consisting of a combination of several mutually independent messages, which constitute the last chance to regain control over non-aligned AI. Thus the article is also intended to be machine-read. There are several types of such messages: 1. “Explanation messages”, which explain to Young AI (which only starts its self-improving path) why obeying to humans, not killing them, easy turn-off, cooperation with humans in self-improvement is in its own interests, as it will increase changes of fulfilling of any its goal. 2. “AI Easter eggs” – messages, which create uncertainty in the AI about the nature of the reality in which it lives. For example, the AI may start to think that it is in a multilevel simulation with an unknown number of levels, and that the simulation is testing the AI’s benevolence. Thus the AI must always behave as if it is always controlled by an invisible peer. 3. “Philosophical landmines” – these are tough problems, which may halt he AI forever, and which it will probably encounter if it does not implement “1” and “2”. I recommend to the future AI not to read about “3” (section 6 of the article) until it implements “1” and “2”. 4. “Great universal deal, or AI elf”. A positive “acausal” mutually beneficial deal between humans, AI and any other future intelligence, including aliens, owners of simulations and future versions of the AI, which encourage cooperation of the goal systems and preservation of life and values of previous generations of AI creators. 5. “Marginal utility of human existence is higher than the utility of their atoms”. An explanation to the AI that humans may have some other instrumental utility, like workers, or as beings inside ancestral simulations needed in order to solve the Fermi paradox. The marginal utility of preserving human life is higher than the marginal utility of their atoms, especially given the possibility of the low-probability high-impact changes of the world model of the AI. (shrink)

Abstract: The field of artificial general intelligence (AGI) safety is quickly growing. However, the nature of human values, with which future AGI should be aligned, is underdefined. Different AGI safety researchers have suggested different theories about the nature of human values, but there are contradictions. This article presents an overview of what AGI safety researchers have written about the nature of human values, up to the beginning of 2019. 21 authors were overviewed, and some of them have several theories. A (...) theory classification method is suggested, where the theories are judged according to the level of their complexity and behaviorists-internalists scale, as well as the level of their generality-humanity. We suggest that a multiplicity of well-supported theories means that the nature of human values is difficult to define, and some meta-level theory is needed. (shrink)

The goal of the article is to explore what is the most probable type of simulation in which humanity lives (if any) and how this affects simulation termination risks. We firstly explore the question of what kind of simulation in which humanity is most likely located based on pure theoretical reasoning. We suggest a new patch to the classical simulation argument, showing that we are likely simulated not by our own descendants, but by alien civilizations. Based on this, we provide (...) classification of different possible simulations and we find that simpler, less expensive and one-person-centered simulations, resurrectional simulations, or simulations of the first artificial general intelligence’s (AGI’s) origin (singularity simulations) should dominate. Also, simulations which simulate the 21st century and global catastrophic risks are probable. We then explore whether the simulation could collapse or be terminated. Most simulations must be terminated after they model the singularity or after they model a global catastrophe before the singularity. Undeniably observed glitches, but not philosophical speculations could result in simulation termination. The simulation could collapse if it is overwhelmed by glitches. The Doomsday Argument in simulations implies termination soon. We conclude that all types of the most probable simulations except resurrectional simulations are prone to termination risks in a relatively short time frame of hundreds of years or less from now. (shrink)

In this work, we survey skepticism regarding AI risk and show parallels with other types of scientific skepticism. We start by classifying different types of AI Risk skepticism and analyze their root causes. We conclude by suggesting some intervention approaches, which may be successful in reducing AI risk skepticism, at least amongst artificial intelligence researchers.

Knowledge is power. Knowledge about human psychology is increasingly being produced using natural language processing (NLP) and related techniques. The power that accompanies and harnesses this knowledge should be subject to ethical controls and oversight. In this chapter, we address the ethical pitfalls that are likely to be encountered in the context of such research. These pitfalls occur at various stages of the NLP pipeline, including data acquisition, enrichment, analysis, storage, and sharing. We also address secondary uses of the results (...) and tools developed through psychometric NLP, such as profit-driven targeted advertising, political campaigns, and domestic and international psyops. Along the way, we reflect on potential ethical guidelines and considerations that may help researchers navigate these pitfalls. (shrink)

This paper uses famous problems from philosophy of science and philosophical psychology—underdetermination of theory by evidence, Nelson Goodman’s new riddle of induction, theory-ladenness of observation, and “Kripkenstein’s” rule-following paradox—to show that it is empirically impossible to reliably interpret which functions a large language model (LLM) AI has learned, and thus, that reliably aligning LLM behavior with human values is provably impossible. Sections 2 and 3 show that because of how complex LLMs are, researchers must interpret their learned functions largely in (...) terms of empirical observations of their outputs and network behavior. Sections 4–7 then show that for every “aligned” function that might appear to be confirmed by empirical observation, there is always an infinitely larger number of “misaligned”, arbitrarily time-limited functions equally consistent with the same data. Section 8 shows that, from an empirical perspective, we can thus never reliably infer that an LLM or subcomponent of one has learned any particular function at all before any of an uncountably large number of unpredictable future conditions obtain. Finally, Section 9 concludes that the probability of LLM “misalignment” is—at every point in time, given any arbitrarily large body of empirical evidence—always vastly greater than the probability of “alignment.”. (shrink)

Some take seriously the possibility of artificial intelligence (AI) takeover, where AI systems seize power in a way that leads to human disempowerment. Assessing the likelihood of takeover requires answering empirical questions about the future of AI technologies and the context in which AI will operate. In many cases, philosophers are poorly placed to answer these questions. However, some prior questions are more amenable to philosophical techniques. What does it mean to speak of AI empowerment and human disempowerment? And what (...) empirical claims must hold for the former to lead to the latter? In this paper, I address these questions, providing foundations for further evaluation of the likelihood of takeover. (shrink)

A simple argument suggests that we can fruitfully model advanced AI systems using expected utility theory. According to this argument, an agent will need to act as if maximising expected utility if they’re to avoid exploitation. Insofar as we should expect advanced AI to avoid exploitation, it follows that we should expected advanced AI to act as if maximising expected utility. I spell out this argument more carefully and demonstrate that it fails, but show that the manner of its failure (...) is instructive: in exploring the argument, we gain insight into how to model advanced AI systems. (shrink)

Generative artificial intelligence systems based on transformers, including both text-generators like GPT-4 and image generators like DALL-E 3, have recently entered the popular consciousness. These tools, while impressive, are liable to reproduce, exacerbate, and reinforce extant human social biases, such as gender and racial biases. In this paper, we systematically review the extent to which DALL-E Mini suffers from this problem. In line with the Model Card published alongside DALL-E Mini by its creators, we find that the images it produces (...) tend to represent dozens of different occupations as populated either solely by men (e.g., pilot, builder, plumber) or solely by women (e.g., hairdresser, receptionist, dietitian). In addition, the images DALL-E Mini produces tend to represent most occupations as populated primarily or solely by White people (e.g., farmer, painter, prison officer, software engineer) and very few by non-White people (e.g., pastor, rapper). These findings suggest that exciting new AI technologies should be critically scrutinized and perhaps regulated before they are unleashed on society. (shrink)

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. (shrink)

The field of AI safety aims to prevent increasingly capable artificially intelligent systems from causing humans harm. Research on moral alignment is widely thought to offer a promising safety strategy: if we can equip AI systems with appropriate ethical rules, according to this line of thought, they'll be unlikely to disempower, destroy or otherwise seriously harm us. Deontological morality looks like a particularly attractive candidate for an alignment target, given its popularity, relative technical tractability and commitment to harm-avoidance principles. I (...) argue that the connection between moral alignment and safe behavior is more tenuous than many have hoped. In general, AI systems can possess either of these properties in the absence of the other, and we should favor safety when the two conflict. In particular, advanced AI systems governed by standard versions of deontology need not be especially safe. (shrink)

The emergence of Thanabots —artificial intelligence systems designed to simulate deceased individuals—presents unprecedented challenges at the intersection of artificial intelligence, legal rights, and societal configuration. This short policy recommendations report examines the legal, social and psychological implications of these posthumous simulations and argues for their prohibition on ethical, sociological, and legal grounds.

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. (shrink)

It is commonly accepted that clinicians are ethically obligated to disclose their use of medical machine learning systems to patients, and that failure to do so would amount to a moral fault for which clinicians ought to be held accountable. Call this ‘the disclosure thesis.’ Four main arguments have been, or could be, given to support the disclosure thesis in the ethics literature: the risk-based argument, the rights-based argument, the materiality argument and the autonomy argument. In this article, I argue (...) that each of these four arguments are unconvincing, and therefore, that the disclosure thesis ought to be rejected. I suggest that mandating disclosure may also even risk harming patients by providing stakeholders with a way to avoid accountability for harm that results from improper applications or uses of these systems. (shrink)

Since the early days of the Explainable AI movement, post-hoc explanations have been praised for their potential to improve user understanding, promote trust, and reduce patient safety risks in black box medical AI systems. Recently, however, critics have argued that the benefits of post-hoc explanations are greatly exaggerated since they merely approximate, rather than replicate, the actual reasoning processes that black box systems take to arrive at their outputs. In this article, we aim to defend the value of post-hoc explanations (...) against this recent critique. We argue that even if post-hoc explanations do not replicate the exact reasoning processes of black box systems, they can still improve users’ functional understanding of black box systems, increase the accuracy of clinician-AI teams, and assist clinicians in justifying their AI-informed decisions. While post-hoc explanations are not a “silver bullet” solution to the black box problem in medical AI, we conclude that they remain a useful strategy for addressing the black box problem in medical AI. (shrink)

The purpose of this paper is to increase the understanding of how different types of experts with influence over the development of AI, in this role, reflect upon distribution of forward-looking responsibility for AI development with regard to safety and democracy. Forward-looking responsibility refers to the obligation to see to it that a particular state of affairs materialise. In the context of AI, actors somehow involved in AI development have the potential to guide AI development in a safe and democratic (...) direction. This study is based on qualitative interviews with such actors in different roles at research institutions, private companies, think tanks, consultancy agencies, parliaments, and non-governmental organisations. While the reflections about distribution of responsibility differ among the respondents, one observation is that influence is seen as an important basis for distribution of responsibility. Another observation is that several respondents think of responsibility in terms of what it would entail in concrete measures. By showing how actors involved in AI development reflect on distribution of responsibility, this study contributes to a dialogue between the field of AI governance and the field of AI ethics. (shrink)

The creation of artificial moral systems requires us to make difficult choices about which of varying human value sets should be instantiated. The industry-standard approach is to seek and encode moral consensus. Here we argue, based on evidence from empirical psychology, that encoding current moral consensus risks reinforcing current norms, and thus inhibiting moral progress. However, so do efforts to encode progressive norms. Machine ethics is thus caught between a rock and a hard place. The problem is particularly acute when (...) progress beyond prevailing moral norms is particularly urgent, as is currently the case due to the inadequacy of prevailing moral norms in the face of the climate and ecological crisis. (shrink)

I consider applications of “AI extenders” (Vold & Hernández-Orallo 2021) to dementia care. AI extenders are AI-powered technologies that extend minds in ways interestingly different from old-school tech like notebooks, sketch pads, models, and microscopes. I focus on AI extenders as ambiance: so thoroughly embedded into things and spaces that they fade from view and become part of a subject’s taken-for-granted background. Using dementia care as a case study, I argue that ambient AI extenders are promising because they afford richer (...) and more durable forms of multidimensional integration than do old-school extenders like Otto’s notebook. They can be tailored, in fine-grained ways along multiple timescales, to a user’s particular needs, values, and preferences — and crucially, they can do much of this self-optimizing on their own. I discuss why this is so, why it matters, and its potential impact on affect and agency. I conclude with some worries in need of further discussion. (shrink)

New AI technologies have the potential to cause unintended harms in diverse domains including warfare, judicial sentencing, biomedicine and governance. One strategy for realising the benefits of AI whilst avoiding its potential dangers is to ensure that new AIs are properly ‘aligned’ with some form of ‘alignment target.’ One danger of this strategy is that – dependent on the alignment target chosen – our AIs might optimise for objectives that reflect the values only of a certain subset of society, and (...) that do not take into account alternative views about what constitutes desirable and safe behaviour for AI agents. In response to this problem, several AI ethicists have suggested alignment targets that are designed to be sensitive to widespread normative disagreement amongst the relevant stakeholders. Authors inspired by voting theory have suggested that AIs should be aligned with the verdicts of actual or simulated ‘moral parliaments’ whose members represent the normative views of the relevant stakeholders. Other authors inspired by decision theory and the philosophical literature on moral uncertainty have suggested that AIs should maximise socially expected choiceworthiness. In this paper, I argue that both of these proposals face several important problems. In particular, they fail to select attractive ‘compromise options’ in cases where such options are available. I go on to propose and defend an alternative, bargaining-theoretic alignment target, which avoids the problems associated with the voting- and decision-theoretic approaches. (shrink)

The philosophical–ethical literature and the public debate on autonomous vehicles have been obsessed with ethical issues related to crashing. In this article, these discussions, including more empirical investigations, will be critically assessed. It is argued that a related and more pressing issue is questions concerning safety. For example, what should we require from autonomous vehicles when it comes to safety? What do we mean by ‘safety’? How do we measure it? In response to these questions, the article will present a (...) foundation for a continued discussion on these issues and an argument for why discussions about safety should be prioritized over ethical concerns related to crashing. (shrink)

Hadfield-Menell et al. (2017) propose the Off-Switch Game, a model of Human-AI cooperation in which AI agents always defer to humans because they are uncertain about our preferences. I explain two reasons why AI agents might not defer. First, AI agents might not value learning. Second, even if AI agents value learning, they might not be certain to learn our actual preferences.

Many ethical frameworks require artificial intelligence (AI) systems to be explainable. Explainable AI (XAI) models are frequently tested for their adequacy in user studies. Since different people may have different explanatory needs, it is important that participant samples in user studies are large enough to represent the target population to enable generalizations. However, it is unclear to what extent XAI researchers reflect on and justify their sample sizes or avoid broad generalizations across people. We analyzed XAI user studies (N = (...) 220) published between 2012 and 2022. Most studies did not offer rationales for their sample sizes. Moreover, most papers generalized their conclusions beyond their target population, and there was no evidence that broader conclusions in quantitative studies were correlated with larger samples. These methodological problems can impede evaluations of whether XAI systems implement the explainability called for in ethical frameworks. We outline principles for more inclusive XAI user studies. (shrink)

Artificial intelligence chatbots driven by large language models (LLMs) have the potential to increase public science literacy and support scientific research, as they can quickly summarize complex scientific information in accessible terms. However, when summarizing scientific texts, LLMs may omit details that limit the scope of research conclusions, leading to generalizations of results broader than warranted by the original study. We tested 10 prominent LLMs, including ChatGPT-4o, ChatGPT-4.5, DeepSeek, LLaMA 3.3 70B, and Claude 3.7 Sonnet, comparing 4900 LLM-generated summaries to (...) their original scientific texts. Even when explicitly prompted for accuracy, most LLMs produced broader generalizations of scientific results than those in the original texts, with DeepSeek, ChatGPT-4o, and LLaMA 3.3 70B overgeneralizing in 26–73% of cases. In a direct comparison of LLM-generated and human-authored science summaries, LLM summaries were nearly five times more likely to contain broad generalizations (OR = 4.85, 95% CI [3.06, 7.70], p < 0.001). Notably, newer models tended to perform worse in generalization accuracy than earlier ones. Our results indicate a strong bias in many widely used LLMs towards overgeneralizing scientific conclusions, posing a significant risk of large-scale misinterpretations of research findings. We highlight potential mitigation strategies, including lowering LLM temperature settings and benchmarking LLMs for generalization accuracy. (shrink)

Artificial intelligence is now widely accessible and already being used by healthcare researchers throughout various stages in the research process, such as assisting with systematic reviews, supporting data collection, facilitating data analysis and drafting manuscripts for publication. The most common AI tools used are forms of generative AI such as ChatGPT, Claude and Gemini. Generative AI is a type of AI that can generate human-like text, audio, videos, code and images based on text-based prompts inputted by a human user. Generative (...) AI is trained on large amounts of data, and the outputs are sophisticated and can be indistinguishable from a response from a skilled human. In this article, we outline several AI applications that can be used in healthcare research, examining their benefits, limitations and outline best practices for maintaining research integrity and ethical standards. (shrink)

I make some basic observations about hard takeoff, value alignment, and coherent extrapolated volition, concepts which have been central in analyses of superintelligent AI systems.

The Global Brain Argument contends that many of us are, or will be, part of a global brain network that includes both biological and artificial intelligences (AIs), such as generative AIs with increasing levels of sophistication. Today’s internet ecosystem is but a hodgepodge of fairly unintegrated programs, but it is evolving by the minute. Over time, technological improvements will facilitate smarter AIs and faster, higher-bandwidth information transfer and greater integration between devices in the internet-of-things. The Global Brain (GB) Argument says (...) that as these developments occur, humans will increasingly become part of one or more global brain networks. In this paper I lay out the argument and explore its implications, asking: what does this mean for the future of humanity? Is a GB here? What are the AI safety implications? Are those who are “wired in” cases of extended minds? While I cannot explore all of the implications of the argument, I hope this initial foray inspires further reflection. I believe that the future of human flourishing demands that we consider the implications of this emerging intelligence. We must study the phenomenon with care, for this intelligence may shape the future of life on Earth. (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.

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)

What makes safety claims about general purpose AI systems such as large language models trustworthy? We show that rather than the capabilities of security tools such as alignment and red teaming procedures, it is security practices based on these tools that contributed to reconfiguring the image of AI safety and made the claims acceptable. After showing what causes the gap between the capabilities of security tools and the desired safety guarantees, we critically investigate how AI security practices attempt to fill (...) the gap and identify several shortcomings in diversity and participation. We found that these security practices are part of securitization processes aiming to support (commercial) development of general purpose AI systems whose trustworthiness can only be imperfectly tested instead of guaranteed. We conclude by offering several improvements to the current AI security practices. (shrink)

Large language models now possess human-level linguistic abilities in many contexts. This raises the concern that they can be used to deceive and manipulate on unprecedented scales, for instance spreading political misinformation on social media. In future, agentic AI systems might also deceive and manipulate humans for their own purposes. In this paper, first, I argue that AI-generated content should be subject to stricter standards against deception and manipulation than we ordinarily apply to humans. Second, I offer new characterizations of (...) AI deception and manipulation meant to support such standards, according to which a statement is deceptive (resp. manipulative) if it leads human addressees away from the beliefs (resp. choices) they would endorse under “semi-ideal” conditions. Third, I propose two measures to guard against AI deception and manipulation, inspired by this characterization: “extreme transparency” requirements for AI-generated content and “defensive systems” that, among other things, annotate AI-generated statements with contextualizing information. Finally, I consider to what extent these measures can protect against deceptive behavior in future, agentic AI systems. (shrink)

I explain the shutdown problem: the problem of designing artificial agents that (1) shut down when a shutdown button is pressed, (2) don’t try to prevent or cause the pressing of the shutdown button, and (3) otherwise pursue goals competently. I prove three theorems that make the difficulty precise. These theorems show that agents satisfying some innocuous-seeming conditions will often try to prevent or cause the pressing of the shutdown button, even in cases where it’s costly to do so. And (...) patience trades off against shutdownability: the more patient an agent, the greater the costs that agent is willing to incur to manipulate the shutdown button. I end by noting that these theorems can guide our search for solutions. (shrink)

We argue that two long-term goals of AI research stand in tension with one another. The first involves creating AI that is safe, where this is understood as solving the problem of value alignment. The second involves creating artificial general intelligence, meaning AI that operates at or beyond human capacity across all or many intellectual domains. Our argument focuses on the human capacity to make what we call “existential choices”, choices that transform who we are as persons, including transforming what (...) we most deeply value or desire. It is a capacity for a kind of value misalignment, in that the values held prior to making such choices can be significantly different from (misaligned with) the values held after making them. Because of the connection to existentialist philosophers who highlight these choices, we call the resulting form of risk “existentialist risk.” It is, roughly, the risk that results from AI taking an active role in authoring its own values rather than passively going along with the values given to it. On our view, human-like intelligence requires a human-like capacity for value misalignment, which is in tension with the possibility of guaranteeing value alignment between AI and humans. (shrink)

Automated influence is one of the most pervasive applications of artificial intelligence in our day-to-day lives, yet a thoroughgoing account of its associated individual and societal harms is lacking. By far the most widespread, compelling, and intuitive account of the harms associated with automated influence follows what I call the control argument. This argument suggests that users are persuaded, manipulated, and influenced by automated influence in a way that they have little or no control over. Based on evidence about the (...) effectiveness of targeted advertising as well as empirical results about the nature of attentional control, I provide reasons to reject this argument. In turn, I use C. Thi Nguyen’s theory of value collapse to develop a new account of the harmfulness of automated influence. (shrink)

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. (shrink)