Which artificial intelligence (AI) systems are agents? To answer this question, I propose a multidimensional account of agency. According to this account, a system's agency profile is jointly determined by its level of goal-directedness and autonomy as well as is abilities for directly impacting the surrounding world, long-term planning and acting for reasons. Rooted in extant theories of agency, this account enables fine-grained, nuanced comparative characterizations of artificial agency. I show that this account has multiple important virtues and is more (...) informative than alternatives. More speculatively, it may help to illuminate two important emerging questions in AI ethics: 1. Can agency contribute to the moral status of non-human beings, and how? 2. When and why might AI systems exhibit power-seeking behaviour and does this pose an existential risk to humanity? (shrink)

Many researchers and intellectuals warn about extreme risks from artificial intelligence. However, these warnings typically came without systematic arguments in support. This paper provides an argument that AI will lead to the permanent disempowerment of humanity, e.g. human extinction, by 2100. It rests on four substantive premises which it motivates and defends: first, the speed of advances in AI capability, as well as the capability level current systems have already reached, suggest that it is practically possible to build AI systems (...) capable of disempowering humanity by 2100. Second, due to incentives and coordination problems, if it is possible to build such AI, it will be built. Third, since it appears to be a hard technical problem to build AI which is aligned with the goals of its designers, and many actors might build powerful AI, misaligned powerful AI will be built. Fourth, because disempowering humanity is useful for a large range of misaligned goals, such AI will try to disempower humanity. If AI is capable of disempowering humanity and tries to disempower humanity by 2100, then humanity will be disempowered by 2100. This conclusion has immense moral and prudential significance. (shrink)

How can one build AI systems such that they pursue the goals their designers want them to pursue? This is the alignment problem. Numerous authors have raised concerns that, as research advances and systems become more powerful over time, misalignment might lead to catastrophic outcomes, perhaps even to the extinction or permanent disempowerment of humanity. In this paper, I analyze the severity of this risk based on current instances of misalignment. More specifically, I argue that contemporary large language models and (...) game-playing agents are sometimes misaligned. These cases suggest that misalignment tends to have a variety of features: misalignment can be hard to detect, predict and remedy, it does not depend on a specific architecture or training paradigm, it tends to diminish a system’s usefulness and it is the default outcome of creating AI via machine learning. Subsequently, based on these features, I show that the risk of AI alignment magnifies with respect to more capable systems. Not only might more capable systems cause more harm when misaligned, aligning them should be expected to be more difficult than aligning current AI. (shrink)

Machine learning researchers distinguish between reinforcement learning and supervised learning and refer to reinforcement learning systems as “agents”. This paper vindicates the claim that systems trained by reinforcement learning are agents while those trained by supervised learning are not. Systems of both kinds satisfy Dretske’s criteria for agency, because they both learn to produce outputs selectively in response to inputs. However, reinforcement learning is sensitive to the instrumental value of outputs, giving rise to systems which exploit the effects of outputs (...) on subsequent inputs to achieve good performance over episodes of interaction with their environments. Supervised learning systems, in contrast, merely learn to produce better outputs in response to individual inputs. (shrink)

The last 5 years have seen a series of remarkable achievements in deep-neural-network-based artificial intelligence research, and some modellers have argued that their performance compares favourably to human cognition. Critics, however, have argued that processing in deep neural networks is unlike human cognition for four reasons: they are (i) data-hungry, (ii) brittle, and (iii) inscrutable black boxes that merely (iv) reward-hack rather than learn real solutions to problems. This article rebuts these criticisms by exposing comparative bias within them, in the (...) process extracting some more general lessons that may also be useful for future debates. (shrink)

Motivational internalism is the thesis that captures the commonplace thought that moral judgements are necessarily motivationally efficacious. But this thesis appears to be in tension with another aspect of our ordinary moral experience. Proponents of the contrast thesis, motivational externalism, cite everyday examples of amoralism to demonstrate that it is conceptually possible to be completely unmoved by what seem to be sincere first-person moral judgements. This paper argues that the challenge of amoralism gives us no reason to reject or modify (...) motivational internalism. Instead of attempting to diagnose the motivational failure of the amoral agent or restrict the internalist thesis in the face of these examples, I argue that we should critically examine the assumptions that underlie the challenge. Such an examination reveals that the examples smuggle in substantive assumptions that the internalist has no reason to accept. This argument has two important implications for the debate in moral motivation: first, it reveals that the motivational externalist needs a new argumentative strategy; and second, it shows that there is nothing especially problematic about a formulation of the thesis that captures the core internalist intuition that first-person moral judgements are necessarily accompanied by motivation. (shrink)

What are the philosophical views of professional philosophers, and how do these views change over time? The 2020 PhilPapers Survey surveyed around 2000 philosophers on 100 philosophical questions. The results provide a snapshot of the state of some central debates in philosophy, reveal correlations and demographic effects involving philosophers' views, and reveal some changes in philosophers' views over the last decade.

This paper discusses the relation between intelligence and motivation in artificial agents, developing and briefly arguing for two theses. The first, the orthogonality thesis, holds (with some caveats) that intelligence and final goals (purposes) are orthogonal axes along which possible artificial intellects can freely vary—more or less any level of intelligence could be combined with more or less any final goal. The second, the instrumental convergence thesis, holds that as long as they possess a sufficient level of intelligence, agents having (...) any of a wide range of final goals will pursue similar intermediary goals because they have instrumental reasons to do so. In combination, the two theses help us understand the possible range of behavior of superintelligent agents, and they point to some potential dangers in building such an agent. (shrink)

Reviews work on moral judgment motivational internalism from the last two decades.

The standard argument to the conclusion that artificial intelligence (AI) constitutes an existential risk for the human species uses two premises: (1) AI may reach superintelligent levels, at which point we humans lose control (the ‘singularity claim’); (2) Any level of intelligence can go along with any goal (the ‘orthogonality thesis’). We find that the singularity claim requires a notion of ‘general intelligence’, while the orthogonality thesis requires a notion of ‘instrumental intelligence’. If this interpretation is correct, they cannot be (...) joined as premises and the argument for the existential risk of AI turns out invalid. If the interpretation is incorrect and both premises use the same notion of intelligence, then at least one of the premises is false and the orthogonality thesis remains itself orthogonal to the argument to existential risk from AI. In either case, the standard argument for existential risk from AI is not sound.—Having said that, there remains a risk of instrumental AI to cause very significant damage if designed or used badly, though this is not due to superintelligence or a singularity. (shrink)

An artificial general intelligence (AGI) might have an instrumental drive to modify its utility function to improve its ability to cooperate, bargain, promise, threaten, and resist and engage in blackmail. Such an AGI would necessarily have a utility function that was at least partially observable and that was influenced by how other agents chose to interact with it. This instrumental drive would conflict with the strong orthogonality thesis since the modifications would be influenced by the AGI’s intelligence. AGIs in highly (...) competitive environments might converge to having nearly the same utility function, one optimized to favorably influencing other agents through game theory. Nothing in our analysis weakens arguments concerning the risks of AGI. (shrink)