This article evaluates the evidence for sentience - the capacity to have feelings - in cephalopod molluscs: octopus, cuttlefish, squid, and nautilus. Our framework includes eight criteria, covering both whether the animal's nervous system could support sentience and whether their behaviour indicates sentience. There is strong evidence of sentience in octopuses and cuttlefish, which are assessed with very high or high confidence in meeting six of eight criteria. There is also substantial evidence for squid (very high or high confidence in five of eight criteria). By contrast, whether nautiluses are sentient remains unknown (high confidence in only one of eight criteria), since this group of cephalopods have attracted little research. This reflects a general pattern: cases where a taxon did not satisfy a criterion were invariably due to insufficient evidence, rather than evidence that the criterion was not met. In no cases were we confident that a taxon failed a criterion. We explore the nuances of evidence for sentience, examining both neural and behavioural markers, drawing from and updating our previous review (Birch et al., 2021), and shedding light on the implications for ethical treatment and welfare within this class of animals while also revealing areas where further research is needed.

In a vision-first story, conscious vision evolved before other kinds of conscious experience. This can be contrasted with an olfaction-first story, in which conscious vision co-opted integrative mechanisms that first evolved for modeling the causes of olfactory stimuli. An olfaction-first story makes good sense of the connection between consciousness and holistic integration across temporal windows in the ∼400ms range.

Some patients, following brain injury, do not outwardly respond to spoken commands, yet show patterns of brain activity that indicate responsiveness. This is 'cognitive-motor dissociation' (CMD). Recent research has used machine learning to diagnose CMD from electroencephalogram recordings. These techniques have high false discovery rates, raising a serious problem of inductive risk. It is no solution to communicate the false discovery rates directly to the patient's family, because this information may confuse, alarm and mislead. Instead, we need a procedure for generating case-specific probabilistic assessments that can be communicated clearly. This article constructs a possible procedure with three key elements: (1) A shift from categorical 'responding or not' assessments to degrees of evidence; (2) The use of patient-centred priors to convert degrees of evidence to probabilistic assessments; and (3) The use of standardised probability yardsticks to convey those assessments as clearly as possible.

An emerging field shows how animal feelings can be studied scientifically.

There is increasing recognition that the welfare needs of cephalopod molluscs and decapod crustaceans are important. Current commercial practices involving these animals include a range of potential threats to their welfare, such as conditions of farming, capture, transport, and slaughter. This article draws from and updates our 2021 review for the UK Government, recommending a range of relatively simple and impactful changes that could benefit welfare while highlighting important research gaps that should be prioritised to facilitate the drafting of guidelines for best-practice.

The edge of sentience is a book about what we should do when a decision problem forces us to draw a pragmatic line between the sentient and the non‐sentient. The commentaries from Wandrey and Halina and from Bayne analyse both my precautionary framework and my claims about specific cases. My responses start with the case of large language models (LLMs), zoom out to general issues about the science–policy relationship, then zoom in again on the case of fetuses.

How will society respond to the idea that artificial intelligence (AI) could be conscious? Drawing on lessons from perceptions of animal consciousness, we highlight psychological, social, and economic factors that shape perceptions of AI consciousness. These insights can inform emerging debates about AI moral status, ethical treatment, and future policy.

Motivational trade-off behaviours, where an organism behaves as if flexibly weighing up an opportunity for reward against a risk of injury, are often regarded as evidence that the organism has valenced experiences like pain. This type of evidence has been influential in shifting opinion regarding crabs and insects. Critics note that (i) the precise links between trade-offs and consciousness are not fully known; (ii) simple trade-offs are evinced by the nematode worm Caenorhabditis elegans, mediated by a mechanism plausibly too simple to support conscious experience; (iii) pain can sometimes interfere with rather than support making trade-offs rationally. However, rather than undermining trade-off evidence in general, such cases show that the nature of the trade-off, and its underlying neural substrate, matter. We investigate precisely how.This article is part of the theme issue 'Evolutionary functions of consciousness'.

Rapid advances in artificial intelligence (AI) have led users to believe that systems such as large language models (LLMs) have mental states, including the capacity for ‘experience’ (e.g., emotions and consciousness). These folk-psychological attributions often diverge from expert opinion and are distinct from attributions of ‘intelligence’ (e.g., reasoning, planning), and yet may affect trust in AI systems. While past work provides some support for a link between anthropomorphism and trust, the impact of attributions of consciousness and other aspects of mentality on user trust remains unclear. We explored this in a preregistered experiment (N = 410) in which participants rated the capacity of an LLM to exhibit consciousness and a variety of other mental states. They then completed a decision-making task where they could revise their choices based on the advice of an LLM. Bayesian analyses revealed strong evidence against a positive correlation between attributions of consciousness and advice-taking; indeed, a dimension of mental states related to experience showed a negative relationship with advice-taking, while attributions of intelligence were strongly correlated with advice acceptance. These findings highlight how users’ attitudes and behaviours are shaped by sophisticated intuitions about the capacities of LLMs—with different aspects of mental state attribution predicting people’s trust in these systems.

We often face grave practical decisions that seem to hinge on whether a system is sentient. This family of cases includes invertebrate animals, people who are unresponsive after brain injury, fetuses, neural organoids, and now AI technologies. We must decide what to do despite ongoing disagreement about the nature of sentience. In our state of uncertainty, we should pragmatically transform the question from “Is it sentient?” to “Is it a sentience candidate, an investigation priority, or neither?”. When a system is a sentience candidate, it is negligent to fail to consider precautions. We should instead evaluate precautions for their proportionality using the “PARC tests”. When we think about the problems in this way, we see that overconfidence about the absence of sentience has repeatedly led decision makers to neglect serious risks. Erring on the side of caution requires many revisions to current practice in many areas of human activity.

Rapid progress in artificial intelligence (AI) capabilities has drawn fresh attention to the prospect of consciousness in AI. There is an urgent need for rigorous methods to assess AI systems for consciousness, but significant uncertainty about relevant issues in consciousness science. We present a method for assessing AI systems for consciousness that involves exploring what follows from existing or future neuroscientific theories of consciousness. Indicators derived from such theories can be used to inform credences about whether particular AI systems are conscious. This method allows us to make meaningful progress because some influential theories of consciousness, notably including computational functionalist theories, have implications for AI that can be investigated empirically.

Sometimes a person, after brain injury, displays sleep-wake cycles but has severely impaired, or entirely absent, responses to external stimuli. Traditionally, attempts have been made to distinguish the persistent vegetative state (PVS) from the minimally conscious state (MCS). However, diagnostic procedures are subject to high error rates and high uncertainty. There is also a realistic possibility that midbrain mechanisms suffice for basic valenced experiences even if cortical injury fully prevents a patient from reporting these experiences. Decisions to withdraw treatment should be based on comprehensive best-interests assessment, not on the PVS/MCS distinction. The method of withdrawing clinically assisted nutrition and hydration (CANH) would not be acceptable for any other sentient being, and alternatives must be explored and discussed by inclusive, democratic processes. In cases where CANH withdrawal has been authorized, clinicians should be guaranteed that hastening death using large doses of sedatives or analgesics will not lead to punishment.

Abstract We should not be complacent about the risks of developing sentient AI in the near future. Large language models (LLMs) already present some risk. Three other pathways to artificial sentience candidates are also worth taking seriously. The first involves emulating the brains of sentience candidates such as insects, neuron by neuron. The resulting virtual brains are sentience candidates if they display the same pattern of behavioural markers that we take as sufficient for sentience candidature in the biological original. A second path involves evolving artificial agents that converge on similar patterns of behavioural markers to biological sentience candidates. A third involves deliberately implementing a minimal version of a large-scale computational feature credibly linked to sentience in humans. All three pathways present ways in which we might come to recognize a system as an artificial sentience candidate. We must be mindful of the possibility of significant decouplings of sentience from intelligence in this area.

Abstract How could a citizens’ panel reach an informed judgement about proportionality? This chapter describes a possible procedure (intended as a realistic, feasible ideal) based on a pragmatic analysis of proportionality. The panel is presented with a shortlist of feasible options on which stakeholders have been openly consulted. To each policy option, the panel applies four tests in sequence: permissibility-in-principle, adequacy, reasonable necessity, and consistency. Proposals that fail a test are set aside. Proposals that pass all four of the ‘PARC tests’ are judged proportionate. The PARC tests induce a division of labour between the panel and its expert advisers. At each stage, the expert advisers provide on-demand input regarding the likely consequences of different policy options, but it falls to ordinary citizens to debate the central evaluative questions. These questions can be easily understood and do not require arbitrating scientific disagreements. Although a government is the ideal implementing agent for such a process, other organizations can conduct similar exercises.

Abstract Given the rate at which AI is developing, and the risks associated with artificial sentience taking us by surprise, we should apply the run-ahead principle: at any given time, measures to regulate the development of sentient AI should run ahead of what would be proportionate to the risks posed by current technology, considering also the risks posed by credible future trajectories. The run-ahead principle may potentially justify strong regulatory action, but a moratorium may go beyond what is reasonably necessary to manage risk. An alternative proposal, involving regular testing to monitor the sentience of our AI creations, is currently unfeasible, due to the absence of tests that can be applied to large language models and other systems with high potential for gaming our criteria. A third approach involves oversight by means of sector-wide codes of good practice and licensing schemes. This path would require a greater level of transparency than we have seen from the AI industry to date. The overarching imperative is to have democratic debate about these questions now.

Abstract ‘Sentientist’ ethical outlooks regard sentience as necessary and sufficient for having interests that matter morally in their own right. Sentientism finds expression in at least three major secular ethical theories (classical utilitarianism and the theories of Korsgaard and Nussbaum), as well as in the idea of ahimsa in Indian thought. Sentientism can be contrasted with various ways of denying the necessity and/or sufficiency of sentience for moral status. The possibility of Vulcan-like beings who have the consciousness aspect of sentience without the valence aspect suggests a qualification to pure sentientism may be needed. A more serious challenge comes from agency-centric and rationality-centric positions. One example is orthodox Kantianism, which allows only indirect duties (formally owed to ourselves) concerning non-rational beings. Another challenge comes from the Abrahamic religions, which give only very limited moral standing to non-human sentient beings. We can, however, find in all of them support for duties of stewardship, including the duty to avoid causing gratuitous suffering.

How should proportionality be assessed in practice? A ‘tyranny of expert values’ occurs when the values of expert advisers determine a policy decision without those values being properly scrutinized by a democratic process. Citizens’ assemblies or panels can be an attractive way to avoid this problem. Moreover, they have advantages over elected assemblies and referendums. These advantages are especially clear when an issue generates deep value conflicts, requires sustained attention and regular revisiting, requires consideration of the interests of beings who cannot vote, and when there are reasons to departisanize the issue. Questions of proportionality at the edge of sentience have all of these properties. Since citizens do not generally have scientific training, careful thought needs to be given to the structure of deliberation, so that they are not forced into a position of arbitrating scientific disagreement. Their focus should be on whether or not a proposed response can be publicly justified as proportionate, not on whether a being is a sentience candidate.

Abstract Reasonable disagreement about sentience requires responsiveness to evidence and argument. It excludes baseless recommendations, dogmatic adherence to refuted theories, and morally abhorrent (e.g. sadistic) positions. However, the uncertainty in this area is such that many very different positions can be held by reasonable people. This chapter examines sources of disagreement that have their origins in the philosophy of mind. Major metaphysical pictures including materialism, epiphenomenalism, interactionism, Russellian monism, biopsychism, and the ‘integrated information theory’ are introduced and their major strengths and weaknesses are considered. The chapter then turns to other axes of disagreement. One concerns the importance of agency and embodiment, real or virtual. Another concerns the scale of functional organization that matters. A third concerns whether the edge of sentience is sharp or blurred.

This chapter turns to sources of uncertainty in the science of consciousness and emotion. To have a science of consciousness at all, we need reliable ways of disentangling conscious and unconscious processing. In the case of vision, long-running debates about blindsight epitomize two major problems: the criterion problem and the problem of confounders. These problems arise even more strongly in the case of valenced experience, since methods for eliciting unconscious analogues of valenced experiences are less mature. In the absence of secure ways of dissociating valenced experience from its unconscious analogues, two rival pictures of the neural basis of valenced experience are likely to persist. On one picture, valenced experience wells up directly from subcortical mechanisms without the need for further cortical processing. On the other, subcortical circuits produce coordinated behavioural responses, but conscious experience only comes with cortical involvement. Unfortunately, current evidence does not allow us to choose confidently between these pictures.