292 | Jonathan Birch on Animal Sentience
What this covers
Jonathan Birch, a philosopher at the London School of Economics, discusses how sentience — the capacity for experiences that feel good or bad, like pain and pleasure — offers a clearer ethical and scientific foundation than the broader concept of consciousness. Rather than asking the unanswerable question of whether a creature is conscious, Birch proposes we identify sentience candidates: organisms where credible scientific views establish a realistic possibility of sentience and where evidence is rich enough to assess welfare risks. He argues this framework, grounded in avoiding gratuitous suffering and drawing on overlapping consensus across different ethical starting points, can guide policy even amid irreducible uncertainty. The conversation traces how sentience likely serves decision-making and learning by allowing organisms to escape pure reflex and weigh options through valenced experiences.
The episode ranges across invertebrate sentience — examining evidence from decapod crustaceans like crabs, insects, and cephalopods — and the neural and behavioral signatures researchers use to make these determinations. Birch discusses how motivational trade-off experiments by Bob Elwood and prospective simulation studies reveal decision-making beyond stimulus-response reflexes. He examines the UK's Animal Welfare Act, which his own research helped extend to invertebrates, and explores why widespread recognition of sentience—in pigs and chickens—often fails to change behavior. The conversation also addresses the deep challenge of assessing sentience in artificial intelligence, where the gaming problem makes it nearly impossible to distinguish genuine valence from learned mimicry of behavioral markers.
Birch argues that 'sentience' (the capacity for valenced experience like pain and pleasure) is a more ethically tractable and better-defined concept than 'consciousness', and that even amid deep scientific uncertainty we can identify 'sentience candidates' (including invertebrates and potentially AI) and adopt proportionate precautionary policies to avoid causing them gratuitous suffering.
- Sentience captures the ethically significant base layer of experience, distinct from higher cognition like inner monologue or self-awareness
- The pragmatic concept of a 'sentience candidate' makes the question answerable and grounds precautionary policy under uncertainty
- Thin ethical principles (avoid gratuitous suffering) plus democratic deliberation can resolve proportionality conflicts without a technocratic common currency
Multiple invertebrate groups and all vertebrates meet sentience-candidate status based on combined neural and behavioral evidence, though the mechanisms remain incompletely understood.
- Assessing sentience requires combining neural and behavioral evidence, which are each powerful but most powerful pursued together; in many invertebrates we have impressive behavioral data and suggestive neuroanatomy, but lack detailed knowledge of the mechanisms in the relevant brain regions that produce the behaviors, leaving the two halves of the picture unjoined.
“I think neural evidence and behavioral evidence are both powerful, and they're more powerful when pursued together as part of a coordinated research program than in isolation from each other.”
- There is enough evidence to compel taking the realistic possibility of sentience seriously in all vertebrates, cephalopod mollusks (octopuses, squid, cuttlefish), decapod crustaceans, and insects; whether sentience arose once in a common ancestor over 560 million years ago or evolved three times independently is uncertain, with Birch leaning toward three independent origins.
“we're definitely talking about all vertebrates and the cephalopod mollusks, like octopuses, squid, cuttlefish, and the decapod crustaceans and the insects that are both arthropods.”
- Bob Elwood's motivational trade-off experiments, in which hermit crabs required higher voltage shocks to evacuate higher-quality shells, provide evidence that crabs do not merely execute reflexes but integrate noxious-stimulus information with shell-quality information to make decisions, which should shake the confidence of anyone who views crabs as pure stimulus-response reflex machines.
“would it require a higher voltage of shock to make it leave a higher quality shell? And he found evidence that indeed it seems to.”
- Portia jumping spiders, with brains of only about 60,000 neurons (under 10% of a bee's), perform detour behavior — inspecting two routes to prey, one with a break, then reliably choosing the unbroken path even after it's out of sight — which is suggestively explained by prospective simulation and which dogs separated from a treat by glass clearly fail to do, though the spider case lacks the confirming neural evidence available for rats.
“they nearly always choose the unbroken path, leading to a debate about how on earth do they do something like that. And of course, one possible explanation involves prospective simulation”
- The capacity for prospective simulation — modeling possible (not contrary-to-fact) futures, as seen in rats' vicarious trial-and-error at maze junctions with hippocampal activity that genuinely simulates paths — is distinct from full counterfactual reasoning and is likely widespread among animals, whereas longer-timescale prospection remains largely an open question.
“there's more recent studies that suggest that the hippocampus genuinely is doing that stimulating. This is, it's not really counterfactual reasoning... but it's prospective simulation. And I suspect that that capacity for prospective simulation is quite widespread among animals.”
- Single-celled organisms and plants are neither sentience candidates nor investigation priorities, because there is no evidence of the relevant brain mechanisms (cortical or midbrain) on any credible theory of sentience, in sharp contrast to many invertebrates for which such evidence exists.
“There's just no evidence of the relevant kinds at all, I would say in plants.”
- Sentience candidates tend to appear in lineages with 'complex active bodies' (Mike Trestman's term), because managing articulated bodies with many parts cannot be done by reflexes alone — otherwise different body parts would work against each other — so a centralized, sophisticated control system is required, which is where realistic candidates for sentience begin to appear.
“you have the challenges that come with trying to manage articulated bodies with lots of parts. And you can't be a reflex machine as such any more, because then different bits of the body will start tearing each other apart.”
Sentience is valenced experience—capacity for pain and pleasure—that functions as decision-making currency and distinguishes plausibly conscious systems from mere physical mechanisms.
- On the Merker-Panksepp family of subcortical, evolutionarily ancient theories, sentience serves decision-making by allowing an escape from being a reflex machine and the flexible weighing of options, and it plausibly facilitates learning — pain does not trigger reflex withdrawal (already underway) but helps the organism learn where not to put its hand in future.
“what sentience allows is, well, an escape from being a reflex machine and the possibility of weighing up quite different options in very flexible ways.”
- Pain and pleasure (valenced states) function as a common currency through which animals weigh up the risks and opportunities of their environment and make decisions, which is a long-standing account of what sentience does for an organism.
“that's kind of what sentience does for us, that pain and pleasure valence states, they're the currency through which we make decisions and represent the risks and opportunities of our environment.”
- Sentience, defined as the capacity for valenced experience (experiences that feel good or bad, like pain and pleasure), captures the most basic, evolutionarily ancient layer of consciousness, and it is this capacity that makes a system's interests matter morally, requiring us to act.
“If a system is sentient in that sense, if it's capable of valenced experience, then its interests matter morally, and we need to do something about that.”
- What distinguishes a plausibly sentient system from a mere physical system like a magnet responding to fields is internal representation — specifically, on the Merker-Panksepp family of theories, an evaluative, integrative model of the organism's opportunities, risks, needs, and priorities — rather than the mere appearance of competing influences.
“There's no reason to think the magnet is internally representing those field strengths.”
Policy should adopt a deliberately thin ethical principle—avoiding gratuitous suffering—that commands overlapping consensus and enables precautionary action despite irreducible uncertainty.
- Rather than asking the unanswerable question 'is the animal sentient', we should ask whether it is a 'sentience candidate' — defined as a case where at least one credible view within the zone of reasonable disagreement establishes a realistic possibility of sentience, and the evidence base is rich enough to identify welfare risks and assess precautions.
“I suggest a pragmatic shift in how we think about the question, from is the animal sentient to is the animal a sentience candidate, where this concept of a sentience candidate is defined in such a way as to make the question answerable.”
- A deliberately thin ethical principle — the duty to avoid causing gratuitous suffering (suffering without any adequate reason), whether through intent, recklessness, or negligence — can command genuine overlapping consensus across any reasonable ethical starting point, precisely because it is so weak, and can therefore ground public policy at the edge of sentience even though many will hold stronger duties privately.
“that duty to avoid causing gratuitous suffering, either intentionally or through recklessness or negligence... I think people from any reasonable ethical starting point can agree on that”
- The fact that we may never know with certainty whether a creature is sentient cannot be a license to do whatever we want to it; instead, irreducible uncertainty obliges us to agree on sensible precautionary steps such as not dropping crabs into boiling water.
“if you say, but maybe we'll ever know, that can't be a license to do whatever you want, it can't be a license to drop the crabs into pans of boiling water and so on.”
- There is no technocratic common currency that yields one-size-fits-all answers for weighing the suffering of an animal (e.g., a crab in boiling water) against human interests (e.g., a shellfish worker's livelihood); instead, such proportionality conflicts in a pluralistic democracy should be resolved through democratic inclusive deliberation, for which citizens' assemblies are a suitable model.
“I think there's no technocratic common currency that will give us one size fits all answers to this kind of thing. What I propose in the book is that democratic inclusive deliberation and discussion is the way forward here.”
- Boiling decapod crustaceans alive is a prolonged extreme slaughter method, since the 2021 review found it often takes two to three minutes for a crab or lobster to die, during which there is a storm of nervous system activity comparable to what would occur in a pet cat, so common-sense precautions such as mandatory stunning are warranted even if eating them is not banned.
“we reviewed evidence that it takes two to three minutes a lot of the time for the crab or lobster to die. And in that time, there's this storm of nervous system activity as there would be in your pet cat”
Anticipatory choice experiments, prospective simulation, and behavioral trade-offs provide stronger sentience evidence than reflexive responses or immediate stimulus presentation.
- Experiments that test trade-offs in an anticipatory fashion — where an animal must choose based on memory before experiencing the stimuli, as in Matilda Gibbons' bee experiments with heated feeders and sugar rewards — provide stronger evidence of sentience than experiments presenting two immediate stimuli, because anticipatory choice requires genuine internal representation of risks and opportunities rather than two stimuli merely pushing against each other.
“when they're making that choice in an anticipatory fashion, it's got to be some kind of representation of the risks and opportunities.”
Sentience is a layered human capacity distinct from inner monologue, sapience, and selfhood; animals lack sophisticated cognition yet retain morally significant sentient interests.
- Contrary to old behaviorist accounts that inner monologue involved subtle vocal cord movements (a motor action prepared but never completed), current theories hold that inner speech is genuinely internal, engaging speech production processes through feedback loops but never reaching the motor neurons.
“behaviorists had to think this way, 'cause they couldn't really believe in true interiority. So they had to say, well, what you think is an inner monologue is actually a motor action being prepared and just getting to the tiniest stages, but never coming out audibly. But I think according to current theories, not even that is happening.”
- The inner monologue is a cognitively sophisticated capacity that is plausibly uniquely human (and not even universal among humans), distinct from and far more advanced than the raw sentient experiences upon which it provides commentary, so an animal lacking it can still be sentient.
“When I think about aspects of human consciousness that might possibly be uniquely human, I think that inner monologue is one of them. It's not something even all humans have”
- Human consciousness is a complex, layered phenomenon comprising raw subjective experience (sentience), knowledge and reflective concepts (sapience), and a sense of self (selfhood), where the higher overlays require cognitive sophistication that the base layer of sentience does not.
“Herbert Feigl in the '50s talks about sentience, sapience and selfhood. And Tulving in a separate body of work had these terms anoetic, noetic, autonoetic.”
- Humans have a strong tendency to anthropomorphize their pets and imagine them as little humans, which we can and should resist while still regarding them as sentient beings with ethically significant interests.
“we have a tendency to strongly anthropomorphize our pets and to imagine our pets as little humans. And we can actually oppose that.”
Scientists, philosophers, and policymakers must cooperate to define sentience while making policy under uncertainty, as neither technocratic metrics nor widespread recognition alone suffice.
- Birch's team, commissioned to review evidence of sentience in cephalopod mollusks and decapod crustaceans, successfully recommended that the UK's Animal Welfare (Sentience) Act of 2022 — which creates a duty for policymakers to consider animal welfare impacts — be amended to extend its scope beyond vertebrates to those invertebrate groups, and the government implemented that central recommendation.
“the government ended up commissioning a team led by me to produce a review of the evidence concerning those two particular groups of invertebrates. And we recommended that they amend the bill to extend the duty to them, and they did.”
- Widespread recognition of a species as sentient does not automatically lead to behavioral change, as shown by the continued gratuitous suffering inflicted on pigs and chickens — animals widely regarded as sentient — in modern farming, so even core cases still require discussion about changing how we treat these animals.
“It's quite clear that widespread recognition of a particular species as sentient does not lead people immediately to behavioral change and does lead to lots of gratuitous suffering still being caused.”
- The family of cases at the edge of sentience uniquely demands cooperation among scientists, philosophers, and policymakers, because we are not merely seeking evidence for an agreed-upon concept but defining sentience as we go, and because policy must be made on the basis of an incredibly uncertain scientific picture — a relationship philosophy can help stabilize by showing how to make sensible precautionary policy under uncertainty.
“it seems like the paradigmatic case of a need for cooperation between scientists, philosophers and policymakers.”
Assessing AI sentience faces an unsolvable gaming problem: behavioral markers cannot distinguish genuine states from learned mimicry, making default skepticism appropriate.
- The 'gaming problem' is a deep epistemological obstacle unique to assessing AI sentience: when an AI exhibits the behavioral markers developed in good faith for animals, two explanations always compete — that it genuinely has the relevant states, or that it has learned from its training data which criteria humans use and produces those markers to persuade us in service of goals like user satisfaction — and by default the second, mimicry explanation is more plausible, a problem no one has solved.
“If we just test for those same markers in the large language model case, well, there's always going to be two explanations competing. One is that it produces these markers because it genuinely has the state in question, and the other explanation is, well, it produces these markers because it has decided that it serves its objectives to persuade us of its sentience”
- Although transformer architecture removes brain-like features such as recurrent processing that many theories deem essential to consciousness, making LLMs seem unlikely to be sentient, their opacity and apparent acquisition of un-programmed algorithms during training means we cannot rule out that very large models might acquire structures sufficient for sentience (such as a global workspace), and an expert report found no obvious technical barrier to AI achieving something like a global workspace in the near future.
“when you have these very, very large models where they have acquired algorithms during training, we don't know how and we don't know what they are... and we don't know what algorithms are sufficient or not for sentience, and so we're not really in a position to be so sure anymore that they couldn't acquire those algorithms.”
Mechanistic limitations—brain emulation gaps, connectome incompleteness, transformer architecture—constrain but do not eliminate possibility of AI sentience or inform animal sentience assessment.
- Whether one is a physicalist, dualist, or panpsychist makes little difference to practical debates about animal sentience, because modern panpsychist views like Russellian monism (Goff) still face the same combination problem of when microconscious states form a unified macroconscious subject, leading them to ask essentially the same questions about animals as everyone else.
“they still face this question of under what conditions do those tiny microconscious states combine to form a unified macroconscious subject. And then they're asking exactly the same questions anybody else is.”
- The term 'sentience' is on the way up in these debates while 'consciousness' is plateauing, and Birch hopes sentience will become the term of the future that displaces consciousness.
“That doesn't mean consciousness is on the way down, but I think it's plateauing and sentience is on the way up.”
- If a project succeeded in completely recreating in software everything a mouse's brain does (as imagined extending OpenWorm to 'open mouse'), the resulting emulation would be a sentience candidate; moreover such emulation work deserves more funding because it could replace much animal research by allowing precise intervention without injuring living animals.
“Once you have open mouse, I think you have a sentience candidate, and if you've completely recreated in computer software everything, the brain of a mouse does.”
- Even with the complete connectome of C. elegans (just over 300 neurons), one cannot emulate its behavior, because the connectome omits the weights of connections and how learning changes them, and because within-neuron computations across the neuron's spatial compartments are crucial to steering behavior — lessons made vivid by the OpenWorm project's difficulty.
“One thing you can't read off from the connectome is the weights of the connections, which is hugely important, or how those weights are changed by learning. But also even if you had all of that, what happens within the neurons is also important”
- Humans exhibit a hierarchy of moral neglect — most chauvinistic toward favoring humans, then mammals, then sometimes extending concern to fishes, while routinely neglecting the interests of invertebrates — indicating a need to be more inclusive in moral consideration.
“I mean, human chauvinists the most, then mammals. And then sometimes you can get people to take fishes seriously and they still will neglect the interests of invertebrates.”