Sentience Part 2 - The Edge of Sentience

Apr 27, 2025

Reasonable disagreement

In the previous essay I spoke about the work of Jonathan Birch, one of the authors of the New York Declaration on Animal Consciousness. Birch has since released a book on sentience. The Edge of Sentience¹ (EoS) gives a great overview of the meaning of sentience and how we can detect it. But more importantly, it considers what to do about it. In the absence of good evidence for sentience, it's easy to convince yourself there's nothing to worry about. But there are urgent concerns about causing suffering in sentient beings, right now and in the near future. Pigs who spend their whole lives in indoor pens bite each other's tails out of boredom2. We already have AIs which have convinced their creators that they're sentient3. And in the future we may be able to create neuron-by-neuron simulations of animal brains that exist only on a computer.

For Birch, sentience is "the capacity to have valenced experiences, such as experiences of pain and pleasure." This requires both consciousness (the ability to have subjective experience) and valence - that those experiences can be good or bad. The scientific basis for consciousness gets a thorough treatment, with good explanations of Global Workspace Theory and midbrain-centric ideas of consciousness like we saw in Part 1, as well as weirder ideas like Integrated Information Theory and Russellian Monism. Birch doesn't gather the theories together to pick a winner (although he does seem to lean towards midbrain-centric theories). The point is to map out the zone of reasonable disagreement - the set of theories that at least have some reasonable chance of being correct. And for Birch, if there is reasonable disagreement about a creature's sentience, we have to have consideration for that creature's wellbeing - be they an animal, a human fetus, an organoid or a simulation.

This doesn't mean we all have to be vegans, or Jainist ascetics sweeping the ground before us lest we step on an insect. But at a minimum we have to avoid causing gratuitous suffering. You might think that's a low bar. But Birch has a few horrifying examples where that bar wasn't cleared.

Kate Bainbridge4 was left in a persistent vegetative state after suffering severe encephalitis. She had a feeding tube inserted and had her mouth suctioned regularly, while her caregivers assumed she was unconscious. But they were wrong. When Bainbridge recovered enough function to type on a keyboard, she revealed that she had been fully aware, but paralyzed. She didn't know why she couldn't eat or drink (and she was constantly thirsty), as no-one had told her. And the suction was agony. As she writes, "I tried to hold my breath to get away from all the pain".

Bainbridge was so damaged by the encephalitis that she could not even direct her gaze to indicate awareness. It's understandable that no-one realized she was conscious. But Birch's message is that where there is uncertainty, we must take proportionate action. In Bainbridge's case, this might have meant taking a few seconds to tell her what procedure she was undergoing, and using the same sedation or anaesthesia that would be used on any other patient.

No-one has ever reported a painful memory from when they were an infant. But should we be confident that newborns can't feel pain? Anand & Hickey wrote in 19875 that "newborns are frequently not given analgesic or anaesthetic during invasive procedures, including surgery". In fact, Birch details how their article (and a campaign by a mother who was horrified to learn that her newborn was operated on without anaesthetic) began a debate in the field, which eventually led to widespread acceptance of the need for precautionary pain management for newborns.

There is certainly room for doubt around newborns consciousness (or level of consciousness). But for a surgical team to assume an infant is effectively unconscious before making an incision seems careless to the point of callousness. In fact, recent work6 suggests that infants can form memories (for which consciousness is a prerequisite). So-called "infantile amnesia" may be caused by the subsequent loss of the retrieval path to the memory.

Implications for animals

So how does the Birch's precautionary approach map onto animal welfare? First, we must have a map of the zone of reasonable disagreement on animal sentience. Conveniently Birch provides this in emoji form!

👧 👶 << 🐒 🐀 🐦 🐟 🐝 >> 🧶 🧫 8

This shows a spectrum of views about the minimum requirements for sentience, from strictest to loosest. Birch's zone of reasonable disagreement is inside the << >> brackets. Each view is represented by an emoji for the simplest animal (or thing) that could be sentient, based on that view. Unreasonably strict views are on the far left: that only humans with natural language are sentient (child), or humans with a fully developed prefrontal cortex (baby) . On the far right are the unreasonably inclusive views. The Petri dish represents the view that non-nervous tissue grown in a lab could be sentient. (Lab-grown nerve tissue organoids are a different story, and Birch devotes a whole chapter to the debate on these.) And the ball of yarn is meant to be the cerebellum, but that doesn't have its own emoji! This represents the view that a cerebellum is sufficient for sentience. The cerebellum is a large and densely-connected structure, but no-one seriously believes it has a role in sentience. For one thing, there is at least one9 case of a person living an almost normal life without one!

Let's look within the zone of reasonable disagreement, and specifically at the rightmost view within the zone (represented by a bee) - the view that a midbrain or something functionally equivalent suffices for sentience. This is the view of Barron & Klein that I mentioned briefly in the previous essay. The vertebrate midbrain plays a role in integrating different sensory inputs and memory areas. This view sees the insect central complex and decapod crustacean hemiellipsoid bodies as functionally similar enough to the midbrain to also confer sentience.

The decapods consists of (among others) crabs, lobsters, shrimp & prawns. (The terms shrimp and prawn are used interchangeably depending on where you live, but strictly speaking they are separate groups within the decapods.) In Birch's view, there is reasonable disagreement about shrimp sentience, so we need to take reasonable measures for shrimp welfare. Now I suspect the mention of shrimp welfare is going to set some eyes rolling. As Andrés Jiménez Zorrilla puts it:

When I first read [a] proposal for a shrimp welfare charity, I thought: "Effective altruists have gone mad — who cares about shrimp?"10

But Zorilla ended up leaving his private equity job to found the Shrimp Welfare Project! They campaign for better living conditions for farmed shrimp and electric stunning before slaughter. Their argument is that the number of farmed shrimp is vast - possibly exceeding the number of every other farmed animal combined. So the improvement in total animal wellbeing from improving the average shrimp's quality of life is huge.

Assuming of course, they are sentient. I suspect the normie view is that they aren't. Personally, I grant they have some non-zero level of sentience, rather than being entirely meat robots. And I suppose I lightly agree that we should avoid causing gratuitous suffering. I'm not confident enough to support legal protection for shrimp. But I was surprised to learn that shrimp welfare is already law in the UK, partly due to Birch's work!

Birch led a team which advised the UK government when they were drafting the Animal Welfare (Sentience) Act 2022. The team recommended that the legislation should cover all decapods, and this was implemented. It's not clear how strictly the rules are applied, but by law scientists have to get Ethical Review Board approval if their experiments include shrimp or other decapods.

Sentience is a continuous variable

Here I want to point out something that seems obvious to me but isn't often stated: That sentience isn't all or nothing, but varies continuously. Humans are more sentient than cattle, cattle more than pigeons, pigeons more than bees. To maintain otherwise is Richard Dawkins' Tyranny of The Discontinuous Mind. Birch does consider the idea of a gradual evolutionary transition in a species from non-sentience to sentience11. But for him, once a species joins the sentience club, it has the same status as everyone else. Birch admits that the "true" prawns in the Dendrobranchiata sub-order have a greatly reduced hemiellipsoid body compared to other decapods, and probably aren't sentient. But in the interest of simplicity his team still recommended they be covered by the legislation mentioned above. He speculates about the decapod lineage having gained sentience, and the prawns then losing it, in a similar way to the ancestors of penguins gaining the ability to fly, which penguins subsequently lost. Except that flying really is all-or-nothing, in a way that shouldn't be applied to sentience.

Birch excerpts a diagram12 to show the reduced hemiellipsoid bodies in shrimp. It's actually a great example of the continuity of sentience. The hemiellipsoid bodies vary hugely from group to group, from relatively huge in the true crabs to medium in shrimp and practically non-existent in prawns. So if hemiellipsoid size really corresponds to sentience it's actually finely graded even within the group.

Sentience-Adjusted Lives of Suffering

The continuity of sentience really matters when we start adding up the numbers of sentient beings. Zorilla says:

Why worry about shrimp in a world where so many mammals and birds live in torturous conditions due to industrial agriculture? The answer is that shrimp farming dwarfs other forms of animal agriculture by sheer numbers. An estimated 230 billion shrimp of various species are alive in farms at any given moment — compared to the 779 million pigs, 1.55 billion cattle, 33 billion chickens, and 125 billion farmed fish.

I think we need to do a bit more than simple addition to get an idea of the harms involved. I'm thinking of something based on the idea of DALYs - Disability-Adjusted Life Years14. DALYs are used to measure the total burden of a disease. Let's take COVID-19 as an example. Firstly, instead of just adding up the number of people killed by it, we estimate the number of life years lost. For each death we subtract the age at death from their estimated life expectancy (75 years is often used as a standard). So the death of a 35-year-old is 40 years lost, but the death of a 70-year-old is only 5 years lost.

Then we add a term for people who aren't killed by the disease but suffer a disability which affects their quality of life. Each disability is assigned a weighting between 0 (no effect) and 1 (as bad as death). Wyper et al's study15 on DALYs in Scotland from COVID put a weighting of 0.219 on long-COVID-like post-acute symptoms. Then you add up all the life years lost, plus the years of disease multiplied by their weighting. In the Scottish study they got a total of around 100,000 DALYs lost to COVID in 2020.

I want to come up with something similar which measures the total suffering caused by some activity (like factory farming) which applies a weighting based on the sentience of the beings involved. We also need a term for the level of harm to each being. For now I won't include a time term - we are just interested in the amount of suffering occurring at a point in time. Let's call the measure SALOSes - Sentience-Adjusted Lives of Suffering. We'll assign a sentience weighting of 1 to an adult human. Any other creature has a sentience between 0 and 1. I'm going to take the existence of an enslaved person in the United States in the 19th century as my benchmark for a high level of harm and assign that a value of 1. Slavery involved total confinement and near-daily torture for many, but I suppose worse forms of suffering are conceivable, so I'll allow values greater than 1. The number of SALOSes then is just the number of beings times the sentience weighting times the harm weighting.

Let's take slavery as an example. In 1860, there were around 3.9 million people16 enslaved in the US. By definition our sentience weighting and harm weighting are both 1, so the number of SALOSes caused by slavery at that point in time was 3.9 million.

How about factory farming? Let's try beef cattle in the US. In 2024 there were around 28 million beef cattle alive. For sentience, I'll give cattle a weighting of 0.05, or a twentieth of a human. I'm not firmly attached to that number but it'll do for a start. The harm level is hard to judge. The cattle are at least well fed and not routinely tortured. But I'll bet they are prodded and whacked to get them to move when needed. And they have less space than they would like and can't choose where to go. I'll put it an 0.2 for now. That gives us 28 million x 0.05 x 0.2, which is 280,000 SALOSes. And I think that's a reasonable result. It's not an abomination on the scale of chattel slavery, but it's not nothing either.

Let's try it for shrimp. Picking a sentience weighting here is hard. It's hard to even place shrimp and humans on the same scale when they're so far apart. I'll have to resort to comparing brain volumes. Birch warns against using neuron counts or size measurements as proxies for sentience, but I think it's good enough for a rough estimate. Eyeballing the images in this study17 of Litopenaeus vannamei (the most commonly-farmed species) gives a brain diameter of around 1mm. Approximating to a sphere that gives a brain volume of 4mm^3. The human brain is around 1000 cm^3, 250,000 times larger. So if we assume sentience is proportional to brain volume, shrimp sentience is 1/250,000. Per Zorilla there are 230 billion farmed shrimp currently alive on earth. I really don't know what conditions are like for them, but it's at least going to be overcrowded. Let's put the harm at 0.1. Multiplying those numbers out gives us 92,000 SALOSes for shrimp farming on earth as a whole. The exact value isn't significant, but I think it supports a position that shrimp farming is not a greater cause of suffering than beef farming. US beef production is only about a fifth of world production18, so beef farming worldwide causes around 1.4 million SALOSes.

Reason doth stretch man's mind upon the rack

For my SALOS model I've assumed more sentience corresponds to more suffering at the same level of harm. So kicking a dog is worse than kicking a lobster, which is worse than kicking a robot lawnmower. As we saw in the previous essay, just registering harmful stimuli is not the same thing as suffering. But what is it about being sentient that makes suffering worse? Part of if is the ability to think about one's suffering. I'd imagine if someone kicked a lobster it might feel something like fear as it fled, but then pretty much get on with its day. But a dog can remember being kicked, so now he will flinch whenever a human comes near. And of course, humans are the masters of this multiplication of suffering. Not only do we suffer, we worry if the suffering will ever end. We can imagine worlds without suffering. We are even pained by the suffering of others. The 17th-century poet and philosopher Margaret Cavendish wrote in A Dialogue betwixt Man, and Nature (1653):19

Though Beast hath Sense, feels pain, yet whilst they live,
They Reason want, for to dispute, or grieve,
Beast hath no pain, but what in Sense doth lie,
Nor troubled Thoughts to think how they shall die.
Reason doth stretch man's mind upon the rack,
With Hopes, with Joys, pulled up, with Fear pulled back.

Samuel Taylor Coleridge in Fears in Solitude (1798)20 finds a beautiful dell in the Quantock hills to relax in, but cannot enjoy it due to his fear of a French invasion:

It is indeed a melancholy thing,
And weighs upon the heart, that he must think
What uproar and what strife may now be stirring
This way or that way o'er these silent hills—

The invasion never came to pass. But the threat cast a constant shadow on Coleridge's life in a uniquely human way.

Behaviour and adaptation

The main motivation in EoS for considering the sentience of other beings is the risk of causing them suffering. So understandably much of the behavioural evidence of animal sentience is based on how they respond to pain. But here we run into the robot lawnmower problem from part 1. If you thought it was important, you could program a robot to respond to damage or avoid things that might cause harm. And of course evolution has programmed many animals to do just that! Responding to harm is such a fundamental aspect of survival that we shouldn't be surprised that it has evolved multiple times, even in species that don't have the cognitive architecture for a conscious response. In fact, even humans have such an unconscious response. If you put your hand on a hot surface you will involuntarily withdraw it. This is a reflex pathway which happens without conscious thought22. Your consciousness does register it, but on a parallel pathway that doesn't cause the movement. For an outside observer it's not at all obvious that this response to pain was unconscious.

Perhaps we should look at behaviour that isn't adaptive for evidence of sentience. If a behaviour looks like the product of a sentient mind but doesn't have any adaptive significance, we're less likely to be fooled. I'm pretty convinced by the mourning behaviour seen in elephants. They will spend time picking over the bones of dead elephant, ignoring the bones of other species23:

They cautiously extend their trunks, touching the body gently as if obtaining information. They run their trunk tips along the lower jaw and the tusks and the teeth—the parts that would have been most familiar in life and most touched during greetings—the most individually recognizable parts.

It's uncannily like the Alas, poor Yorick scene where Hamlet examines the skull of a man he once knew:

Alas, poor Yorick! I knew him, Horatio: a fellow
of infinite jest, of most excellent fancy: he hath
borne me on his back a thousand times; and now, how
abhorred in my imagination it is! my gorge rims at
it. Here hung those lips that I have kissed I know
not how oft.

There's nothing adaptive about the elephants' behaviour. It's not like they're scavengers who pick meat off bones. The only reasonable explanation is that it's a kind of mourning, and they are feeling a kind a grief. Evolutionarily, I suppose it's a side-effect of recognizing and caring for other elephants which spills over into death.

Final thoughts

I'd better wrap this up before I write a book about the book. There is much in the book I haven't mentioned, including a whole section on the possible sentience of AI and the welfare of artificial intelligences. I will only remark that if an AI apocalypse comes to pass, I hope they spare the folk who worried about us hurting the AIs.

Normally a review will recommend whether or not to buy the book. No need in this case, as the book is free! Birch has very kindly made it free for anyone to download a PDF copy. Click "Read Now" on edgeofsentience.com. And I very much recommend you do read it. The section on the science of sentience and consciousness is a great introduction to the field. I wish I'd read it before I wrote Part 1. The writing on the ethical considerations is urgent and convincing. The parts about government policy and proposals for Citizens’ Assemblies weren't really for me, but I’d recommend them anyone working in a related policy field.

I ended the previous essay skeptical of detecting sentience from behaviour. In some ways The Edge of Sentience has deepened my skepticism. Birch makes it clear just how difficult it is to guess the mind of a creature who cannot speak. But it's also made it clear that when we affect another being's welfare, we can't risk cruelty where there is a chance that being can suffer.