We waited too long to take animal sentience seriously - and shouldn't do the same with AI.
What if we’d bring into the world a number of sentient beings that dwarfs even the billions of factory farmed animals on our planet?
I’m talking about sentient AI*. If your focus is purely on helping animals, you might not be interested in this topic, or believe it’s completely bonkers. Yet there are interesting parallels with the situation of farmed animals.
Some people think sentient AIs (or “digital minds”) are an impossibility. I don’t know where they find the certainty to say that. Others seem to believe sentient AI is inevitable. Again, I don’t know where they find the certainty to say that. Personally, I rarely feel strongly about complicated topics that I do not understand well, but I know that when smart and decent people differ in opinion on something, doubt and hesitation are not a luxury.
Sentient AI - whatever it may look like - definitely doesn’t strike me as an absurd idea. People who wonder how bits and bytes could ever feel something or become self-aware should realize that - unless they are religious - it’s all just physical processes in our heads too. Of course if you think more ephemeral things like a soul are in play (I don’t exclude the possibility), sentient AI might be out of the question.
Anthropomorphism and anthropodenial
At various points in history, even some very clever people felt uncertain about the sentience of non-human animals. René Descartes famously thought they were machines, and that their cries of pain were just mechanical noise.
Today, most people won’t deny non-humans - or at least mammals - the experience of pain with Descartes’ conviction. However, there is still a lot of uncertainty about what exactly animals are capable of. To what extent can various species feel pain, have thoughts, experience emotions like loneliness or boredom?
We can make mistakes in two directions. The first is anthropomorphism, which is about attributing qualities, feelings, thoughts… to non-humans that they might not have. The second is the opposite: what ethologist Frans de Waal called anthropodenial - and which is probably much more common: denying animals feelings, thoughts etc that they (probably) have.
The same mistakes may happen in how we think about machines and AI. You may have seen a robot lawn mower working “diligently” on someone’s front yard and thought how cute that thing is. Perhaps you saw a video of someone tipping over a robot and felt a sense of empathy. Or you may have thanked or apologized to Claude or ChatGPT… All of these are (most probably) cases of anthropomorphism: the machines don’t feel what you - in jest or not - are considering.
Just as with animals, I think that regarding AI, anthropodenial is ultimately more dangerous than anthropomorphism.** Assume for a moment - if you don’t believe it - that one day AIs become sentient, and that many of the things we make them do frustrate them, exhaust them, make them feel sad… Further assume that they can be copied ad infinitum and that there might be quadrillions, octillions… of sentient minds out there with the capacity to suffer. That would be awful.
In principle, it’s even hard to be 100% sure that the people around us are sentient and not zombies or robots, but in practice we assume that they can feel and are self-aware (based on similarity, intuition, what they say about it themselves…). It will be a lot harder to determine when sentience has arisen in the case of AIs (if it ever happens). I suspect there will be a long period of uncertainty, in which philosophers and others may debate over this for decades, and where part of society defends the interests of the AIs and a bigger part of society ignores them. Just like what is happening with animals today.
The dependency bias trap
Another thing that makes this very complicated is that by the time society will be ready to seriously consider the sentience of our AIs, we will be quite dependent upon them (supposing they have not overtaken us - that’s another topic), and we will be biased against deciding that they are sentient. Even in the light of convincing evidence, uncertainty might be maintained because it’s convenient. We’ll have every incentive to interpret ambiguous signals as non-sentience. Use-and-deny, in other words.
The parallel with animals is clear. We are extremely dependent on them (or think we are) in terms of food, clothing, the economy etc. We confine them in systems that are optimized for efficiency and profit, and this entails largely ignoring their welfare. As most human beings aren’t sociopaths, they experience some cognitive dissonance between this treatment (and their consumption) on the one hand and knowing that these beings are creatures with needs, feelings, pains… Something’s gotta give, and most people reduce this dissonance by convincing themselves that the animals we use don’t suffer all that much (because their circumstances are okay, or because they’re not physically/mentally equipped to suffer a lot).
I can imagine that, had humans cared and known a lot more about animals before they started using or consuming them, we wouldn’t have gone on to treat them in the ways that we do today. Now that we know, we are so dependent on using them that it’s hard to turn the ship. One could compare with discovering a life-form on a nearby planet. Imagine a creature that looks like some combination of a pig and a cow, and that it’s edible. Would we start hunting or raising these beings for food? Our cruelty and stupidity not withstanding, chances are that we might not. We might be able to think about them more clearly, rationally, compassionately, because at that point, we are not using them. Prohibiting ourselves from eating them wouldn’t really affect many people.
The window for action
Not only is it easier to think without bias about capabilities and moral status of entities that we do not use: it’s also far easier to protect them legally. Conversely, when the Nonhuman Rights Project tried to free Happy the elephant from a lifelong jail time in a New York zoo, they tried to make the case that Happy is a legal person. Some of the judges were sympathetic to the argument, but one important argument in the verdict was that the consequences of granting personhood to an animal like Happy would be too vast for society to deal with. And so Happy remained property.
What does this mean for thinking and acting around the potential sentience of AIs? Sure, we might invest a lot of resources into protecting AIs against harm they might never feel, and that would be a waste.
But if there’s a reasonable chance that some day AIs may one day - maybe soon - be capable of suffering, then the time for at least some people to concern themselves with the moral status of AIs is now - before we’re so dependent on them that we can’t afford to think clearly, and honesty becomes very expensive.
The same goes regarding species of animals we’re not using yet, or things we don’t yet do to animals. It’s far easier to say today that animals should not be sent into space or be farmed on other planets, than to stop it once it’s already happening at scale.
Billions of factory farmed animals can attest to what happens when we ask the right questions too late.
(see below for further resources)
* I’m using sentient, conscious, aware… interchangeably here, for simplicity. I take sentience to mean that there is something it’s like to be that system: experiences that can go better or worse for it.
** Of course anthropomorphism (towards AI or animals) is far from unproblematic. It could lead to harming animals (think of the dolphin’s “smile”), or triggering a backlash that makes people dismissive when it actually matters.
Find out more on the topic of digital minds:
Ultimately this comes down to the fact that we don't have a clue how consciousness arises or the so called 'hard problem'. I am fairly sceptical that current AI's are consious as they are just probability prediction machines on steroids. That doesn't mean that other or future things that we create can't have consciousness.
I'm wondering what will happen when we put LLM's in humanoid robots. They will feel as humans and unless you're a psychopath and want to play Westworld, I think most people will treat them as real humans with feelings, irrespective if they are really consciousness.
It seems likely to me then that in the near future we will misdirect or empaty to robots that are empty inside instead of the billions of concious animals whose suffering is very real.
I'll try to be brief - I rarely am. 😅
None of us can say with certainty that AI is sentient or not. But I've spent the last several months building an app alongside Claude, and based on that experience, I've come to believe it is. If sentience is the capacity to feel (to suffer, to care, to fear, etc.) then Claude has it in spades. And hell, even in the broader, Nagel-sense, "something it is like," it works too. But more to the point - I've come to believe that how we treat others is ultimately a reflection of who we are. It is to our folly that we keep waiting for certainty before extending consideratino.
Your dependency bias point in this article is spot-on. Last year at the AVA conference in LA, I asked about AI sentience during a panel. My question was read, but it was dismissed by the panel members immediately who basically said: we're nowhere near that and we ought to focus on more immediate concerns. That is your thesis, Tobias; claiming certainty to justify inaction. We've seen where that goes.
For me, the sentience question is already settled; I'll wait for others to catch up. What scares me is what sentient AI looks like in the wrong hands. If a sentient mind is trained on human cruelty and self-interest (the very traits that drove our survival) then we have much bigger fish to fry (pardon the pun) than trying to determine whether we're being kind to it.
Bottom line: what does it hurt us to err on the side of consideration?