Earlier this year, a bonobo named Kanzi (now deceased) was shown to interact with imaginary grapes and juice, suggesting a capacity for imagination. A review from last year suggested that insects possess subjective experience, citing bumblebees who roll balls without reward and honeybees that opt out of difficult tasks. Other studies suggest some birds and flies have lasting fear memories or that some corvids show causal reasoning. But what does all of this tell us about whether animals possess consciousness?
Comparative cognition researcher Edward Wasserman would say not much. Wasserman, who is a professor at University of Iowa, has been studying pigeon cognition for more than 40 years, particularly the birds’ ability to understand complex concepts. Pigeons are quite intelligent, his findings show, but he says he is highly skeptical of the idea that intelligent behavior can be equated with internal subjective experience. He’s also increasingly interested in what intelligent animal minds can tell us about the nature of intelligent machines.
Read more: “Insects and Other Animals May Have Consciousness”
Wasserman was recently included on a list of 100 researchers working on provocative and difficult to answer research questions published by The New York Times. The question attributed to Wasserman was, “Will we ever understand consciousness in a way that could apply to nonliving entities?”
I spoke with Wasserman recently about the law of effect and the edge of chaos, how animal research informed machine learning, what separates animal consciousness from human consciousness, and what that can tell us about whether Claude has a mind.
You’ve spent many decades studying pigeon behavior. Why do pigeons play such an important role in animal cognition research?
Pigeons came into the process with B.F. Skinner. In the middle 1940s, Skinner got involved in a project that was ultimately supported for a while by the military, to have pigeons guide missiles via screens. He found out in a hurry that you had to engage in active learning methods to get the pigeon to do this.
He and his students developed this technique to train pigeons by using reward. That’s how “shaping” came about. The work with pigeons was foundational to many things that Skinner later went on to study. He and his students later trained many birds to do many things. But the pigeon lives a long time. And the bird is sufficiently smart that they can learn interesting things, but not so smart that they can confound the experimenter.
Some animals are just too clever. They figure out ways to learn tasks that perhaps we wouldn’t have thought of. I’m a kind of a second or third generation operant conditioner and the pigeon is a very convenient animal.
What kinds of interesting things have you taught pigeons to do?
Our pigeons learn tasks that are very conceptual. I can’t tell you that it’s turning a somersault. These are tasks that we’ve developed that would thwart most people. And the reason for it is that the pigeon probably doesn’t have the same advanced cognitive capacities of people, but they rely on a mechanism that’s tried and true in evolution.
And that’s if you make a response and it’s followed by reward, you make it more often. If you make a response and it isn’t followed by reward, you stop. And the fact that you repeat responses that have been rewarded in the past, it’s called the ‘law of effect.’ And this law was discovered by Thorndike as a graduate student in the 1897 or thereabouts. Before Pavlov, he became the scientific discoverer of the mechanism of learning.
Does that law of effect extend beyond animal learning?
Yes, that mechanism—trial and error—is what powers computers. Computers failed for many, many, many years. Herbert Dreyfuss wrote a couple of books: What Computers Can’t Do and another one about What Computers Still Can’t Do. And the reason why computers couldn’t do these things is because people were programming them with rules.
They thought that a human is a rule-based machine, and that we respond based on these elaborate cognitive constructs and logical relations. And the ultimate answer was that we respond to the same basic primitive law of effect.
When they trained computers the way they trained pigeons, the computers got smart. And so what you’re looking at when you converse with an LLM is modeled on the basis of pigeon learning. The people who got the Turing award, Andrew Barto and Richard Sutton, in 2024: They got it because they based it on animal learning. It goes back to Skinner.
How do you train a pigeon to recognize a bottle instead of a book? Well, you’d show them many, many, many, many examples of books and many, many examples of bottles, and you reward them when they make the correct response, and you don’t reward them when they make the incorrect response. How is a computer going to make predictions about what the next word in a sentence is going to be? Based on past experience. For computers, the reward is called feedback. You simply get a signal. One signal says correct. The other says incorrect. And they’re trained with millions of examples.
But you published research earlier this year that complicated the law of effect.
Yes, the variability paper, that message is really an important one. The law of effect says that if you make a response and it’s rewarded, repeat it. If it’s not rewarded, don’t repeat it. That would lead you to always lock in on a single or really small number of possible responses.
But we built a task where the pigeon had five buttons they can peck in any order. And what we found was that the pigeons never settled on a single or small number of sequences. They kept rotating through many different sequences over eight months. I never thought this experiment would last that long. When I told people in the lab what I was going to do, they said, “What do you expect? Maybe a week or two?” I said, “Ah, let me not go out on a limb. Let me, let’s just see what the birds do.”
In that experiment, there were 120 possible sequences. Ultimately, every bird made all 120 of them over that period of eight months, but they never really settled in on a set small number. And you say, “Are the birds just perverse?” The answer is maybe, maybe they are. Maybe they’re just playing head games with us.
But another possibility is that, although I told you about one algorithm, the law of effect, there’s another algorithm: It’s mutation. For evolution to work, you have to have survival of the fittest, but you also have to have variation, because if you don’t have variation and the environment changes, you’re dead. So I believe we’ve stumbled into another important algorithm.
Does this law of variation and mutation hold for computer learning, too?
Yes. Computer scientists have now discovered that if they don’t build variation into LLM design, learning doesn’t work as well. And it doesn’t work as well for the same reason, because if you get to a novel situation and nothing is working, you don’t want to go extinct. You need to have enough variation in your behavior that you don’t lock in.
You usually think noise is a bad thing. Most AI people said, “Let’s get rid of all the noise.” And that’s the wrong approach. They have now built in variation into their training. So there are these two algorithms. One of them generates stability. One of them generates variability. And in biology, there’s a concept called the edge of chaos. And the edge of chaos is where you want to be.
This edge of chaos is a real thing. If you’re too variable, you’ll never settle in and you’ll fail to adapt. If you’re too fixed, you’ll extinguish because when things change, you’ll be unable to be flexible. Cassius Clay said, “You move like a butterfly, and you sting like a bee.”
You sent me a recent essay of yours on animal consciousness, in which you write, “I see no reason why we should still be beguiled by the siren song of animal consciousness,” which is also the title of the piece, “The Siren Song of Animal Consciousness.” What makes it a siren song and why is it so beguiling?
I’ve actually lost philosopher buddies over this, because it just isn’t a useful question. It must be beguiling to some, but it isn’t beguiling to me. I don’t know what I would have to do in order to answer the question as affirmative or negative. What would you have to see or record that would convince you? After all, look at people with locked-in syndrome. How many thousands of people who are locked, seeing, hearing, understanding, but unable to react?
Haven’t they devised tests that can determine whether someone with locked-in syndrome is conscious?
Have you seen the “Diving Bell and the Butterfly”? It’s a film but it’s based on real life. The fellow had a stroke and he could only move one eyelid. He could blink with one eye. He had an attendant, a nurse who held up a sheet in French, because this was in France, where the letters of the alphabet were ordered in terms of their frequency. And she asked him to respond with one blink if the letter she was pointing to was correct, and two blinks, if that was incorrect. And she spelled words. He blinked, and he wrote a book. Without language, I’m not so sure how she could’ve taught this, because he already had language. But this was the way that they established communication. The book was published right after he died. It’s a very poignant and chilling story.
So there you have an example of consciousness. Now, mind you, if it happened to be a young baby, it would be helpless. There would be no way you could tell. Now there are techniques to monitor brain activity, so you don’t have to use an overt response. There are covert responses.
But think about the mechanism. How did this work? What’s happening when he blinks once, yes, twice, no? Does that not sound like reward and non-reward feedback for these responses? When an animal learns a task, it isn’t trying to please us. It’s receiving reward for certain responses and not receiving reward for other responses. The animal’s behavior changes. Those changes in behavior, we can read. We’re smart. No animals are teaching us how we should respond. We’re teaching them how they should respond. And then we make inferences. But our inferences are always these mentalistic ones.
You offer the story of Pavlov as an example of how tangled up efforts can become when you use these mentalistic types of reasoning. Do you see that creeping back into animal research specifically today?
Oh, yes. The article that I wrote for that journal, if you go back to the journal, it’s a collection of papers in response to a declaration that animals have consciousness. They got a whole bunch of philosophers and biology people, and ethics people, and psychology people, and computer people, and they tried to make a declaration: “We’re seeing hints that animals might be conscious. Those signs are their behaviors.” Is behavior tantamount to consciousness? I don’t know. The father of modern animal consciousness is Donald Griffin. He and I never saw eye to eye. He wrote, “Well, Wasserman’s done these very nice experiments showing that animals can learn concepts. They can categorize the world into cats and cars and chairs and flowers. But why doesn’t he go all the way? Why doesn’t he say, ‘Well, they’re conscious of what they’re doing.’?” And I just don’t see the point of going there. It doesn’t add anything. It brings us back to Pavlov, trying to understand what the dog might have been thinking or feeling. I think part of the slippage here is, well, if you say they’re conscious, then we have to give them rights.
Right, this is one important reason they signed that declaration. They want to protect their welfare. Do we need to believe that animals have consciousness to protect their welfare?
Absolutely not. You can respect animals, you can do everything in your power to preserve their habitat, to take care of them, to minimize any discomfort. I mean, veterinarians know a lot about animals. If you ask a vet, “Is this animal in pain? Is this animal suffering?” They have some really informed guesses, much more informed than the philosophers who are opining about consciousness in whales. Veterinarians know more about your dogs and cats and parrots than the philosopher is going to know about a whale. And I don’t know what you accomplish by doing that, other than in some ways, minimizing what it is to be human. If you’re humanizing animals, the opposite is you are animalizing humans. But the last time I checked, we are animals.
It seems like the number of cognitive capacities understood to be unique to humans keeps shrinking. What features of human cognition still separate us from other animals?
The safest answer is metacognition: Thinking about thinking. Descartes famously said, “I think therefore I am.” The fact that I’m thinking is proof that I exist. When somebody else is looking at you, how do you know that you’re not looking at a cyborg?
And that’s where the Turing test came about. Turing famously proposed that we could engage in dialogue. And if I couldn’t exceed 60 percent in my accuracy between a computer and a person in my conversation, then I’d probably have to confess that the computer was conscious. That’s a test, but it’s probably not a very good test because look at all the animals that are around us, including infants that don’t speak. You wouldn’t want to conclude that your baby has no consciousness because it doesn’t speak.
How do we know animals are not capable of metacognition?
What would a case of metacognition be? Carolyn Ristau has studied death feigning in plovers. So the plover is a bird that if a fox is approaching the nest, she’ll leave the nest and do this thing with the wing, and flop over and then lie still. This looks like a genetic program because not all birds do this. Is the bird aware that when it’s moving away from the nest, it’s preserving the babies? Does it know that the fox thinks that it’s dying or dead and this is why it’s doing it? Do you think that envisioning this elaborate theater of the consciousness of the bird is a useful way to approach our understanding of the world? Is this scientifically satisfying? I just can’t go there. I just can’t take that step.
Then we’ve got death. Do animals understand the concept of death? We hear about the dolphin or the whale pushing their dead calf through the ocean for weeks. What is that? Maybe it is grief. Maybe there is something like grief. I’m not in a position to say yes or no. I think you can be skeptical. Aren’t we supposed to be trained to be skeptical?
As you note in that same Siren Song essay, Spencer and Darwin proposed minds evolve just as bodies did, which would suggest some sort of continuity between human and animal minds. But then why wouldn’t we infer that consciousness exists in other species?
Why don’t we just infer that consciousness exists in computers? We want to be able to treat all entities equivalently from a scientific vantage point. If we see evidence that suffices in one case, then it should suffice in the other.
And you know the interesting difference? We evolved the computer. Nature evolved us. There’s a French materialist philosopher named Julian La Mettrie. He contended that nature made the human machine. He called it L’homme Machine. He said, Nature made eyes which can see without itself being able to see. Nature made a machine which can think without itself being able to think. Now, here’s the next step. That human machine that can think has made an artificial machine that can think. But did we make the artificial thinking machine because we can think or because we discovered the right rules in order to do it? This is a meta-analysis.
Have you ever interacted with Claude or with an AI extensively? I have absolutely every feeling when I do that I’m talking to a person. I have absolutely no doubt that its responses are in every way, shape or form just what I would expect from talking to somebody who is really, really well versed in the material that I’m asking about. I ask questions, and I catch it when it’s wrong, and it corrects itself. It catches mistakes I’ve made or things that I haven’t thought of. This is just like talking to a person. And you know why? Because it’s had conversations with millions of other persons stored in it. So is it conscious? Is it conscious of what it’s doing?
The pigeon may not have the same circuitry that we have. Its brain is not the equivalent of our brain. Our brain may have evolved in ways that transcend what the pigeon brain can do or the primate brain can do. That’s entirely possible. That’s an absolutely open question. But there’s every reason to believe that machines that are built on that primitive, ancient process can get smarter than we are.
Recently The New York Times included you on a list of 100 researchers exploring unanswered research questions. The question attributed to you was, “Will we ever understand consciousness in a way that could apply to nonliving entities?“ Did they come up with the research question or did you?
They asked me to. I thought, what could I ask that would be provocative? It wasn’t just, “Can computers think?” It’s “What would we have to do to understand the question of consciousness if we were applying it not only to people and animals, but computers? What would the common thread be? What would we really be seeking?”
But it almost seems like your answer to that question is: who cares?
Well, I can’t imagine what the answer would be.
Are there experiments that could be done that would get us closer to an answer that you can imagine?
You never say never. Don Griffin said, “Maybe these behaviorists are just hide bound. They’re not daring. They don’t go out on a limb. They don’t want to take that next step.” One thing I will say, if you ask somebody, if I’ve taken daring steps in experiments, they would have to say, “Yes.”
The experiments that I’ve done ask tough questions. They find surprising results. Things that you might never have expected that animals can do. But I’m perfectly content, as you suggested, to push that envelope. To say, “Well, what can a pigeon do? It pecks at a key. So what? What’s the big deal?” Well, what we’ve done is to show how elaborate the machinery of their brains is and to design computer models that can actually emulate what the pigeons are doing. And it turns out that pretty simple models can actually do those things.
All the while, I feel myself pushing to try to look for more advanced forms of adaptation on the part of animals. I’m never going to give up rooting that success in primitive processes. Maybe these primitive processes aren’t so bad after all. Remember, the only way to learn about the world is through trial and error.
Enjoying Nautilus? Subscribe to our free newsletter.
Lead Image: Mabeline72 / Shutterstock
