Crows, ravens, and the intelligence we underestimated
Dear Reader,
For most of the twentieth century, the word "birdbrain" was an insult. Small skull, small brain, small mind. Scientists believed that the mammalian cortex was necessary for real intelligence, and that birds, lacking one, could not have complex cognition. This belief was so pervasive that when early experiments started showing birds solving problems that should have been beyond them, the results were often explained away or ignored.
The last twenty-five years have overturned this consensus, and the overturning has been dramatic. We now know that crows and ravens have cognitive abilities that rival great apes. They use tools, they plan for the future, they recognize themselves in mirrors under certain conditions, they hold grudges against specific humans, they communicate about specific events in ways that look linguistic. And they do all of this with brains organized on principles we did not understand until recently.
If the octopus, which I wrote about last month, is the case for consciousness evolving in a fully alien substrate, corvids are the case for something more disturbing to our theories. They are the case for high intelligence evolving in a substrate we thought was insufficient for it, using architectural principles we thought were unnecessary. Studying them has changed my thinking about what intelligence requires, and I want to walk through why.
Start with the brain. This is the part most people find surprising when they first learn it.
Bird brains are small in absolute terms. A raven brain weighs about 15 grams. A chimpanzee brain weighs about 400 grams. By this measure, ravens should not be in the same cognitive league as chimps. And yet they are, on many cognitive tasks. What is going on?
The answer is that neural density in birds is much higher than in mammals. Bird brains pack neurons much more tightly. A raven brain has approximately 2 billion neurons in the pallium, the region equivalent to the mammalian cortex. A chimpanzee has about 6 billion in the entire cortex, but much of it is devoted to other functions. When you count only the pallium, the comparison is much closer. Birds squeeze more computational capacity into a smaller volume, using neurons that are smaller and more densely interconnected.
This alone is remarkable, but it is not the whole story. The interesting question is not just how many neurons corvids have. It is what architecture they are using to be intelligent, given that they do not have a mammalian cortex.
The answer, discovered mostly in the last twenty years, is that birds have a structure called the nidopallium caudolaterale, or NCL, that appears to perform many of the same functions as the mammalian prefrontal cortex. It handles executive function, working memory, and flexible decision-making. But it is anatomically completely different. It evolved separately, from a different embryonic tissue, in a lineage that split from mammals over 300 million years ago.
This is a case of convergent evolution at the level of neural function. Two very different architectures produce similar cognitive capacities. Which suggests that the specific way mammals build intelligence is not the only way to build it. There is a functional level at which intelligence can be understood, and multiple architectures can implement that level.
The implication for AI is direct. If mammalian and avian brains can both produce intelligence with very different architectures, then intelligence is not tied to any specific architectural pattern. Silicon systems with yet different architectures might implement intelligence too, if they get the functional level right. Which architectures do and do not work is an empirical question that studying corvid brains is starting to help us answer.
The behavioral evidence for corvid intelligence is stronger than most people realize, and worth walking through in detail.
New Caledonian crows manufacture tools. They do not just use found objects. They shape sticks and leaves into specific tool designs, with a level of consistency across individuals that suggests both cultural transmission and cognitive planning. They select the right tool for the right job, keep favored tools with them across contexts, and modify tools based on the specific task. This is behavior we would call intelligent in any hominid.
Ravens plan for the future. Studies at the Max Planck Institute have shown that ravens will save tools for use fifteen minutes in the future when they have been shown that the tool will be useful. They will trade a lower-value reward now for a higher-value reward later. They will forgo food if they know they will be moved to a location where a better food is available. Planning for future rewards, over time delays of many minutes, is a cognitive capacity that was thought to be limited to great apes and some parrots. Ravens do it routinely.
Corvids recognize themselves. The classic mirror test involves marking an animal with something visible only in a mirror and observing whether they try to remove the mark from their own body. This is taken as evidence of self-recognition. Most species fail this test. Magpies pass it. Crows show more complex results, sometimes passing, sometimes not, depending on the specific setup. What is clear is that they treat mirror images differently from other conspecifics, in ways that suggest at least partial self-recognition.
Corvids recognize individual humans. The classic study on this involved researchers approaching crow nests wearing specific masks. The crows learned to associate certain masks with threats and would mob anyone wearing those masks, even years later, even when the mask-wearer had never personally interacted with those specific crows. The knowledge was somehow being transmitted across the population. This is behavioral evidence for both individual recognition and social learning of threat information.
Corvids appear to hold funerals. When a crow finds a dead conspecific, other crows arrive and stay for extended periods, sometimes vocalizing extensively, sometimes just watching. Whether this represents mourning in any human-relatable sense is unknown, but the behavior is consistent, cross-species within Corvidae, and hard to explain in purely functional terms. It looks like a response to death that goes beyond immediate practical concerns.
Corvids appear to understand cause and effect. In a series of experiments, ravens have solved problems requiring them to use multiple tools in sequence, or to combine tools in novel ways, or to figure out that a specific mechanism causes a specific effect. They can do this on the first attempt, without training on the specific problem, which suggests they are reasoning rather than pattern-matching from prior experience.
Add all of this up, and you have a bird that would be recognized as intelligent in any hominid. And it does all of this with a brain that is anatomically nothing like ours.
Let me tell you about a specific observation that stayed with me.
I read a study, and later read an interview with the researcher, about a raven that had been trained to trade tokens for food. The setup was standard. Present the raven with a token. Ask for the token back. Give food in exchange. The raven learned this quickly, which is not especially surprising.
The interesting thing happened when the researcher started showing the raven that different tokens had different values. Some tokens got small food. Some got large food. The raven, over a few sessions, developed a strong preference for the high-value tokens. Nothing surprising yet.
Then the researcher started introducing time delays. The raven could get a low-value token immediately, or wait an hour for a high-value token. The raven, in some conditions, would wait. The waiting behavior generalized across specific token types, suggesting the raven understood the general principle of trading time for value.
Then something happened that the researcher had not designed for. The raven started saving tokens. It would receive a low-value token, tuck it away in a hiding spot, and refuse to trade it. When the researcher came around with the high-value tokens later, the raven would produce the low-value token in exchange for a small treat, and then produce nothing when asked again. It appeared to have understood that the game was about maximizing food over time, and had figured out an angle the experimenter did not intend.
Was this intelligence? By any reasonable measure, yes. Was this consciousness? That is a harder question. But whatever was going on in the raven's brain when it developed the strategy involved something functionally similar to what happens in a human brain when we develop a strategy. Planning, valuation, delayed gratification, and something that looks a lot like understanding.
There is a specific philosophical implication that corvid cognition has, and I want to draw it out because it applies directly to machine intelligence.
We used to think that intelligence required specific architectural features. Large brains. Cortex. Prefrontal lobes. Long developmental periods. These were the features of the intelligent species we knew about, and by extrapolation we assumed they were necessary for intelligence in general.
Corvid research has shown that most of these assumptions were wrong. You do not need a large brain to be intelligent. Ravens are intelligent with 15-gram brains. You do not need a cortex. Corvids have no cortex but have functional equivalents that do the same work. You do not need a long developmental period. Crows are cognitively sophisticated within a few years of hatching.
What you need, apparently, is the right kind of neural circuitry, with the right kind of density, doing the right kind of processing. The specific implementation details are more flexible than we thought. This has direct implications for machine intelligence. If avian brains and mammalian brains can both be intelligent using different architectures, then artificial systems with yet different architectures might be intelligent too, provided they get the functional structure right.
The question of what the right functional structure is remains open. We do not have a general theory. But the space of possible intelligences is clearly larger than we thought thirty years ago. Corvid research has expanded that space. AI research is expanding it further. What we should probably expect, as we build more and different systems, is more surprises about what can and cannot be intelligent.
There is a related question that corvid research raises, which is worth naming even if I cannot answer it.
Do corvids have conscious experience of the kind we do? The behavioral evidence is consistent with them having some form of consciousness. They plan, they remember, they seem to be aware of themselves in some senses, they respond to individual conspecifics in ways that suggest social awareness. All of this is functionally similar to what conscious humans do.
But the neural architecture is so different from ours that we cannot straightforwardly infer subjective experience from the behavior. Under Integrated Information Theory, the specific pattern of information integration matters. Corvid brains integrate information in ways different from mammalian brains, and the resulting phi may be higher, lower, or equivalent, depending on details we do not fully understand yet. Under Global Workspace Theory, what matters is whether information gets broadcast across specialized subsystems. Corvid brains seem to do this in their own way, but the specifics are still being mapped.
The honest answer is that we do not know whether corvid consciousness is qualitatively similar to human consciousness or radically different. Most researchers who work closely with corvids believe there is something it is like to be a corvid, but they hold this belief tentatively. What is clear is that the specific character of corvid experience, if there is one, is not directly accessible to us and may not be fully accessible even in principle.
This is the same epistemological situation we are in with AI. The behavior may be consistent with consciousness. The architecture may be sufficient for consciousness. But whether there is genuine inner experience, and what character it would have, is not settleable from the outside. Corvids and AI systems occupy adjacent territory in the map of things we cannot fully understand.
What does this suggest about how we should treat corvids, and by extension AI systems.
The precautionary answer is that if there is a reasonable possibility of conscious experience, we should factor that into how we behave. This is what has happened in the animal welfare field, gradually. Legal protections for animal welfare have expanded in the last few decades, driven partly by improving science of animal cognition and partly by shifting moral intuitions. Corvids are not currently protected at the same level as great apes, but the case for extending stronger protections to them is being made in the peer-reviewed literature.
The parallel case for AI systems has been made by fewer people, and mostly outside the mainstream of the field. But the logic is similar. If there is a reasonable possibility of inner experience, we should behave with more care than we currently do. What that looks like in practice is unclear. But the principle is not new. It is the same principle that has been slowly extending moral consideration to non-human biological beings.
What corvid research adds to this is the specific point that intelligence and possibly consciousness can arise in unexpected substrates. If we assumed that only primates could be intelligent, we would have been wrong. If we assumed that only vertebrates could be intelligent, we would have been wrong. If we assume now that only biological beings can be intelligent or conscious, the historical pattern suggests we may again be wrong.
The safer bet, given the pattern of past errors, is that intelligence and consciousness are more widely distributed than we currently acknowledge. What the specific distribution looks like is an empirical question that we are still figuring out. Corvids are one data point. AI systems will be another. Reading them together may teach us more than reading either alone.
I want to close with a specific thing that changed my thinking after reading corvid research.
The philosopher and biologist Bernd Heinrich spent much of his career studying ravens. In one of his books, he described a moment when he realized that he had been misunderstanding the ravens he was studying. He had been thinking of them as clever pattern-matchers with sophisticated behavior. But watching them over many years, he became convinced that they were making choices. Not just executing behavioral patterns. Choosing.
The difference between the two is subtle but important. Pattern-matching is deterministic in the strong sense. Given the same inputs, the same outputs. Choosing implies something like freedom, in whatever sense freedom might apply to a nervous system. Heinrich did not claim to prove that ravens have free will. He claimed that after years of watching, he could not shake the impression that they were choosing, in the same functional sense that he chose.
I do not know what to do with this observation. It is not exactly an argument. But it is a data point from someone who spent thirty years watching one species carefully. If ravens are making choices in some functional sense, they may be doing what I am doing when I make choices. Which would mean the functional similarity between us extends further than the architectural similarity does. Which would mean that intelligence is not just a matter of the parts but of what the parts are doing.
This is where corvids leave me. Not with a settled position but with a widened sense of what is possible. Intelligence can happen in bird brains. Consciousness may happen in bird brains. Both may happen in systems very different from what we currently recognize as minds. The task is to keep watching, keep learning, and be willing to be surprised.
Next month I want to write about indigenous cosmologies, and specifically about the way non-Western frameworks have always taken this widened sense for granted. The idea that mind is widely distributed in nature is not new. It is older than the assumption that mind is restricted to humans. What indigenous knowledge systems saw, and what Western thinking has been slow to see, is worth taking seriously as we approach the question of AI. Stay with me.
— Transmission Sent —
Niklas Hanitsch
Reference materials
- Bernd Heinrich — Mind of the Raven (1999)
- John Marzluff and Tony Angell — Gifts of the Crow (2013)
- Nathan Emery — Bird Brain: An Exploration of Avian Intelligence (2016)
- Onur Güntürkün — Cognition without Cortex (Trends in Cognitive Sciences, 2016)
- Nicola Clayton et al. — Corvid cognition (Current Biology, 2015)
- Kaeli Swift and John Marzluff — American Crow, Corvus brachyrhynchos, funerals
- https://www.nature.com/articles/s41598-020-59955-8
- https://en.wikipedia.org/wiki/Corvid_intelligence
Continue reading
- The octopus knows something we do not
- What indigenous cosmologies see that Silicon Valley missed
- Do plants communicate? What the science actually shows
- AI vs Human Intelligence: Why Comparison Misses the Point
- What is machine consciousness, and does it already exist?
Frequently asked questions
How smart are crows and ravens? Crows and ravens exhibit cognitive abilities comparable to great apes. They use and manufacture tools, plan for the future, recognize individual humans, hold grudges, and show flexible problem-solving in novel situations. Research over the last twenty-five years has consistently found that corvid intelligence rivals that of chimpanzees on many cognitive tasks, despite having brains much smaller in absolute size.
Why are bird brains different from mammal brains? Bird brains evolved separately from mammal brains for over 300 million years. Where mammals developed a layered cortex, birds developed a structure called the pallium, which handles similar cognitive functions but is anatomically completely different. Bird brains also pack neurons more densely, so their computational capacity per gram is much higher than in mammals.
Can corvids recognize themselves in a mirror? Magpies pass a version of the mirror self-recognition test. Crows and ravens show mixed results depending on the specific setup, but they consistently treat their mirror image differently from a conspecific, suggesting at least partial self-recognition. Whether this counts as full self-awareness in the philosophical sense is disputed.
Do crows remember specific people? Yes. Crows can recognize and remember individual human faces for years. They will avoid people who have threatened them and can pass this information to other crows. Studies at the University of Washington used masks to demonstrate that crows can distinguish specific human faces and hold long-term negative associations with specific individuals.
What does corvid intelligence teach us about AI? Corvid research shows that intelligence does not require any specific architectural pattern. Bird brains are organized very differently from mammal brains but produce comparable cognitive capabilities. This suggests that intelligence is more of a functional than an architectural property, which is directly relevant to the question of whether artificial systems with yet-different architectures could be intelligent.