Humans Are Like Sharks, Artificial Intelligences Are Like Dolphins
In the search for intelligence, all that twitters is not gold.
It only takes a moment for a dream to become a nightmare. Blissful solitude became terrifying isolation the moment she saw the grey fin protruding from the waves. Frantically, she began to swim but there was no escaping. While her body was meant for land, millions of generations of evolution had shaped the pursuing predator into the ideal swimmer. She looked back just in time to see the silhouette of its long, streamlined body closing in. Her mouth opened to scream. But what came out was a laugh. For she felt not the slicing of teeth, but a rush of air as the animal leapt over her. Only now could she see that it was not a shark. It was a dolphin! It only takes a moment for a nightmare to become a dream.
Such is the resemblance of dolphins and sharks, that cases of mistaken identity are commonplace. They both have the same grey colouration. They have the same distinctive streamlined shape — long, narrow and smooth like a torpedo. Alongside their infamously deceptive dorsal fins, they also share a pair of pectoral fins at their fronts and a similar tail shape at their backs. At a glance, it would be natural to imagine that they are closely related animals.
But looks can be deceiving! Looking beneath the surface, we find that dolphins share more of their deepest traits with us than sharks. Like us, they are mammals that birth live young and nurse them with milk, while sharks lay eggs. They breathe air through lungs and would drown if unable to reach the surface, while sharks breathe underwater through gills. They have calcified skeletons like ours, while sharks have only cartilage. The resemblance of dolphins to sharks is only skin deep, while their likeness to us goes down to the bone. A dolphin can imitate a shark, but it functions like a human.
Today, imitation is a live issue. For the first time, we are encountering artificial intelligences (such as ChatGPT) with the potential to win Alan Turing’s famous imitation game by convincingly mimicking humans in conversation.1 This is inspiring far-reaching debate about the nature of their cognition, with some even concluding that they must be sentient. Discussions of these questions can quickly become speculative and unmoored from past knowledge which may seem irrelevant to understanding these futuristic systems. After all, what could the past teach us about something so unprecedented? As it turns out, the answer is quite a lot! We just have to know where to look.
A Tale of Three Creatures
Fifty million years ago, the ancestors of dolphins, sharks and humans roamed the Earth. There was no question then about which pair was most similar. While the ancestors of sharks were much like the sharks of today, the ancestors of dolphins and humans were both small, furry animals the size of cats. There were some hints of their differing progenies. While the ancestral dolphin had feet, its bone structure suggests it spent time in the water, perhaps living like another of its descendants — the hippopotamus. Meanwhile, the ancestral human was more likely to be found climbing a tree, much like its primate descendants. Yet no one could have doubted that they shared a kinship, just as surely as no one would have mistaken the dolphin’s ancestor for the shark’s.
How then did the descendants of a small furry animal come to look so much like sharks? The answer is convergent evolution. Over time, the dolphin’s ancestors came to spend ever more time in the water. Any advantage that equipped an individual of the species better in this new environment, such as being a slightly better swimmer or being slightly harder to spot in the water, was rewarded by that individual being more likely to flourish and produce numerous offspring with the same advantage. Over millions of generations, this gradually led the population to consist of individuals who were ever more adapted to their aquatic environment. The accumulated effect of these adaptations was the evolution of creatures who resembled sharks far more than their own ancestors. These are the animals we call dolphins.
Why did this process of evolution lead to creatures that look so much like sharks? Because sharks’ own traits are themselves the result of a much earlier process of adaptive evolution to the same aquatic environment. Dolphins and sharks both evolved their grey colouration to help them blend into the ocean since individuals of this colour were more likely to successfully catch prey and enjoy long, prolific lives. Dolphins and sharks are similar in shape and have fins and a tail instead of feet because that structure is best suited to fast and efficient swimming, so individuals with this shape were more likely to thrive and reproduce than their less streamlined cousins. The resemblance between sharks and dolphins is not a coincidence, but neither is it the result of a deep familial relationship. The common characteristics arose independently because they are the best response to the common selective pressures placed on both types of animals as a result of their shared way of life as marine predators.
By contrast, dolphins and humans appear superficially so different because of divergent evolution. If we go back far enough in time, we will find a single animal that was an ancestor of both us and dolphins.2 But over tens of millions of years, two of this creature’s lines of descent came to live very different lives. One line remained on land while the other took to the seas. The adaptations that suited these different environments are very different and so the individuals of each line came to look ever more different from each other. The eventual result was two types of animals that bear little superficial resemblance — humans and dolphins. Yet, the fundamental traits of the shared primogenitor are retained in both, explaining why deep commonalities remain.
The story of humans, dolphins and sharks teaches us a general lesson about evolution. Convergent evolution occurs when animals from different origins are exposed to the same selective pressures. It leads to creatures that are superficially similar but retain deeper differences. By contrast, divergent evolution happens when animals from a common origin are exposed to different environments, leading to creatures that are superficially different but retain deceptively deep similarities. Equipped with this lesson, let us now return to questions of artificial intelligence.
Machine Evolution
Of course, fifty million years ago, artificial intelligences had no ancestors. They are created out of nothing by human engineers. Yet, they too develop their traits through a process of evolution.
An artificial intelligence is initiated as nothing more than a basic structural template that is populated by billions of numbers. These numbers determine its behaviour; they provide instructions for how the system is to respond to different inputs. Each number is analogous to one letter in the DNA of an organism; collectively, they act like the genome.3 Initially, the numbers are set to default values and the output of the system is gibberish. However, the system is then trained by making small changes to the numbers, like mutations in the DNA of organisms. When changes improve the performance of the system, they are accepted and the updated system replaces the old one; otherwise, they are rejected. This is analogous to how mutations that improve the adaptedness of an organism to its environment come to displace older genes through natural selection, while detrimental mutations quickly die out. Over an immense number of iterations of this process, the improvements accumulate until the system is effective at performing the task it was trained for, just as the result of natural selection over millions of years is the evolution of well-adapted organisms.
The capabilities of the system evolved in this way depend on the definition of improved performance used in training. This dictates the selective pressures which shape the evolutionary process, just as the environment of a species determines the pressures which direct its evolution. For large language models, such as ChatGPT, success is defined by how authentically and impressively the text output of the system imitates that of a human. This results in a process of convergent evolution. The system becomes ever better at mimicking the aspects of humans that are essential to conversation, just as the evolution of the dolphin led it to increasingly share the traits of a shark that are vital to being a successful aquatic predator. For billions of years, convergent evolution has successfully constructed imitants in nature. With hindsight, it should be no surprise that the same phenomenon should prove the key to winning Turing’s imitation game. We have simply had to wait for computational power, large-scale data collection and algorithmic ingenuity to develop to the point where we can reproduce this ancient process in a computer.
The Search for Intelligence
If the only aquatic animals we knew of were fish, it would be natural to assume that no animal could live permanently in the ocean without the ability to breathe underwater. It would feel obtuse to imagine a sea creature that could drown. If the descendants of a lunged animal were to come to live in the ocean, it would seem inevitable that they would eventually evolve gills. Yet, in reality, land animals have evolved into sea creatures without this ever happening. Indeed, this has happened not just once but multiple independent times — not only for dolphins and whales, but also for seals, dugongs and manatees, turtles, and sea snakes, among others. Convergent evolution is a master at defying intuition.
Similarly, since the only beings we have ever known that are capable of coherently conversing are humans, it is natural to assume that no being could hold a conversation without the human trait of conscious thought. Yet, we can now see that this is presumptuous. Just as convergent evolution can produce animals that can imitate sharks without needing gills, so it may produce entities that can imitate humans without needing sentience. Like a terrified swimmer sighting a fin, we can be too quick to jump to conclusions based on appearances alone.
Conversely, as the deceptively deep similarities between humans and dolphins show, divergent evolution can mask profound commonalities under superficial differences. If we are searching for the beings most similar in thought to us, perhaps we should look not to our best mimics but to our closest relatives — chimpanzees and bonobos. That is not to say that bonobos belong in our offices writing our emails, any more than pigs (as relatives of dolphins) belong in aquariums. Imitation and resemblance are important when assigning roles and assessing utility. But it does mean that we should be cautious about projecting our rich inner lives onto mimicking machines more readily than are prepared to with our closest animal kin. Let us always remember: in the search for intelligence, all that twitters is not gold.
In fact, this same animal was also the ancestor of a wide range of other mammals as diverse as bats, horses, moles, tigers and capybaras.
More context on how DNA and genes provide the instructions for organisms can be found in the following post:
Thanks Paul, excellent outline of convergent and divergent evolution. I does seem that ChatGPT is just a clever “imitaton”, but because:
(a) the “hard problem” of what makes consciousness (or entities are conscious) is still “hard” (and I believe unsolved), then
(b) one of several candidates could be the case, and thus
(c) because one of those candidates is that any complex system with the quantity of nodes, connections and feedback circuits similar to a human (or even planetarium worm brain) might be sufficient for the “emergence” of consciousness, then ..
(d) it might be that a highly complex AI “dark cloud” will mysteriously generate its own silver lining of emerging shiny sentience (i.e. we can’t say).
So, while “all that twitters may not be gold” it could equally be the case that every complex “dark cloud might have a sentient lining”.
Of course if such an AI system was to be embodied and able to sense and move around in our environment and then set free to evolve, it may through convergent evolution be even more likely to develop some form of consciousness, though even then (because of the “other minds” problem) we would still never really know “what it is like to be a robot”.