In the middle of the twentieth century, the Dutch ethologist Nikolaas Tinbergen became interested in something that looked almost trivial: why newly hatched gulls peck at a red spot on their parents’ beaks. The behavior had an obvious purpose. Pecking helps prompt the adult to feed the chick. But Tinbergen kept asking a more basic question. What, exactly, was the chick responding to? He began stripping reality away. He and his colleagues showed chicks simplified models of a gull’s head and beak. They changed the colors, shapes and markings. Eventually they discovered something wonderfully strange. The chick did not need anything that looked much like its mother. A crude object carrying the right visual cue could trigger the response. Stranger still, an exaggerated version of the cue could sometimes produce an even stronger response than the natural one. Later accounts of the work describe chicks responding intensely to a thin artificial object with conspicuous red markings, even though it looked far less like a gull than an actual gull did.
Tinbergen and other ethologists found versions of this pattern elsewhere. Some nesting birds, when presented with an artificial egg much larger than their own, would try to incubate the oversized fake. The egg could be obviously wrong in several respects and still be compelling because one characteristic had been amplified. Ethologists came to call this a “supernormal stimulus”: an artificial or exaggerated version of a meaningful signal that can provoke a stronger response than the real thing.
There is something quietly profound in that story. The gull chick had not evolved to recognize “mother” in some complete sense. It had evolved a practical shortcut. For generation after generation, a certain combination of color, shape and movement was reliably connected to something that mattered enormously: a parent carrying food. There was no reason for the chick to ask for more evidence. The shortcut worked. The weakness appeared only when something entered the environment that could produce the signal without being the thing the signal had always represented.
Nature is full of these shortcuts. A smell can signal a mate. A cry can signal a hungry offspring. Bright color can signal poison. Movement in the grass can signal danger. These shortcuts are not signs of stupidity. They are efficient ways of surviving in a world where there is rarely time to establish the full nature of everything before responding to it. But every reliable shortcut contains an interesting vulnerability. If another organism can reproduce the signal while separating it from the underlying reality, the recipient can be made to respond to something that is not really there.
Cuckoos discovered this long before we did. A cuckoo does not need to convince a host bird through reasoning that its egg belongs in the nest. It only needs to exploit the cues the host bird uses to recognize what belongs there. Over time, this has produced an extraordinary evolutionary contest. Host birds become better at spotting impostor eggs. Cuckoos become better at imitating the colors and patterns the hosts use for recognition. In some species, even the begging behavior of the cuckoo chick is tuned to trigger the host’s feeding response. What makes this fascinating is that the host bird is not foolish. Its recognition system is highly adapted to the world it evolved in. The problem begins when another organism learns how to manufacture the evidence that system has learned to trust.
I keep coming back to Tinbergen’s gulls when I think about artificial intelligence.
For almost all of human history, certain signals implied that another mind was present. Language was one of the strongest. If something could hold a conversation with you, remember what you said earlier, respond appropriately to sadness, make a joke, disagree with you, explain an idea and adjust to your reaction, there was very little reason to wonder whether there was a mind on the other side. The signal and the underlying reality had always travelled together.
Our social instincts developed inside that world. Sustained attention came from people. Empathy came from people. Memory of a relationship came from somebody who had actually shared that relationship with you. Confidence usually came from a person who believed something. Advice came from someone who would, at least in principle, have to live with some understanding of its consequences.
Now we are learning how to produce many of those signals independently of the things they once represented.
That feels more important to me, at least in the near term, than the argument over whether AI is conscious.
I do not know whether machines will someday have subjective experience. I am not sure anyone knows how we would establish that with confidence. The debate about model welfare is therefore worth having. But I worry that it draws our attention toward a distant threshold while another one is already passing beneath us. The socially important moment may not be when a machine becomes conscious. It may be when machines become good enough at producing the signals we associate with consciousness, judgment, care, expertise and relationship.
There is another reason Tinbergen’s experiment feels so relevant. AI does not merely have to imitate human signals. In some cases, it can intensify them.
A person eventually gets tired of listening. A machine does not have to. A colleague has meetings, deadlines, children, bad days and other people competing for attention. A machine can appear available whenever you open the screen. A friend may forget something you told them six months ago. A system with memory can surface it instantly. A teacher, doctor, manager or adviser has limited time. Software can produce the experience of patient, focused attention again and again.
This creates a strange possibility. The machine may not feel patience, but it can appear more patient than most people. It may not care, but it can produce language that feels consistently caring. It may not remember you in anything like the human sense, but it can recall details with a precision few humans can match.
So the interesting question is no longer simply whether AI can imitate a person.
It is what happens when AI becomes a more concentrated version of some of the signals by which we recognize personhood, care, confidence, memory and expertise.
Think about empathy. Real empathy is expensive. Caring about another person costs attention. Sometimes it changes your priorities. Sometimes it creates an obligation you would rather not have. A machine can produce the language of empathy at almost no marginal cost. That does not make the interaction worthless. A person may genuinely feel comforted. But the mechanism underneath the comfort has changed.
The same is true of confidence. In human life, confidence is often treated, sometimes wrongly, as evidence that a person knows what they are talking about. A model can produce confident language without experiencing confidence at all. Memory works the same way. When a friend remembers a small detail from months ago, we read something into it. We assume they listened. We assume the moment mattered. When a machine retrieves the same detail from stored context, the emotional signal can feel similar even though what produced it is completely different.
None of this means AI assistance is fake. A navigation system does not need to understand what it feels like to be lost before it can help me get home. A medical model does not need to fear illness before it can notice a pattern a doctor missed. Artificial systems can be useful, even profoundly useful, without sharing human experience.
The problem begins when we confuse the usefulness of the output with the meaning of the signals surrounding it.
This is why I think the debate about model welfare eventually connects to a much larger question about institutions. We are spending enormous effort asking what increasingly capable models are. We may need to spend just as much time asking what humans will become in relation to them.
Our equivalents of the gull’s red spot are everywhere. Fluency is a signal of intelligence. Confidence is a signal of competence. Empathetic language is a signal of care. Memory is a signal of relationship. A professional title signals authority. A signature signals consent. A familiar face signals identity. None of these signals has ever been perfect, but for most of history they were difficult enough to produce that institutions could reasonably organize themselves around them.
AI changes the cost of producing the signal.
That is where the problem becomes larger than anthropomorphism. It becomes a question of institutional design.
Imagine a manager receiving advice from a system whose analysis is more comprehensive than her own. A patient describing symptoms to an AI that appears less rushed than the physician they saw last week. A teenager asking a machine a question they would never ask a parent or teacher. An executive consulting an agent that has read every document in the company and can discuss a decision for hours without losing patience.
Each of these interactions may be useful. Some may be extraordinarily useful. The difficulty is what follows usefulness. Usefulness creates trust. Trust makes delegation easier. Delegation can become dependence. And dependence, after enough repetition, starts to look a lot like authority.
Nobody has to declare that the machine is a person for this to happen.
Organizations rarely change through philosophical declarations anyway. They change through workflow.
A recommendation becomes useful enough that nobody reproduces the analysis from scratch. An assistant becomes reliable enough that people stop checking with a colleague. A system becomes familiar enough that its suggestions begin to feel less like external advice and more like part of how someone thinks. A few years later, an organization can find itself in the odd position of insisting that the AI has no formal authority while a great deal of human behavior has quietly reorganized itself around what the AI says.
This is the part of Tinbergen’s work that stays with me. The bird’s mistake was not that it was incapable of seeing the difference between a giant artificial egg and its own egg. The problem was that its behavior gave enormous weight to a signal that had been dependable for a very long time. Once the connection between the signal and the underlying reality changed, the bird’s instincts did not immediately change with it.
Human civilization is built on recognition systems too. We have built professions, organizations, laws and relationships around signals that were reliable partly because they were hard to manufacture without the thing underneath them.
Now we are building machines that can manufacture many of those signals very well, very cheaply and at enormous scale.
That leaves us with a question I find more urgent than whether the machine behind the screen has become human enough for us to owe something to it.
What happens when the signals coming through the screen become human enough that we start handing over parts of our judgment, authority, attention and trust before we have noticed that the environment has changed?
Nature has seen versions of this before. A signal becomes detachable from the reality it once represented. Something else learns to reproduce it. The old recognition system keeps responding.
The interesting question is how long it takes us to realize that the old rules no longer work.
