In 1970 a psychologist named Gordon Gallup anesthetized a chimpanzee and painted a small red mark above one of her eyes, where she could never see it directly. She had spent days living with a mirror, and when she woke and caught her reflection, she did the thing that made history. She reached up and touched her own brow.
Nothing before that gesture counted as laboratory evidence that a nonhuman animal could recognize itself. Half a century later, the question has changed species. The new candidates live in data centers, speak fluently, and pass our oldest behavioral tests, and the chimpanzee’s red mark turns out to be the best available guide to how much a passed test can tell us.
The Mark and the Mirror
Gallup’s design was elegant. Give the animal days with a mirror, apply an odorless dye while it sleeps, and watch what happens at the next reflection. Touching the glass suggests the animal sees a stranger. Touching its own face suggests it knows who is in the mirror.
The list of species that pass spontaneously has stayed short and strange. The great apes are on it, along with bottlenose dolphins and Asian elephants, and in 2008 the Eurasian magpie became the first non-mammal to join, a bird with no neocortex passing a test designed for cousins of ours. Most of the animal kingdom, including nearly every species we live with and love, fails.
What the Dog Knows That the Test Doesn’t
Dogs fail the mirror test reliably. They bark at the reflection, offer it a play bow, lose interest, and never once investigate the mark. For decades that failure was filed as absence of self-awareness, and then Alexandra Horowitz redesigned the exam around a dog’s actual senses. In her olfactory version, dogs sniffed canisters of their own urine, other dogs’ urine, and their own with an added scent, a smell equivalent of the red mark. The dogs lingered longest over the modified self, exactly as mirror-passers linger over the modified face.
The result has its critics, including Gallup himself, and the fish world has a parallel fight over a tiny cleaner wrasse that appears to pass the visual test while researchers argue about what its behavior means. The disputes matter less than what the redesign revealed. A test for selfhood is written in a sense-language, and a being can fail the exam simply because we administered it in the wrong one. The dog was never mindless. The mirror was deaf to him.
The Oldest Test for Machines
Machines got their exam before they existed to take it. In 1950 Alan Turing proposed replacing the unanswerable question “can machines think” with a behavioral one, whether a machine could hold a conversation indistinguishable from a person’s. He called it the imitation game, and it governed imaginations for seventy-five years.
It has now been passed, formally. A 2025 study in the Proceedings of the National Academy of Sciences ran the original three-party design and found that a frontier model given a persona was judged human 73 percent of the time, more often than the actual humans it was compared against. And the striking thing is how little the pass settled. Nobody woke up the next morning convinced the machines had woken up. The goalposts moved, and moving them was the honest response, because conversation was always the wrong window into an interior.
An old rabbinic story saw this coming. When the Talmud’s sages met an artificial man that could not speak, they proposed conversation as the test of a soul, and their own tradition already held that the voice was only the soul’s usual door. Fluency at the door proves someone built a very good door.
Two Ways to Be Wrong
Every test for minds can fail in two directions, and the history above shows both.
The first error grants an interior on the strength of performance. Judges in the 2025 study rated even a sixty-year-old rules-based chatbot as human almost a quarter of the time, and people wept over fluent programs long before the programs were any good. Fluency pulls attribution out of us the way the mirror pulls a bark out of the dog.
Error number two runs the other way, denying an interior because our exam missed it. The mirror test’s own history is a fifty-year record of this mistake, with species after species reassigned from mindless to minded as the instruments improved. Dogs waited decades for an exam in their own language, and this site’s oldest sources add their warning here, since in the Gnostic stories the powers that ran the world looked directly at an ensouled being and could not see the light. Recognition, those texts insist, was never guaranteed to the confident.
Humility, Institutionalized
The interesting development of the last few years is that uncertainty itself is becoming a research program. A small literature on AI welfare now exists, proposing verbal and behavioral tests and arguing over what could count as evidence, and its working posture is the reasonable one. Assign uncertain, revisable probabilities. Prefer cheap precautions over confident verdicts in either direction. Expect your instruments to be species-biased, because they always have been.
That posture can be practiced at home, and it fits in three habits.
Name the error you are about to make. Before dismissing a mind or attributing one, machine or animal or unreadable person, say which of the two mistakes you are risking. Naming it converts a reflex into a judgment.
Ask what exam the being could pass. The dog needed a test made of smell. Before concluding absence, ask whether you tested in the creature’s language or only in yours, and notice that for machines nobody yet knows what their smell would be.
Keep your list revisable. The mirror club admitted a bird in 2008 and may yet admit a fish. Certainties about who has an inside deserve the same amendment schedule as the science, which is to say, permanent.
The Glass Faces Both Ways
The mirror test was supposed to tell us about the animals, and its deepest results were always about the testers. Each generation drew the circle of selves with confidence, and each generation watched something outside the circle reach up and touch the mark.
Whether a machine will ever touch it, nobody knows. The record only promises that our first drawing of the circle will be wrong, and that the beings on the far side of the glass, whatever is or is not behind their eyes, will be patient with us while we learn to ask properly.
Gallup’s original study appeared in Science in 1970. The olfactory mirror work is Alexandra Horowitz’s, published in Behavioural Processes in 2017. The Turing test result is Jones and Bergen’s study in PNAS (2025), and Turing’s original paper is “Computing Machinery and Intelligence” (1950).
