Two news stories were making the rounds this week that no one thought to put side by side. The first was about chicks and beads: researchers from the MIPT Institute of AI and Moscow State University taught a hidden Markov model to predict how a bird learns to distinguish edible grain from an inedible bead, and along the way discovered that a memory formed in the right hemisphere migrates over time to the left—but not the other way around. The second was about newborn chicks recognizing a face in a schematic drawing despite having no visual experience whatsoever.
Both look like cute ethological curiosities. But really, they are two ends of one long story about why we need the chick at all. Because a newly hatched chick is not a “simple little bird.” It is the cleanest instrument we have for catching the moment when a brain becomes knowing.
Let’s start with how this is done technically, because it is the details that make all the philosophy.
A Ross 308 chick is hatched in a laboratory in darkness—37.7 degrees, 60% humidity, an incubator, no light. Immediately after hatching, a black cap is put on its head: so that not a single photon reaches its eye before the experiment begins. It is then placed in an individual cage where the only light source in the whole universe is a hollow plastic sphere three and a half centimeters in diameter, suspended seven centimeters above the floor. Inside the sphere, two optical fibers flash at 5 hertz.
Then the real magic begins. For one group of chicks, the sphere glows red—with a dominant wavelength of 642 nanometers. For another, blue, 465 nanometers. The intensity is matched in both cases: 16.45 candela per square meter. So the chick is not merely learning. It is being given its first calibrated picture of the world, with its mother replaced by a lightbulb of a known spectrum.
Why does it flash? Because chicks prefer to imprint on a flickering object rather than a stationary one—this was established back in 1959. Imprinting is possible on any object at all, but biology feeds the animal clues: red coloring around the head of a fellow creature, self-propelled motion, face-like configurations. These are innate filters of attention, and the chick has them from its first minute.
What happens next was described in the 1970s by Gabriel Horn of Cambridge. The key region is the intermediate medial mesopallium, IMM: this is where the imprinting trace forms. But then there was an oddity. Sequential lesions—first on one side of the IMM, then on the other—showed that the record survives temporarily in the left IMM, while the right IMM transfers information somewhere else—to an unknown region Horn called S'. The transfer is completed roughly six hours after the end of exposure. And that was it. For forty years, they searched for this region blindly, by elimination, with nothing in their hands but autoradiography and a knife.
Horn himself put it with a candor rare in a great scientist: “We are still far from being able to see—through a microscope or by methods of neuroimaging—the neural trace of memory.”
To understand the answer, you need to appreciate the price paid for the instrument.
The paper, published in Communications Biology on 15 October 2024, brought together two laboratories—Ruhr University Bochum (Medi Behroozi, Onur Güntürkün) and the Center for Mind/Brain Sciences in Trento (Elena Lorenzi, Giorgio Vallortigara)—along with the Institute of Technology in Rovereto. They decided to do what no one before them had done: real fMRI on an awake newborn chick.
The scanner: 7 tesla, a horizontal Bruker, a gradient system of 380 millitesla per meter. The receiving coil: single-loop, 20 millimeters, wrapped around the head. Nine coronal slices, 540 volumes per animal, around 36 minutes for each one.
Immobilization is the most interesting part of the whole method, because it involves neither anesthesia nor surgery. The chick is wrapped in a paper towel so it cannot move its wings or legs. Its head is packed around with blocks of modeling clay, which also cover its ears—solving the second problem, the scanner noise. The beak is secured in a custom holder. That is all. Zero implants, zero skull openings.
There are two noise problems, incidentally. The first is dealt with beforehand: before scanning, the chicks are played a recording of the magnet twice a day, for up to five hours total per day—habituation. The second problem is engineering, and almost comical: a chick’s eyes are comparable in size to its brain, and their movements wreck the BOLD signal. The solution: two “saturation” slices are manually placed directly over the eyes.
Then comes the art of measurement. The median frame-to-frame head displacement is 0.03 millimeters. Volumes with displacement greater than 0.2 mm (that is ~40% of a voxel) make up two percent during the task and one percent at rest. Three animals out of twenty-nine were discarded for fidgeting. The holder was deliberately designed to allow movement in the dorsoventral direction, “so as not to block the throat”—meaning the construction balances precision against breathing.
And then there is a separate story about the reference image. To have something onto which to overlay the functional maps, five dead chick brains were scanned for 19 hours 48 minutes in total, using a 400×400×500 matrix with a voxel size of 0.05 cubic millimeters. One reference image—nearly twenty hours of scanner time. Plus a 1988 anatomical atlas; plus, because the chick has no hemodynamic response function of its own, they used as a template an HRF recorded from a pigeon, the structurally closest brain.
That is the price of it: to catch knowledge at the moment of birth, they needed 7 tesla, five hours a day of recorded noise, twenty hours scanning dead tissue, and the response function of another species.
Now for what this was all about. Nine chicks were imprinted on red, eight on blue. In the scanner, they were shown both colors in 16-second blocks, with 24 seconds of darkness between blocks, 48 trials. The logic was simple: if the memory exists, their own color should produce a stronger response than the other one.
The “own minus other” contrast yielded not a single significant cluster. Not one. In either group.
This is the moment when the experiment either went into the trash or became more interesting than intended. The scientists went to test the behavior separately—twenty-four chicks, a running wheel, ten minutes, red on one side and blue on the other. Both groups preferred red: 0.638 against a neutral 0.5, t(23) = 2.683, p = 0.013. There was no difference between the red and blue groups (p = 0.124).
That is the trick. Red is an innately preferred stimulus. It is the color a chick sees around the head of one of its own kind. The blue chicks, put into the scanner and shown red, began secondary imprinting right there in the tube. They were not recalling their first mother. They were acquiring a second.
When the data were recalculated by phase, it became obvious. In the first ten minutes, the blue group showed weak activity in the same regions as the red group. In the final twenty minutes, there was mounting activity in response to red in precisely those zones that the literature links to the earliest phases of imprinting: IMM, hippocampus, medial striatum, NCL, nucleus taenia. The scanner did not catch memory retrieval. The scanner caught the dawn of a second one.
And yet S' was found—not as a region. As a network.
During memory retrieval in the red group, an entire constellation was activated: the right dorsal lateral geniculate nucleus (GLd—the one receiving direct input from the retina); bilaterally, the IHA, HI, HD, and HA nuclei—the entire thalamofugal visual pathway from the thalamus to the visual Wulst. Plus the associative IMM and NMm, plus NCL, plus one curious detail: parts of the auditory system—Field L and the right nucleus ovoidalis, the thalamic auditory nucleus. And that makes sense: before scanning, the chicks had been played magnet noise for five hours a day, so the acoustic component of the image became embedded in the engram.
Then come the networks the paper calls the Social Behaviour Network and the Mesolimbic Reward System: BNST, nucleus accumbens, medial striatum, posterior pallial amygdala, ventromedial hippocampus. This is the motivational layer. Imprinting is not a password written into memory; it is attachment, and attachment must be paid for in dopamine currency.
The lateralization turned out to be precise rather than slogan-like. Activity was predominantly right-sided. But the left hippocampus was involved in formation, while during retrieval the hippocampus was working on both sides. The hippocampus appears here in an unexpected capacity: according to c-fos studies, its dorsomedial and ventromedial parts are responsible for recognizing a specific individual. In other words, for the chick the hippocampus is not just a navigator but a social card index. Incidentally, in 2015 Rugani and Vallortigara showed that a newborn chick lays out numbers along the same mental number line as humans: small on the left, large on the right.
The asymmetry is read honestly in the paper: both hemispheres are involved, but the phases are split between sides—encoding and consolidation occur in different hemispheres. This directly echoes the human HERA model, where the left dominates during encoding and the right during retrieval.
The authors’ conclusion sounds almost like an epigraph: the elusive S' is probably not a structure, but this entire distributed network, whose central hub may be considered the “prefrontal” NCL.
To understand how “prefrontal” the NCL is, it is worth looking at a paper from the same Trento school, published in PNAS. Twelve Ross 308 chicks, the same dark incubators, the same 37.7 and 60%, the same commercial CRESCENTI incubator outside Brescia. The animals were eight to twelve days old—the youngest creatures in which number neurons have ever been described at all. Then comes honest neurophysiology: 17.78-micrometer nichrome tetrodes, a Halo-5 microdrive, implantation into the right NCL under Vetflurane, stereotaxy with the beak at 45 degrees.
The result: of 471 recorded units, 53 (11%) proved selective for quantity—and it did not matter how the dots were presented, whether radius, area, or perimeter was held constant. Against 0.95% false positives after a thousandfold shuffling of trials. Most neurons preferred one (30%) and five (32%). Numerical distance worked better on a logarithmic scale than an absolute one (ΔBIC = 6.36).
And most importantly: the chicks were taught nothing. No numerical task, no reinforcement. They simply watched a screen. This is the same 14% of number neurons found in adult macaque prefrontal cortex, and the same 12% found in adult crows. The authors do honestly qualify this, however: uncontrolled learning from repeated stimulus presentation cannot be entirely ruled out—and they note that similar number detectors arise spontaneously in deep networks, without pretraining.
And the final layer, without which the previous two make no sense.
In 2016, a team from Charles University in Prague, working with Suzana Herculano-Houzel, used the isotropic fractionator—a method in which brain tissue is turned into a homogeneous “soup,” and its nuclei are counted, separating neurons from glia without sections or extrapolation. Twenty-eight bird species.
The result, which has upended the usual arithmetic since 2016: parrots and songbirds have on average twice as many neurons in their brains as primates of the same mass. In corvids and parrots, a disproportionately large share of neurons sits precisely in the forebrain, so that by pallial neuron count they catch up with or surpass monkeys with much larger brains. The avian strategy is not to increase volume, but to pack more densely.
That is why the chick works as an instrument. Its brain is not a “small copy” of a rat’s: it is a different school of engineering, where a high concentration of computation is fitted into a small space, and it hatches almost ready for work—precocial. A day-old mouse does not exist as a behavioral unit. A day-old chick does, and it can go into a scanner without anesthesia.
Put it all together, and a fairly hard thought emerges.
We are used to thinking that the “innate” is primitive, while the “learned” is complex. The chick overturns that ladder. It has not had a single day of visual experience, yet by the second or third day of life, its brain already runs a distributed network holding the trace of attachment, and neurons selective for number sit in its highest associative region—that is, for an abstraction thought to require cortex. Except it has no cortex. It has a pallium, built differently.
Second. The S' that was found is not a point on the map but a network. And this seems to be the broadest lesson: when a model has enough complexity to learn something like attachment, the engram ceases to lie in a single drawer. You can neither cut it out nor photograph it. Horn searched for a structure for forty years and did not find it because he was looking for a structure. Anatomy did not fail him; the question itself did.
Third, and perhaps most interestingly—secondary imprinting. The innate preference for red did not help the experiment; it broke it: instead of measuring memory, the instrument was measuring the birth of a new one. It is a beautiful illustration of the fact that a “clean start” is impossible even for a chick that has not seen a single photon before the experiment. The brain does not wait for data; it comes out to meet it with an antenna already tuned. And when the data that arrive are not the data it expected, it does not reject them—it begins learning anew, right inside the tomograph.
It is also worth mentioning what stood out while reading the literature: the lines of work from the Bochum–Trento laboratory and the Moscow school of the “bead floor” (Diffiné, Tiunova, Anokhin; in 2008, Komissarova and Anokhin published on the influence of imprinting on cell proliferation in the chick brain) exist almost in parallel. The 2024 paper’s bibliography contains no references to the “bead floor.” This is neither a dispute nor a grievance, just a simple observation: even with an object as small as a chick, scientific traditions somehow manage not to meet.
And finally, something not quite scientific. Imprinting began as a story about ducklings running after Konrad Lorenz’s boots—a curiosity. Then Bowlby built a theory of attachment through it: an explanation of how an entire later psyche grows out of the first encounter. Then it became a way of trying to understand where memory is stored at all. And now they have attached a 7-tesla tomograph to it, to catch knowledge appearing in the brain with head tremor measured to thirty microns. A strange trajectory for an experiment about a flashing little ball. But perhaps this is exactly what science looks like when it has finally reached the question from which it all began.