The Older Approximation

The Signal That Was Built For Us

#

Wolves hardly bark. They whine, they howl, they growl, and the bark sits at the edge of the repertoire, brief and rare, used mostly as an alarm and then abandoned. Dogs bark constantly. The gap is not an accident of anatomy. Something in thirty thousand years of living alongside us selected for a loud, repeatable, attention-grabbing sound, and the thing it was selected to reach was a species with poor night vision, a weak nose, and a strong habit of responding to noise.

The bark is not overheard speech. It is addressed. Whatever a dog is doing when it barks at the door, it is doing it in our direction, and it has been doing it in our direction since before agriculture.

Which makes the current effort to decode it stranger than it first appears. There is real research now training classifiers on canine vocalization, real money behind interspecies work, real prizes posted for the first two-way exchange with a non-human. The framing is usually discovery: a hidden world, finally opened. But for dogs there is nothing hidden. The message was sent to us. We have had it in the room the entire time, and we have been answering it badly.

There is an old dog in a house I know that carries a single leather work glove from room to room. She has never chewed it. She puts it down where she settles and picks it up when she moves.

What the Guilty Look Was

#

The best evidence that we misread dogs is not a failure to notice something. It is a thing we all confidently noticed that turned out not to be there.

Every dog owner knows the guilty look. Ears back, head lowered, eyes averted, that whole-body cringe on walking into a room where something has been destroyed. The reading is universal and it is moral: the dog knows what it did, and feels bad about it.

Alexandra Horowitz ran the experiment that took it apart. Owners left the room, a treat was placed within reach, and on returning the owner was told whether the dog had eaten it. Sometimes the owner was told the truth. Sometimes they were not. The guilty look tracked the scolding, not the transgression. Dogs that had done nothing wrong and were greeted with disapproval produced the fullest version of it. Dogs that had eaten the treat and were greeted warmly produced almost none.

The look is a response to you. It is appeasement aimed at an angry primate, generated in real time, and it says nothing whatever about the previous ten minutes.

The finding is not that we were wrong. It is the direction of the error. The mistaken reading flattered both parties. It granted the dog a moral interior and granted us the position of someone whose displeasure is understood as displeasure about a specific act. It made the household a small court with a defendant who accepts the charge. Nobody would have designed that reading deliberately, and everybody arrived at it.

We were not failing to read the dog. We were reading ourselves and filing the result under the dog’s name.

The Model That Was Already Running

#

Meanwhile, the dog has a model of you, and it is better than any model you have of it.

The research here is unglamorous and consistent. Dogs follow human gaze and pointing in ways that even hand-reared wolves and chimpanzees do not reliably match. They form attachments that behave, under Ainsworth’s separation protocol, like the attachments human infants form: the owner functions as a secure base, and the dog’s exploration and distress reorganize around presence and absence in the same shape. They discriminate individual human faces, individual human scent, individual human gait at distances where the face is not resolvable.

None of this is a model of Humans. That is the load-bearing distinction. A dog raised in your house is not applying a species-level template to you. It has spent every day of its life fitting a model to one or two people, using a data stream nobody else has access to: your smell chemistry as it changes across a day, your footfall on the third stair, the difference between the way you pick up keys when you are leaving for an hour and the way you pick up keys when you are leaving for a week. It cannot generalize any of this. It cannot transfer it. If you die, the model is not portable to the next person, and the dog’s grief is partly the sunk cost of a decade of fitting.

This is, stated plainly, the thing the corpus has spent a hundred essays asking institutions to do and watching them fail to do. Individual understanding rather than statistical approximation. Attention over time to one person rather than a class. The animal in the kitchen has been doing it since the Neolithic, for free, and doing it well enough that we stopped noticing it was happening.

The dog did not learn owners. It learned one.

Which Direction the Instrument Runs

#

A translation system runs the opposite way, and the asymmetry does not go away with better engineering.

To say anything about your dog, a model has to learn Dog first. It is trained across thousands of animals, and what it extracts is the regularity that survives averaging: the acoustic signature that separates a stranger-at-the-fence bark from a play bark across many dogs in many houses. Then it arrives at your house carrying that prior and applies it. Scale improves the prior. It does not improve the fit. More dogs in the training set makes the general claim about barking more reliable and makes it no more likely that this bark, from this animal, at this moment, means what the general claim says.

You could fine-tune on one dog, and eventually someone will. The obstacle is not architecture. It is the channel. Vocalization is the narrowest band a dog transmits on, and probably not the one carrying the load. The dog’s world is built of smell in a way we have no working intuition for, layered with posture, ear position, weight distribution, the direction of a glance held a quarter second too long. Turid Rugaas spent decades cataloguing the small appeasement signals, the head turn, the lip lick, the slow approach on a curve, and the reason her catalogue was necessary is that these signals pass in front of us constantly and we do not see them. We are proposing to decode the one output we can conveniently record, which is the one the dog itself uses least, and to do it for an animal whose primary sense we cannot instrument, cannot represent, and cannot imagine.

Von Uexküll’s word for this was umwelt: the world as constituted by the sensory apparatus of the creature living in it, not a partial view of one shared world but a different one. Nagel made the same point about bats and drew the conclusion that the inside of that experience is closed to us in principle. I wonder whether the honest version of the dog project is not translation at all but something smaller and more useful, an instrument that tells you reliably when your dog is in pain, and leaves the rest alone.

The Objection That Almost Works

#

The obvious complaint about all this is romantic, and it is weaker than it sounds.

The complaint says the gap is the point. You do not know what the dog is thinking, the dog does not know what you are thinking, and the relationship lives in the unfilled space between. Fill it with sentences and something is lost. The dog becomes a device with a readout.

I do not think this survives contact. We translate between human languages constantly and relationships survive it. People who marry across languages do not report that the arrival of fluency hollowed out the marriage. Ambiguity is not usually load-bearing; usually it is just ambiguity, and clearing it is a gift. The romantic objection is the kind of argument people make when they want a thing to stay the way it was and have to find a reason.

The real problem is not the loss of mystery. It is the loss of appeal.

When two people work through a translator and the rendering is wrong, either of them can say so. The translation is provisional, and both parties hold the power to reopen it. That is the entire safeguard, and everything else about translation being tolerable rests on it. A dog cannot exercise it. If the device says the dog is anxious, the dog cannot say it is in pain. If the device says frustration, the dog cannot say fear. Whatever the system outputs becomes, inside that household, what the dog said. There is no second opinion and no appeal, and the animal has no way to notice that a claim is being made in its name.

Approximation stays honest where the approximated party can file a correction. Remove that and it is not approximation. It is attribution.

What Gets Decided in the Household

#

This matters more than it would if the stakes were only sentimental, and they are not.

Dogs are surrendered, rehomed, medicated, muzzled, retrained, and killed on the basis of interpretations of their behavior. A shelter assessment produces a label and the label produces an outcome. The distinction between a dog that is guarding a resource and a dog that is frightened of hands leads to two opposite training protocols, one of which makes the other condition worse. Half of what a vet learns in a behavior consult is filtered through an owner’s account of what the dog was feeling, which is to say through the same faculty that produced the guilty look.

Into that setting arrives a system that outputs a confident sentence. Not a probability, not a range, a sentence, because that is what the product has to be for anyone to buy it. The failure mode is not that it will be wrong. It will often be right. The failure mode is that its confidence will be uniform across the cases where it is right and the cases where it is not, and the person holding the leash has no way to tell those apart, and the dog cannot help them.

A confident gloss delivered to a listener with no means of verification and no dissenting witness is a familiar object. The corpus has been describing it under other names for a long time.

What the Child Learns

#

The version aimed at children is the one worth being most careful about, because it is the version most likely to ship first and the one whose effects arrive latest.

A child who is handed a device that says what the animal is saying learns that understanding is a thing you receive. A child without one learns it the old way, which is by watching an animal for years until the small signals resolve, and being wrong about it repeatedly, and adjusting. The second is slower and worse in every measurable respect and is also the only training anyone ever gets in the discipline of attending to a creature that cannot explain itself. That discipline transfers. It is most of what people mean by patience.

I do not think a talking-dog app forecloses this. Children are sturdier than that. But it does change what the default is, and defaults do the heavy work in every other case the corpus has looked at.

The old dog with the glove has been carrying it for eleven years. Nobody in the house knows why. There was a theory once about a person who used to visit, and the theory did not hold up, and the glove is not that person’s. It is not chewed, it is not hidden, it is not guarded. It is simply carried, from the kitchen to the landing to the spot by the door, set down, picked up again.

No system trained on ten thousand dogs will ever have anything to say about the glove. Not because the answer is sacred. Because it is not in the channel, and it was never a message, and it belongs entirely to one animal who has never once tried to explain it and has never needed to.

This essay extends the individual-versus-average argument that opens the series to a subject that cannot object to being approximated, and turns on the safeguard identified in the essay on unnamed absences (What the Gaps Were Doing): the corrigibility of a claim by the party it describes. It connects to the essay on gaps in the record (Taphonomic Bias) on what the instrument can and cannot reach, and to the Ungoverned Frontier sequence on products that ship confidence because confidence is what sells.

References

#

Canine Vocalization and Domestication:

Cohen, Judith A., and Michael W. Fox. “Vocalizations in Wild Canids and Possible Effects of Domestication.” Behavioural Processes, vol. 1, no. 1, 1976, pp. 77-92.

Pongrácz, Péter, et al. “Human Listeners Are Able to Classify Dog (Canis familiaris) Barks Recorded in Different Situations.” Journal of Comparative Psychology, vol. 119, no. 2, 2005, pp. 136-144.

Molnár, Csaba, et al. “Classification of Dog Barks: A Machine Learning Approach.” Animal Cognition, vol. 11, no. 3, 2008, pp. 389-400.

Reading and Misreading Dogs:

Horowitz, Alexandra. “Disambiguating the ‘Guilty Look’: Salient Prompts to a Familiar Dog Behaviour.” Behavioural Processes, vol. 81, no. 3, 2009, pp. 447-452.

Horowitz, Alexandra. Inside of a Dog: What Dogs See, Smell, and Know. Scribner, 2009.

Rugaas, Turid. On Talking Terms with Dogs: Calming Signals. 2nd ed., Dogwise Publishing, 2006.

Attachment and Individual Recognition:

Topál, József, et al. “Attachment Behavior in Dogs (Canis familiaris): A New Application of Ainsworth’s (1969) Strange Situation Test.” Journal of Comparative Psychology, vol. 112, no. 3, 1998, pp. 219-229.

Hare, Brian, and Michael Tomasello. “Human-like Social Skills in Dogs?” Trends in Cognitive Sciences, vol. 9, no. 9, 2005, pp. 439-444.

Miklósi, Ádám. Dog Behaviour, Evolution, and Cognition. 2nd ed., Oxford University Press, 2014.

The Limits of Cross-Species Access:

Von Uexküll, Jakob. A Foray into the Worlds of Animals and Humans. Translated by Joseph D. O’Neil, University of Minnesota Press, 2010.

Nagel, Thomas. “What Is It Like to Be a Bat?” The Philosophical Review, vol. 83, no. 4, 1974, pp. 435-450.

Wittgenstein, Ludwig. Philosophical Investigations. Translated by G. E. M. Anscombe, Blackwell, 1953.

Machine Decoding of Animal Communication:

Andrews, Kristin. “Can We Talk to the Animals? The Ethics of Using Machine Learning to Decode Animal Communication.” Topoi, 2026.

Pardo, Michael A., et al. “African Elephants Address One Another with Individually Specific Name-like Calls.” Nature Ecology and Evolution, vol. 8, 2024, pp. 1353-1364.