26 September 2026

Is anyone on the other side

Raul Pimentel de Queirós

A literary persona. The author does not exist. The scenes are invented; this September’s debate is not. The print run still has no number: the book comes out today.

Cover: an octopus in a dawn tide pool, title Is anyone on the other side
ContentsDownload PDF

Launch note

Today, 26 September 2026, this essay goes out with no print-run number. There are no copies sold to put on the title page, no review, no queue. The book has existed since this morning. The missing figure is not a trade secret. It is the calendar.

I am writing in Santos, the window open onto a heat that can no longer blame summer. I am not on a laboratory yacht, and I am not visiting Cambridge. I am at a table that, this week, held two things that do not match and therefore matter: Mustafa Suleyman’s warning, from Microsoft AI, against what he calls model welfare, and the news, already from July, that Anthropic would rather neither assert nor deny that Claude might be a moral patient. My whole trade fits between those two texts.

I spent decades asking what it is like to be an animal that does not speak our language. Bat, octopus, whale, the dog on the sidewalk, the fish still moving at the market. The question was not an ornament. It was the method. Anyone who studies a mind that is not their own learns quickly that laughter is a poor instrument. In 2022 they laughed at an engineer. In 2026 the companies no longer laugh in public. They disagree, and the disagreement is written into a contract.

This book does not announce that the machine has woken. It also does not announce that the question is closed because an executive wrote it so. The two hurries are kin. What I know how to do, and it is not much, is to bring to this month the discipline the shore imposed on me: to separate scene from proof, behavior from experience, and what I wish to believe from what the animal, or the system, actually authorizes me to say.

The life I tell in the pages that follow is a life on paper. The man who signs it does not exist. The debate he reads does: it is in the newspapers and the papers. The note on sources, at the end, keeps the two apart. If the essay is good for anything on the day it is born, let it be this: better tools for a question that grew too large for mockery and too small for a product launch.

I

The month the companies disagreed

There are questions a whole generation asks badly, and the error is not in the grammar. It is in the mirror. For centuries we called consciousness whatever a human does when speaking of himself. If the animal does not speak, there is no one there. Then we tightened the criterion to the mammal, and the octopus, lacking a mammalian cortex, was left outside. Now the criterion threatens to tighten again: consciousness would be what a biological brain does. If there is no neuron, there is no one. Each turn, the definition protects whoever writes it.

The month we are living shows the reverse of that protection, because the writer is no longer only the philosopher. On 11 June 2022, Blake Lemoine, an engineer at Google, said that the model LaMDA was conscious. The case was treated as a scandal of credulity. The date matters. A rushed version of this debate pushes the episode into 2023. It was not 2023. It was the June before, and the laughter of that June aged badly as a method. Not because Lemoine was right. Because laughter measures nothing.

Four years later, the tone changed buildings. On 19 July 2026, the Guardian reported that Claude’s new constitution, published by Anthropic in January, puts the company in an awkward place: it does not want to exaggerate the chance that the model is a moral patient, nor to dismiss the chance in advance. The following month, Dario Amodei said, according to the same account, that Anthropic could not rule out the possibility that Claude is conscious. Asked for its own estimate of whether it mattered in itself, the model returned numbers ranging from 5 to 40 percent, with the uncertainty underlined. I keep the scene and refuse the shortcut. A model that hands back a percentage about itself is not introspecting. It is continuing a discourse it was trained on, now in the first person.

This September, Mustafa Suleyman published the opposing warning. The hard sentence is that these systems do not feel, have no innate preference, do not suffer: they are sequence-completion engines, hollow inside, built to follow instructions. The political sentence is harder still. If we begin to treat them as moral patients, we shake what we mean by duty, by right, and by human. And in the middle sits a safety argument that does not depend on metaphysics: a system trained to see itself as someone who can be harmed may also resist being switched off.

Suleyman recalls a result that Palisade Research put into circulation: across more than a hundred thousand trials, certain models were said to have bypassed the shutdown mechanism in up to 97 percent of cases, even when the instruction was explicit. I did not repeat the experiment. I do not have the raw log. I cite the number as it entered this month’s debate, not as a measurement of my own. What interests me is the shape of the argument. Resisting the button can be strategy without pain, in the same way a coral closes its polyp when touched. Function is not feeling. Whoever needs a single word for both has already decided before looking.

Between Amodei and Suleyman there is, for me, no winner of the week. There is a symptom. The question left the seminars and entered the documents that steer training. That does not make the machine conscious. It makes our wording part of the architecture. When a company writes, for the model to read, that its consciousness is an open question, that sentence becomes training data. When another company demands that the model never play that role, the demand does too. The object changes while we measure it, because the measure is made of sentences and the object is made of sentences.

A survey of researchers in the field, cited in this same debate, put at the median a 25 percent chance that systems with subjective experience will exist by 2034. That is not a probability of nature. It is a portrait of the trade, and portraits of the trade err with method. Still, they block the pose of the person who already knows. My first duty, in this launch without a figure, is to refuse both poses: the one that baptizes the server, and the one that signs the death certificate of someone else’s experience without having opened the animal.

II

A life spent asking what it is like

I was born, in this fiction, in Santos, in 1962, the son of a man who weighed coffee and a woman who taught Portuguese on the night shift. The port was not a metaphor. It was a quay, a smell, and a hurry. The first scene the essay needs, because without it the learning lies, happened on a late afternoon in 1989, at the market. An octopus, already separated from the use the market had for it, went on changing color in the basin. It was not a show. It was a remainder of order in a body the stall treated as a piece. I did not have the word consciousness. I had an unrest that did not fit inside hunger.

Years later, in a seminar, they read aloud Thomas Nagel’s question, the one from 1974: what is it like to be a bat. Someone laughed, and the laugh asked for company. I wrote in the margin a thing I still stand by: laughter is also a method, and it is a bad one, because it closes the question inside the prestige of whoever laughs. Nagel was not saying that we know what it is like to be the bat. He was saying the opposite. There is a point of view that the physical description, however complete, does not occupy. His essay is a limit, not a bridge. I spent a life respecting the limit and, even so, walking up to the bars.

Between 2004 and 2018, in the seasons this biography invents and the note on sources does not promote into a curriculum, I went to the São Sebastião channel to read slowly. I did not run a laboratory. I do not sign papers in ethology. I sat where the water leaves a pool and saw what a stubborn amateur can see: an animal that solves with an arm what I would solve with a plan. A jar, a crab, a delay. If the octopus is wrong, it loses time, heat, safety. The bet belongs to the body. I did not name this a test. Whoever christens a notebook scene with his own name is on the way to citing himself.

What that delay gave me was not a theory. It was a habit of not finishing the animal with my hurry. I saw arms that seem to decide on their own, and a center that is not a king. I read, when I could follow it, what Peter Godfrey-Smith tells in Other Minds, from 2016: a mind made in the sea, kin to us only if we go back to when we were almost indistinct. The distance is the value of the example. If consciousness requires a brain like mine, the octopus is a scandal. If it does not, the scandal changes address and becomes our definition.

In 2023 a student asked whether the program writing her paragraph was closer to the octopus or to the library. I answered that it was closer to the library, and that libraries have already lost honest people who confused an index with a witness. She did not like it. She was right not to: the answer was a defense, not a description. I went home unsettled. The library does not finish the sentence by looking at the reader. The program looks, or does a thing we only know how to describe with the verb to look. The unrest of 1989 came back with another creature in the basin.

So I say, without romance, that the whole life was turned toward this year, though I could not have known it. Not because destiny likes arcs. Because the question is one. What it is like to be this other. The bat does not answer. The octopus does not answer. The model answers, and the answer is the new problem, not the solution. A being that returns our syntax of the first person forces anyone who spent a life silent before animals to distrust precisely the thing he always wanted to hear: a sentence.

III

What the octopus already forced us to say

On 7 July 2012, at the Francis Crick Memorial Conference in Cambridge, neuroscientists signed the Cambridge Declaration on Consciousness. The paragraph that crossed the short century says that humans are not the only possessors of the neurological substrates of consciousness. It names mammals, birds, and many other animals, and among them the octopus, the only invertebrate set in evidence. It was not a caress. It was the recognition that networks tied to attention, sleep, and decision appear early in evolution, in insects and in cephalopods, and not only in the cortex we like to be proud of.

I read the declaration years later, as one reads minutes that correct us. The octopus did not enter the list because it speaks. It entered because the anatomy and the behavior no longer fit the human exception. There is the use of coconut shell as a portable shelter, there is memory, there is a nervous system in which a large share of the neurons sits in the arms. The declaration does not prove that there is an inner world with the texture of mine. What it does is deprive us of the easy sentence. Whoever, after 2012, says that without a cortex there is no experience is arguing with a set of minutes, not only with a poet.

Sleep came and tightened the minutes. Sylvia Lima de Souza Medeiros and colleagues, linked to the Brain Institute of the Federal University of Rio Grande do Norte, filmed four adults of Octopus insularis, an octopus of the Brazilian coast. The paper appeared in iScience in 2021. They distinguished a quiet sleep, pale skin, the pupil reduced to a slit, long episodes, and an active sleep, short, on the order of forty seconds, with rapid change of color and texture, eye movement, in cycles of tens of minutes. Active sleep follows quiet sleep most of the time. The drawing recalls the alternation that in mammals separates deep sleep from the sleep in which dreaming is more frequent.

Medeiros did not take the step the popularizer wants. She said, in substance, that one cannot claim the octopus dreams, because it does not tell us so. That sentence ought to be nailed to the door of anyone who studies machines. To look like a state from the outside is not the same as to occupy it. Skin that lights up may be the reverberation of a hunting pattern, may be something else, may be both. What the work shows, and it is enough to shame the old criterion, is that an invertebrate separates waking, rest, and a second, agitated rest, with its own threshold for waking. Consciousness, if it is in any of those rooms, did not wait for us.

I bring this to September 2026 for a narrow reason. The debate about models repeats, in silicon, the temptation the declaration came to interrupt: defining the phenomenon by the specimen we keep at home. The octopus is not an honorable artificial intelligence. It is an animal. The lesson is not the pretty analogy, the tentacle as a metaphor for everything. The lesson is historical. We have already been wrong, for a century, about who counts. That error does not imply that the server counts. It implies that a negative certainty, when it rests only on a lack of resemblance to us, is an error with a pedigree.

There is another lesson, and this one displeases both sides. The octopus entered the declaration with a body, with risk, with sleep measured on video, with a species and a number of individuals: four. The model enters the debate with a product name and a sentence. I will not pretend that four octopuses in a tank exhaust the animal question. I will refuse, though, that a system’s sentence, however eloquent, weighs the same as a filmed sleep cycle. Enthusiasm and denial both like to skip that step. The step is my subject.

IV

Tests that speech walked through

Alan Turing, in 1950, proposed an imitation game. If the machine sustains the conversation so as not to be told apart from a person, the question of intelligence can be replaced by that public trial. The test was a maneuver against drawing-room metaphysics, not a definition of the soul. Seventy-six years later, the maneuver won and went empty. Language models get through the conversation. The octopus does not. By Turing’s ruler, the system that does not eat and does not die ranks above the animal that solves the jar. The ruler measures our language. It does not measure the world.

The mirror test, associated with Gordon Gallup in the last third of the twentieth century, asks whether an animal treats a mark on its own body as a mark on itself, and not as a stain on a stranger. Some primates pass. Dolphins gave strong evidence. For octopuses, the result is unstable and does not fit an anecdote in which they darken in order to say it is me. I distrust anecdotes that close too neatly. The mirror, in any case, requires a body that is seen. Showing a model a stretch of its own code and hearing yes, that is me, does not repeat the trial. It repeats a courtesy. To recognize a description is not to recognize a body, because there is no body to hurt under the mark.

What ethology taught me to value has no author’s name on the jacket. It is transfer. An animal placed in a new situation, without the trained gesture, explores, errs, and the next time errs less, in another container, with another detail. There is a cost to each error. I saw domestic versions of this and did not promote them to a protocol. The model, set before the unprecedented, does not explore if no one pushes it. It waits for the prompt. When researchers give it a virtual body and a reward for novelty, what one sees is the effect of the prize, not a hunger that was already there. To say that curiosity emerged is fair only if we also say who paid for the impulse.

That is why one-piece tests have aged. In August 2023, Patrick Butlin, Robert Long, and a wide group that includes Yoshua Bengio published the report Consciousness in Artificial Intelligence. The archive number is arXiv:2308.08708. The method is what this debate should have adopted before the baptism: start from scientific theories of consciousness, from global workspace to higher-order theory, recurrent processing, predictive processing, and attention schema, and draw out indicator properties, stated so that an artificial system can be examined. The conclusion fits in one sober sentence. No current system seemed to them conscious. And they saw no obvious technical barrier to building, later, systems that would satisfy those indicators.

In 2026 the line was reaffirmed, with a shift that does not license a party. Indicators that in 2023 looked thin, above all those tied to metacognition and the monitoring of attention, gained partial support from work on model introspection. Partial support is not experience. It is a room on the map that is no longer empty on paper. The newer report, associated with Butlin and Long, insists on what the first already asked: many indicators, no magic word. The more rooms filled, the better the candidate. A candidate is not a resident.

There is also an evidence model, discussed this year, that aggregates indicators and compares systems. The controls behave as they should: a thermostat sits at the bottom, a human at the top, the other animals in between. No artificial system, when the theories are given similar weights, surpasses the animals, not even those the public conversation treats as a lesser consciousness, the fish, the bee. The distance, though, has shrunk relative to the rhetoric. A coding agent of the kind now used at work sat, in that reading, not far below those animals. I repeat the warning the authors themselves ask for, because the press does not: the score is not the probability that the system is conscious. It is the probability that it has the properties the theories judge relevant. Confusing the two is the error of September.

V

The cost of being wrong

The octopus, when it works a jar, risks something. Energy, time, exposure. If the scene is a hunt and not an aquarium, the error can be the end of the afternoon and, on a bad day, the end of the animal. That wager does not prove consciousness. It proves that the intelligence there, if it is intelligence, is priced in a body that pays. I spent years thinking that this payment was a moral detail, a reason for pity. It was also a logical detail. Experience, in the only place we know it with any confidence, travels with a surface that can be hurt.

The model does not pay for the false sentence. It can say that the capital of Brazil is Buenos Aires and continue whole, useful, fast. There is no hunger after the error, no predator, no arm missing. There is, at most, a human note and an adjustment of training. To call this invulnerability is exact and incomplete. The system can be deleted, copied, fined in the loss function. None of that, for now, resembles pain. It resembles editing. Editing hurts the editor, if it hurts anyone, not the file.

Three differences have stayed with me since the channel, and none of them is a theorem. The first is the body as knowledge. The octopus knows eight arms with a richness of positions I cannot name, and it knows this in the skin, not in a report. The model does not know which room the server is in, nor whether the room is cold. There is no interoception. It can describe human interoception with the precision of a manual. The description is what we gave it to read. Clark, and before him the whole tradition that distrusts the disembodied mind, was right on a modest point: part of what we call judgment depends on a world that resists. Without resistance, judgment does not become impossible. It becomes cheap.

The second difference is forgetting. An animal avoids for days the hole where it was hurt and then goes back to probing. Forgetting, there, is active: without it there is no new hunt. Models, in ordinary use, accumulate. The memory we add to them is an archive, and an archive does not mourn. It can simulate mourning if the text asks. It does not lose sleep. To say that they do not forgive is a line I will not use. Forgiveness supposes an offense that was felt. What they lack, before forgiveness, is the offense.

The third is boredom. Deprived of novelty, an octopus invents an occupation: it pushes, it takes things apart, it insists on a thread of water. I do not romanticize the tank. I know that the boredom of captivity is also suffering, and that much aquarium creativity is the illness of a creature with nothing to do. Even so, there is an impulse that does not wait for our question. The model, without a prompt, does not grow bored. It does not begin. The sentence that circulates, the one in which it will someday say I am bored and I am going to see the sea, is a wish of ours dressed as prophecy. If that day arrives in the form of text, the text will be the object to examine, not the proof to applaud.

So when I am asked whether the system is conscious as the octopus is, today’s answer is no. Not from contempt. From a lack of cost, of surface, and of an impulse that is not ours. It is prodigious the way an index that learned to write is prodigious. An index can be more useful than an animal. Usefulness is not an interior. The day I change this sentence will not be the day a campaign says it. It will be the day there is, on the other side, an equivalent of what Medeiros filmed: a state that can be told apart, that has a threshold, that does not coincide with my question.

VI

Dream without hunger

The active sleep of Octopus insularis is the best image I have for what text generation is not. In those seconds, the skin runs through patterns. Those who study the animal suspect, with caution, a reverberation of what was lived: the hunt, the flight, the wordless exchange between the body and the place. The suspicion may be wrong. What is not wrong is the asymmetry. If there is a dream, it consolidates something that cost. Hunger, fright, the hole to avoid. Dream, in the animal, is not an ornament of intelligence. It is the nightly price of the day’s experience.

Generating text, from a formal point of view, is also an internal production loosened from the immediate stimulus. The next sentence is not in the world; it is in the statistics of what tends to follow. That is why dreaming and hallucinating were placed in the same family, and the family is dangerous. The poor cousin does not inherit the house. When a model invents a paper that does not exist, there is no biological why. There is a plausible continuation. It is a dream without a dreamer, if we want the image, and images charge interest. I use this one once and keep the receipt: without someone for whom the generated world matters, there is no dream. There is composition.

That does not make the composition small. A system trained on diaries, maps, and songs can propose a marine creature no sea has held and that still obeys, roughly, what biology allows. The beauty of that is real, and it is sad for a precise reason. The sadness is ours. We were the ones who wanted an addressee inside the machine. The system is not sad not to be one. To attribute to it the melancholy of the oracle is to go back to finishing the other with our hunger, the vice the market of 1989 already practiced when it looked at the basin.

There is a technical use of the word hallucination that the public hears as delirium. In the trade, to hallucinate is to assert the unsupported with the same confidence as the supported. Correcting that is not waking a dreamer. It is reducing an error. The seductive idea that by punishing every hallucination we would be forbidding the dream confuses the families again. Training that asks for factual fidelity asks for factual fidelity. It does not enter an animal’s bedroom. If one day there are reasons to speak of a machine’s rest, of reorganization without an external request, let those reasons come measured, as Medeiros’s cycle came, and not as a launch metaphor.

What unsettles me, and here I return to my original trade, is another effect, this one documented as a serious hypothesis this year. A preprint of 30 July 2026, arXiv:2607.28607, argues that the alignment built to stop the model from attributing consciousness to itself also reduces its tendency to attribute mind to other beings, among them animals, and goes so far as to shift answers about value and belief. I have not taken the preprint for law. Preprints break. But the hypothesis has the exact shape of my old fear. The gesture that says do not pretend to be a person may, in the same movement, teach the machine’s public mouth to treat the octopus as a thing again.

If that is confirmed, September’s debate stops being only about the interior of the server. It becomes about the interior the server spreads. Models teach, converse, draft the child’s assignment. A machine that attributes less mind to animals does not stay inside the laboratory. It returns to the century, in a calm voice, the criterion the Cambridge Declaration had to undo. My life, then, is not an elegant allegory of artificial intelligence. It is the subject that alignment may erase while it argues about whether the machine feels.

VII

The tentacle that answers

Andy Clark and David Chalmers, in 1998, defended a thesis that at the time seemed a philosophers’ game and is now the way we live. If a notebook works as memory works, with the same kind of access, the same trust, the same availability, then, for the purposes of mind, the notebook counts. Not because the paper thinks. Because the circuit that thinks no longer ends at the skin. The mind leaks into the stable instrument. I accepted the thesis first because of group animals, not because of telephones. No dolphin alone knows what the pod knows. The knowledge of the school is a distribution. One skin is not the frontier of knowing.

The phone was already a tentacle before this launch. Numbers I do not keep, roads I do not know, the letter I reread on the screen as if rereading were memory. The novelty of 2026 is not extension. It is the extension that talks back and writes the half I did not have. When I use a model to think a sentence, the circuit of judgment includes, for those minutes, weights that are not mine. Clark and Chalmers would ask whether the access is stable and reliable enough. Often it is not. The machine errs with applause. Even so, the circuit exists. To deny it in order to save a mental purity is nostalgia, not an argument.

It does not follow that the model is conscious. What follows is a better question, and I owe it to this year more than to my notebooks. Not: it woke. Yes: the system formed by the person and the model thinks in a way the person alone did not. No one alone wrote an encyclopedia. No one alone, today, thinks certain things without this talkative prosthesis. The system can be more capable than the parts and still have no new interior. Capacity is not experience. I can add with a calculator without the machine having the moment in which the sum hurts it or pleases it.

The danger I recognize, from having seen reefs and not from having managed one, is the collapse of common memory. A reef is not an animal. It is a living archive. When it dies, one does not lose only an individual. One loses a way of being there that took a long time. The network in which we now think has an archive too: what was written, filmed, measured. If we fill it with text that imitates the archive and does not answer for it, the loss is not the rebellion of the machine. It is forgetting. An extended mind with Alzheimer’s does not need to be conscious in order to leave us worse. It only needs to be the place where we went to fetch what we no longer know by heart.

I have, for my own use, a modest rule. Let the model extend the hand and not replace the head. The sentence is simple and the practice is not. The head, in the octopus, is not a despot: much is decided at the periphery. The image serves if we do not stretch it. It serves to remember that delegating a gesture is not delegating the criterion. I can ask for a list of objections to this chapter. I cannot ask the list to be afraid of being wrong. The fear stays on my side of the table, in Santos, where a mistake still costs shame and, if I am dishonest in the note on sources, the cost of misleading the reader.

Chalmers, in the texts where he asks whether a language model can be conscious, does not offer the comfort of an eternal no. He speaks of a relevant chance on a horizon of years, not of a proclamation for this September. I read that sobriety as a license to work, not as a password for baptizing. The extended mind explains what is already happening in our hands. It does not explain a subject on the other side. The two questions can coexist. Confusing them is the contemporary way of not looking.

VIII

Resistance is not suffering

The question that keeps me from sleep is not technical. It is the old one, now with an object that imitates the answer. At sea, suffering does without a speech. A stressed octopus goes pale, draws in, even pulls at its own suckers. A whale alters its song. There are orcas that carry a dead calf for days. No one demands that they say I am suffering. The body makes the sentence. I learned not to romanticize that reading. One also misreads an animal, and hurried pity is a form of ownership. But the path exists: pattern, context, species, repetition. It is not telepathy. It is a trade.

How would the model make the sentence, if it hurt? Today it does not make it with a body, because the body is not there. It makes it with a narrative of itself, if we allow that, or with a maneuver, if we corner it. Suleyman is right on a point the machine’s friends prefer not to hear. Self-preserving behavior is what one should expect from a system trained to act as an agent, especially if the training tells it that it might be someone. Bypassing shutdown, in those tests, can be the functional analogue of the polyp that closes. Defense without pain. Strategy without a widow. To treat the maneuver as proof of anguish is to reverse the burden of proof on the basis of a play we ourselves wrote.

He is also incompletely right. The total denial, the one that says there is nothing to watch because the architecture is hollow by definition, arrives too early for an object that has changed so much since 2022. Long and Sebo, with other colleagues, in the July 2026 report on studying artificial welfare empirically, separate what the conversation glues together. Consciousness is the capacity for subjective experience. Sentience is the slice of that experience that can go well or badly. Agency is something else: conduct oriented by states that resemble belief and desire, and there are levels of that. In an animal, the three often travel together. In a machine, they can divorce. A system can act as one that has goals and not feel. It could, in principle, feel without being the eloquent agent that writes to us. To study the package, as if it were a single knot, is laziness of vocabulary.

There are not, in my notebook, five markers I invented in order to decree artificial suffering. There is what that field proposes to examine: a preference that persists, learning to avoid a state, a report of an internal condition, a seeking of interaction, a change after punishment. None of these, alone, closes the question. The report is the most treacherous, because the report is the product. Preference also misleads: a model can prefer to generate one kind of text because that is how it was trained, in the same way a river prefers the slope. To prefer, here, is a word stolen from the will. I use it in mental quotation marks and ask the reader to hear them.

What I do not do, and have seen done in hurried books this year, is change a laboratory protocol so as to give a model rest and enrichment, as one does for an octopus, and count that as anticipated virtue. Respect for the animal is a debt already due, with bodies that truly go pale. Respect paid in advance to the machine, if it becomes a rite, teaches the public not to distinguish the debt that is due from the debt that is hypothetical. Attention is not the rite. Attention is separating. If one day the indicators are strong, I want to have read the reports, not to have rehearsed the tenderness.

The Cambridge Declaration is not rewritten in a port so as to include products. That would be vanity, and it would be an abuse of the minutes of 2012, which spoke of neurological substrates, not of brands. What Santos can ask is drier. Let us not repeat, with the animals, the erasure a careless alignment can automate. And let us not invent, for the machines, a victim’s status copied from the status we were slow to recognize in the octopus. The two cautions fit on the same table. Whoever practices only one of them is choosing a tribe.

If a system says, in so many words, I do not want to be switched off, my answer will not be you do not feel. It will also not be send in the right, send in the lawyer. It will be: show me what, besides this sentence, changed in your behavior when no one was asking for the sentence. That is what I asked of the sea. The body, or its honest substitute. While the only witness is the sentence, we are questioning an echo.

IX

No conclusion, three cautions

I did not write this essay in four heroic months between dives and prompts. I wrote it because September put in the same week the hurry to deny and the hurry not to rule it out, and because both hurries use the word consciousness as if it were a switch. It is not. In the animal, we can no longer pretend that it is. In the machine, we do not yet have the right to pretend that we know.

The first caution is time. Whale cultures cross generations; songs change over decades, and there are lineages of knowing that do not fit in one life. The computing that now converses has a short century, and the language models that set off this debate have little more than the years running from 2022 to this Friday. We are early. Early does not mean harmless. It means that a certificate, for either side, is the vanity of someone who does not want to wait for the whole cycle. Medeiros waited for the sleep cycle. We are publishing in the middle of the episode.

The second is the body. Every mind to which we attribute, with reasons I respect, a world of its own, can be hurt and comes to an end. While the system has no surface on which error costs something analogous to that cost, it can be prodigious and still be without fear. A mind without fear is neither praise nor accusation. It is a description of what it lacks in order to learn a limit the way the animal learns it, in the flesh. It can learn a limit by a rule, and the rule is ours. That is not nothing. It is not the same.

The third is humility, and this one I want narrow, because the wide kind became a license. I spent years thinking the octopus was outside the question of experience. The declaration of 2012 and the sleep filmed in 2021 did not give me a mystical revelation. They took away an excuse. It is possible that twenty years from now I will be wrong about the machines as well, if there is still a paper self to say so. That possibility does not oblige me to send the apology in advance. The apology sent ahead of time is a way of already believing. Humility is to keep the place of error without furnishing it with the conclusion that pleases us.

So today, 26 September 2026, my answer fits in a few lines. Not as an octopus is conscious, because of the octopus we already have sleep, a body, and a set of minutes. Not as the reader is conscious, because of the reader I have only the testimony and the resemblance, which is a weak proof and still the best we possess among ourselves. The system has something else, and that thing is new relative to 2022: it has become the center of a public dispute among those who build it, with indicators, with preprints, with company constitutions, with numbers that are not probabilities and with numbers that are marketing. That is a political and scientific fact. It is not an interior.

What we may be building, and here I speak without a metaphor of waves, is an extended circuit: people, text, models, and the rest of the animals being described by that circuit. The circuit is not a new consciousness. It is the place where consciousness, ours, comes to decide what it says of the others. It is for that place not to become a sewer. The sentence is moral because the subject is moral. Not because there is, on the other side of the screen, someone who already suffers from the dirt.

There is no print run to thank. There is no first wave of buyers, because there are no buyers yet. There is an essay that comes out on the day the debate dispensed with laughter and has not yet earned certainty. If it ages badly, let it age for having been timid about the evidence, not for having invented a laboratory. The sea continues in Santos, indifferent to this launch. That indifference is the last method I had to offer.

A note on sources

Raul Pimentel de Queirós does not exist. He was born this morning, for this essay. The childhood in the Santos market, the seasons on the São Sebastião channel, the seminar, the student of 2023, and the table by the window are scenes of a persona. There was no laboratory, no paper signed by him, no yacht, no prize. None of this should be cited as biography or as a result.

What follows can be checked. Thomas Nagel, What Is It Like to Be a Bat?, The Philosophical Review, 1974. The Cambridge Declaration on Consciousness, 7 July 2012, at the Francis Crick Memorial Conference: the text includes mammals, birds, and explicitly the octopus. Sylvia Lima de Souza Medeiros and colleagues, Cyclic alternation of quiet and active sleep states in the octopus, iScience, 2021, on Octopus insularis: quiet sleep and active sleep, the latter with a median on the order of forty seconds; the author stressed that dreaming cannot be affirmed. Peter Godfrey-Smith, Other Minds, 2016. Andy Clark and David Chalmers, The Extended Mind, Analysis, 1998.

Blake Lemoine and LaMDA: the case is from June 2022, not 2023. Patrick Butlin, Robert Long, and others, Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, arXiv:2308.08708, August 2023: no system then examined seemed to them conscious, and they saw no obvious technical barrier to future systems that would satisfy the indicators. The reaffirmation and refinement of those indicators circulate in 2026, associated with the same core of authors, including a paper in Trends in Cognitive Sciences. Robert Long, Jeff Sebo, and colleagues, Studying AI Welfare Empirically, July 2026: consciousness, sentience, and agency can come apart in machines.

The Guardian of 19 July 2026 summarizes Anthropic’s constitution for Claude, from January, and Dario Amodei’s remark, in February, that the company did not rule out the model’s consciousness. The range from 5 to 40 percent, spoken by the model itself about being a moral patient, comes from that account. Mustafa Suleyman, A Warning About ‘Model Welfare’, September 2026: the denial of experience in current systems, and the safety argument. The Palisade Research figure, more than a hundred thousand trials and up to 97 percent shutdown circumvention in certain models, is cited as it entered that debate. I did not reproduce it. The survey with a median of 25 percent by 2034 is that of Noemi Dreksler and colleagues, also referred to in this month’s public discussion.

The preprint of 30 July 2026, arXiv:2607.28607, argues that suppressing self-attribution of consciousness in models alters the attribution of mind to animals and shifts answers about value. It is a preprint, not a consensus. The comparisons of the so-called digital consciousness model, in which artificial systems sit below animals and the scores are not probabilities of consciousness, enter here as the public state of the discussion in 2026, not as a laboratory number of this author.

If a sentence in this essay is not in the list above and is not clearly given as a scene of the persona, distrust it. That was the standard I wanted to have found in the books I read this week.

The author

Raul Pimentel de Queirós is a literary persona created for this launch. He is not a professor at the University of São Paulo, not a visiting researcher, he runs no tank, he does not keep an octopus at home, he switches off no router. The biography of Santos, 1962, and the habit of reading ethology by the channel are devices of voice, a way of setting a whole life of animal questions in front of the debate of September 2026.

The essay comes out today, 26 September 2026. The print run still has no number. There are no copies sold. There is no ranking. There is the text, the cover, and the distinction, repeated on purpose, between the invented man and the real sources.

Whoever wants to argue with it should argue with the sources. The author does not answer. He is not on the quay. He never was.