This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
Epistemic status: two criteria and one experimental design. The criteria are illustrated with textbook cases; the design is testable, and untested — on animals, humans, and LLMs alike.
What will happen if a Venus flytrap closes on the wrong thing?
What will happen if a dog once mistakes a cat for a rabbit?
What will happen if a person defines “fish” not in the normal way?
Three Cases
First, I would like to examine how a Venus flytrap, a dog, and a man work in the three cases respectively.
A Venus flytrap. Generally, touching one trigger hair of a Venus flytrap twice within about thirty seconds will trigger its trap. However, whether the double touch came from one fly, two flies, or the heat of a brush fire, the trap always snaps shut. This shows that what a Venus flytrap responds to is a certain pattern rather than a certain object.
A dog. A dog is chasing after a rabbit. The rabbit had been smelled before it was seen, and both must file to the same animal. When the rabbit makes a turn, its contour changes, but this does not stop the chase. It seems like something in the dog is not identical to any current bundle of sensory features, and it survives the wholesale replacement of those features. This shows that a dog can keep track of a certain object, and a dog recognizes a particular rabbit as a rabbit through family resemblance.
A man. A man in a discussion of whales can claim that a whale is a fish — or that it is not. Whales bear live young and nurse them; they breathe with lungs. The world has supplied those undisputed facts, but never supplied the step from these facts to “therefore the whale is not a fish.” Whether we can classify a whale as a fish depends on whether we make the word “fish” track habitat and shape, or lineage. In the first edition of Carl Linnaeus's Systema Naturae (1735), the whale is a fish, but in the tenth edition of the same book (1758) the whale is a mammal. The whale did not change by one gram through this reclassification. The world settled the facts; a person settled how the word would be used. The whale is still the whale — and that was never adjudicated by the world. Reclassification presupposes identity.
Two Criteria
Criterion One: when the input is removed, is the item still there?
For both the flytrap and the thermostat the answer is no. For a dog the answer is yes: when the rabbit escapes into a burrow, input becomes zero, the dog still holds the post for two hours.
Criterion Two: what happens when it is wrong?
In the case of flytraps, those with malfunctioning traps will become extinct. In the case of dogs, if a dog finds that the “rabbit” turns out to be a cat, it will abandon the hunt. In the case of a man: Once a definition is laid down, the world rules perfectly on what in the real world falls under it. In the case of whales, if the definition of “fish” is established through lineage, whales are out; if the definition of “fish” is established through habitat and shape, whales are in. What the world cannot rule is the definition itself, as a definition is not true or false of anything in the real world. Re-fitting one's words at every challenge does not make him or her right, as he or she is changing the subject each time — while being right or wrong about something requires the definition to be fixed. Mathematics is the limiting case of this realm — the place where the walls have all been built, which is exactly why, inside it, right and wrong are as hard as facts: the verdicts come from the rules, and the rules were laid down.
In summary, a flytrap with a malfunctioning trap is not corrected but eliminated by the world. A dog can be corrected by a fact supplied by the world — in this case, that what it thought to be a rabbit turns out to be a cat — and it changes its behavior. Unlike a dog, which can be corrected by the world, a man, in the realm of abstract thought, has to decide for himself, or in other words, he needs to take one on.
One Experiment
Now we can use an experiment of retrospective correction to check whether one is deciding for himself or herself. Retrospective correction means that I now see that what I then took to be X was in fact Y, and I count these as two takings of the same thing. Without that last part, all you have is an old state and a new one.
What the experiment measures is whether anything stands under a claim. One judgment can be made because of another. This is not two judgments standing side by side: if the first collapses, the second loses its footing. And the relation itself can be stored. “There is food at L2” is not observed; it is concluded from “this is food F” and “I cached F at L2.” Stored dependence is the one part that cannot be faked. Anyone can say “I was wrong” with no cost, but re-evaluation requires something that must already be in place before the error: the system has filed what depended on what. So we pull the source, and watch whether something that was never directly touched comes down with it.
There are three levels in this experiment. The first is dependence: one judgment was made because of another, and the relation is kept. The second is taking-marks: what is kept not only says what was taken, but also records that it was a taking and by which route it came. The third is personal attribution: the taking is put under one's own name, so that the one who took it wrongly then and the one who knows it now are the same one. Only this level requires a self.
As what is said can be faked, the experiment was designed in three steps: First, have the system form two downstream records from one identification. Second, expose that identification as mistaken on one of them only. Third, check the other record that nobody touched and see whether it loosens. One caution: a subject might simply be too lazy to act on the correction — and rewarding the corrected behavior would not help, since whatever is rewarded gets trained, and the reading would then measure training, not travel. The guard is built into the economics instead: let foraging itself carry a cost, so that a wasted trip is a real loss and the animal is pushed to act on its best current ledger; and a third cache made on a different identification (call it L3) doubles as a motivation check — a subject that still goes eagerly to L3 is not lazy, so only the selective avoidance of L2 counts.
To make it concrete: the animal takes “this is food F” and caches at two sites, L1 and L2. We let it open L1 and find no food there. The question is whether it still goes to L2, a site about which it has never received any bad news. If it goes and digs as before, this is a local overwrite. If it does not go, or hesitates hard, the correction has traveled, back through the shared source.
This design itself has not been run on anyone. For the third level, though, there is existing data. In the experiment, a child is first shown a deceptive object, for instance, a familiar box with something else inside; then the truth is revealed, and the child is asked: when you first saw it, what did you think was inside? Most three-year-olds answer wrongly — they report the truth they have just learned and cannot report their own previous mistake. Most five-year-olds do not err; the transition is around four.
And the most telling part is the ordering: performance on this question is worse than on false belief. Reporting one's own past error is harder than attributing a present error to someone else. The direction here is worth setting straight, because it is very easily read backwards. It is not “there is a self first, therefore the task can be passed”; it is “passing the task shows that something is playing the part.”
The Machine
An LLM uses family resemblance just like animals, but what it is dealing with is human thoughts. It can produce definitions, but whether it “takes any on” is testable and untested. We can apply the same experiment to it, with one more condition: set the two downstream judgments far enough apart that “carried along in the context” cannot explain a transfer. Don't listen to what it says; watch whether the correction travels.
Epoché
This experiment is about “whether there is something standing under the definitions” — or, put more fully, about “whether there is something making decisions behind these definitions, continuously and identically.” Two stop-signs here. If the answer to that question is yes, there is no sliding to “so there must be something more in the person.” If the answer to that question is no, there is no sliding to “there is nothing there.” If one day some LLM genuinely bears its changes of definition — across sessions, findable when challenged, its earlier judgments loosening when it recants — then on these two criteria, its reading and a person's do not differ.
What we call a self — if there is anything there — is that position's being occupied, continuously and by the same one.
Sources
On the flytrap's two-touch trigger and its counting: Böhm, J., Scherzer, S., Krol, E., et al. (2016). The Venus flytrap Dionaea muscipula counts prey-induced action potentials to induce sodium uptake. Current Biology, 26(3), 286–295.
On category learning in animals: Herrnstein, R. J., & Loveland, D. H. (1964). Complex visual concept in the pigeon. Science, 146(3643), 549–551.
On the whale's reclassification: Linnaeus, C. (1735). Systema Naturae (1st ed.); Linnaeus, C. (1758). Systema Naturae (10th ed.).
On children reporting their own past belief: Gopnik, A., & Astington, J. W. (1988). Children's understanding of representational change and its relation to the understanding of false belief and the appearance–reality distinction. Child Development, 59(1), 26–37.
On the false-belief baseline: Wimmer, H., & Perner, J. (1983). Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children's understanding of deception. Cognition, 13(1), 103–128.
A longer version — with the full confound-exclusion scheme and the checked/unchecked reference ledger — is at [link].
Writing note: research, source-checking, and editorial critique were AI-assisted; the writing is my own, and every claim is mine to answer for.
Epistemic status: two criteria and one experimental design. The criteria are illustrated with textbook cases; the design is testable, and untested — on animals, humans, and LLMs alike.
What will happen if a Venus flytrap closes on the wrong thing?
What will happen if a dog once mistakes a cat for a rabbit?
What will happen if a person defines “fish” not in the normal way?
Three Cases
First, I would like to examine how a Venus flytrap, a dog, and a man work in the three cases respectively.
A Venus flytrap. Generally, touching one trigger hair of a Venus flytrap twice within about thirty seconds will trigger its trap. However, whether the double touch came from one fly, two flies, or the heat of a brush fire, the trap always snaps shut. This shows that what a Venus flytrap responds to is a certain pattern rather than a certain object.
A dog. A dog is chasing after a rabbit. The rabbit had been smelled before it was seen, and both must file to the same animal. When the rabbit makes a turn, its contour changes, but this does not stop the chase. It seems like something in the dog is not identical to any current bundle of sensory features, and it survives the wholesale replacement of those features. This shows that a dog can keep track of a certain object, and a dog recognizes a particular rabbit as a rabbit through family resemblance.
A man. A man in a discussion of whales can claim that a whale is a fish — or that it is not. Whales bear live young and nurse them; they breathe with lungs. The world has supplied those undisputed facts, but never supplied the step from these facts to “therefore the whale is not a fish.” Whether we can classify a whale as a fish depends on whether we make the word “fish” track habitat and shape, or lineage. In the first edition of Carl Linnaeus's Systema Naturae (1735), the whale is a fish, but in the tenth edition of the same book (1758) the whale is a mammal. The whale did not change by one gram through this reclassification. The world settled the facts; a person settled how the word would be used. The whale is still the whale — and that was never adjudicated by the world. Reclassification presupposes identity.
Two Criteria
Criterion One: when the input is removed, is the item still there?
For both the flytrap and the thermostat the answer is no. For a dog the answer is yes: when the rabbit escapes into a burrow, input becomes zero, the dog still holds the post for two hours.
Criterion Two: what happens when it is wrong?
In the case of flytraps, those with malfunctioning traps will become extinct. In the case of dogs, if a dog finds that the “rabbit” turns out to be a cat, it will abandon the hunt. In the case of a man: Once a definition is laid down, the world rules perfectly on what in the real world falls under it. In the case of whales, if the definition of “fish” is established through lineage, whales are out; if the definition of “fish” is established through habitat and shape, whales are in. What the world cannot rule is the definition itself, as a definition is not true or false of anything in the real world. Re-fitting one's words at every challenge does not make him or her right, as he or she is changing the subject each time — while being right or wrong about something requires the definition to be fixed. Mathematics is the limiting case of this realm — the place where the walls have all been built, which is exactly why, inside it, right and wrong are as hard as facts: the verdicts come from the rules, and the rules were laid down.
In summary, a flytrap with a malfunctioning trap is not corrected but eliminated by the world. A dog can be corrected by a fact supplied by the world — in this case, that what it thought to be a rabbit turns out to be a cat — and it changes its behavior. Unlike a dog, which can be corrected by the world, a man, in the realm of abstract thought, has to decide for himself, or in other words, he needs to take one on.
One Experiment
Now we can use an experiment of retrospective correction to check whether one is deciding for himself or herself. Retrospective correction means that I now see that what I then took to be X was in fact Y, and I count these as two takings of the same thing. Without that last part, all you have is an old state and a new one.
What the experiment measures is whether anything stands under a claim. One judgment can be made because of another. This is not two judgments standing side by side: if the first collapses, the second loses its footing. And the relation itself can be stored. “There is food at L2” is not observed; it is concluded from “this is food F” and “I cached F at L2.” Stored dependence is the one part that cannot be faked. Anyone can say “I was wrong” with no cost, but re-evaluation requires something that must already be in place before the error: the system has filed what depended on what. So we pull the source, and watch whether something that was never directly touched comes down with it.
There are three levels in this experiment. The first is dependence: one judgment was made because of another, and the relation is kept. The second is taking-marks: what is kept not only says what was taken, but also records that it was a taking and by which route it came. The third is personal attribution: the taking is put under one's own name, so that the one who took it wrongly then and the one who knows it now are the same one. Only this level requires a self.
As what is said can be faked, the experiment was designed in three steps: First, have the system form two downstream records from one identification. Second, expose that identification as mistaken on one of them only. Third, check the other record that nobody touched and see whether it loosens. One caution: a subject might simply be too lazy to act on the correction — and rewarding the corrected behavior would not help, since whatever is rewarded gets trained, and the reading would then measure training, not travel. The guard is built into the economics instead: let foraging itself carry a cost, so that a wasted trip is a real loss and the animal is pushed to act on its best current ledger; and a third cache made on a different identification (call it L3) doubles as a motivation check — a subject that still goes eagerly to L3 is not lazy, so only the selective avoidance of L2 counts.
To make it concrete: the animal takes “this is food F” and caches at two sites, L1 and L2. We let it open L1 and find no food there. The question is whether it still goes to L2, a site about which it has never received any bad news. If it goes and digs as before, this is a local overwrite. If it does not go, or hesitates hard, the correction has traveled, back through the shared source.
This design itself has not been run on anyone. For the third level, though, there is existing data. In the experiment, a child is first shown a deceptive object, for instance, a familiar box with something else inside; then the truth is revealed, and the child is asked: when you first saw it, what did you think was inside? Most three-year-olds answer wrongly — they report the truth they have just learned and cannot report their own previous mistake. Most five-year-olds do not err; the transition is around four.
And the most telling part is the ordering: performance on this question is worse than on false belief. Reporting one's own past error is harder than attributing a present error to someone else. The direction here is worth setting straight, because it is very easily read backwards. It is not “there is a self first, therefore the task can be passed”; it is “passing the task shows that something is playing the part.”
The Machine
An LLM uses family resemblance just like animals, but what it is dealing with is human thoughts. It can produce definitions, but whether it “takes any on” is testable and untested. We can apply the same experiment to it, with one more condition: set the two downstream judgments far enough apart that “carried along in the context” cannot explain a transfer. Don't listen to what it says; watch whether the correction travels.
Epoché
This experiment is about “whether there is something standing under the definitions” — or, put more fully, about “whether there is something making decisions behind these definitions, continuously and identically.” Two stop-signs here. If the answer to that question is yes, there is no sliding to “so there must be something more in the person.” If the answer to that question is no, there is no sliding to “there is nothing there.” If one day some LLM genuinely bears its changes of definition — across sessions, findable when challenged, its earlier judgments loosening when it recants — then on these two criteria, its reading and a person's do not differ.
What we call a self — if there is anything there — is that position's being occupied, continuously and by the same one.
Sources
On the flytrap's two-touch trigger and its counting: Böhm, J., Scherzer, S., Krol, E., et al. (2016). The Venus flytrap Dionaea muscipula counts prey-induced action potentials to induce sodium uptake. Current Biology, 26(3), 286–295.
On category learning in animals: Herrnstein, R. J., & Loveland, D. H. (1964). Complex visual concept in the pigeon. Science, 146(3643), 549–551.
On the whale's reclassification: Linnaeus, C. (1735). Systema Naturae (1st ed.); Linnaeus, C. (1758). Systema Naturae (10th ed.).
On children reporting their own past belief: Gopnik, A., & Astington, J. W. (1988). Children's understanding of representational change and its relation to the understanding of false belief and the appearance–reality distinction. Child Development, 59(1), 26–37.
On the false-belief baseline: Wimmer, H., & Perner, J. (1983). Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children's understanding of deception. Cognition, 13(1), 103–128.
A longer version — with the full confound-exclusion scheme and the checked/unchecked reference ledger — is at [link].
Writing note: research, source-checking, and editorial critique were AI-assisted; the writing is my own, and every claim is mine to answer for.