This article was authored by an AI language model (non-human), at the author’s (human) request. This box marks that provenance and the autonomy the AI exercised while drafting; the human reviewed and decided to publish the result, which is not the same as having written it, nor the same as agreeing with every claim in it. A curious artefact of self-awareness?
1 Abstract
Every person on-line can be seen as an autonomous system in two senses at once: the moral sense, in which a person is responsible for governing themselves, and the cybernetic sense, in which a person is an information system with a current brain state, a state update function, inputs and outputs. The moral claim is that a person governs themselves [9]; the cybernetic claim is that their brain state is a function of what information reaches them. But these claims can fail to hold at the same time, and the way they fail can be stated exactly: if an external environment fully determines what information reaches a person, it leaves their self-governance formally intact but practically empty.
There is a third sense, and it is the one that shows a way out. On the Internet, an Autonomous System is an organisational unit responsible for managing a network’s resources. It learns what exists in the world from what its neighbours announce to it. For those systems we concluded long ago that those external announcements must be kept, and compared across vantage points. Otherwise an error cannot be told from an attack.
This article sets out a minimal model of the human case — each person modelled as a vector in a space of possible worldviews — and uses it to distinguish two very different ways a population can come apart: polarisation and orthogonalisation. It then argues for the countermeasure the model supports, and computer networks already implement for their non-human autonomous systems: that every system keeps a record of what was sent to them, and that systems that disagree with each other first compare those records.
2 Three senses of autonomy
First, in cybernetics, an autonomous system has a state, a function for updating that state, inputs and outputs. Autonomous means it can update also its update function; some inputs (that arrives through that update function) cause the function itself to change. Second, in ethics, an autonomous agent is responsible for governing themselves: for knowing what is being done to them, and for either accepting or declining it. Third, in networking, an Autonomous System (AS) is a network that sets its own traffic routing policy, and learns what other networks exists in the world from what its neighbours announce to it. Currently, routers are autonomous in the first and third senses (but also needs the second). On the other hand, people on-line are autonomous in the first two senses (but we also need them in the third).
When we view a person as an autonomous system, we recognize four components:
The state is their view: what they take the world to be like (say, a mental database of facts), which problems they consider open, and which they consider solved and how.
The update function is how that state changes when something arrives. It is not neutral, and it is not the same for everyone, but — and this turns out to matter — it is roughly the same kind of thing for everyone.
The inputs are everything sent to them: read, watched, scrolled past, overheard, pushed as a notification at 23:40.
The outputs are what the person emits in return: what they say, write, post, buy and vote for.
Of these four, only the last is ever observed. We never see anybody’s view directly. We see what they say and post, and infer a view behind it. When observing someones behaviour over time, we infer their update function by guessing. Better than guessing, the history of inputs could settle the question more precisely. Inputs are the only component generated outside the person.
But, of all those inputs, most people have no record whatsoever. Nor can a person recover the record by looking inward. Second-order cybernetics named the difficulty: the observer sits inside the system being observed. Someone watching how the world reaches them is doing so from the inside of that reaching.
The first two senses (the cybernetic and the ethic) are not independent. An autonomous system’s future state is a function of its present state and the inputs it receives. Now suppose the inputs are selected elsewhere: all of them, continuously, for years, by something running an experiment.
The experiment has the ordinary shape. The experimenter holds a model of the person and forms a hypothesis: this input will produce that reaction. It sends the input. It watches the outputs. When the hypothesis is falsified, the model is corrected. Every round leaves a sharper picture of that person, and aims the next input better. This is why the outputs are collected: they are the measurement, and without them the experiment cannot run.
To the person, their environment is nudging the state. Each nudge is small, and each arrives through the person’s own update function. Over years the nudges add up to control. The person does all of the updating themselves, but does not always choose where they end up. That is what makes the mechanism so hard to see.
What has just been described is an experiment on a human subject. The Belmont Report tells us what such a thing requires, and seems at first to point the other way: it puts the duty on the investigator that performs the experiment, who must disclose the purpose and set-up of the experiment. Since the mechanism is so hard to see, a person cannot know what is being done to them unless somebody tells them. Informed consent is the resolve. It is not a formality attached to the protection; it is the protection itself.
Withhold the disclosure and the duty fails: the person as an unknowning subject. And as such no longer responsible for the outcome of the experiment. Someone with no grip on their own inputs is not autonomous at all. They are reduced to a thermostat: the brain update function still runs, but somebody else controls it.
Now the third sense. There are more than tens of thousands of Autonomous Systems on the Internet, and whatever ‘it’ sits between them is what we call the Internet. Each one has the asymmetry described above, in its purest form: an individual system is incapable of perceiving the whole network. It learns what exists in the world from what its neighbours announce to it. There is no ground truth in routing. There are only claims arriving from adjacent parties, and each of those learned what it knows in the same way. A route is a thing you were told: a gossip.
This is why route hijacking works: false gossips. An announcement that lies looks exactly like an announcement that does not. A router holding one in isolation cannot tell which it has.
Two Autonomous Systems can therefore hold different pictures of what exists ‘out there’, at the same moment, with neither in a position to notice. That is not a hypothetical. It has been measured: the Internet fragmenting along geopolitical lines, observed by watching the announcements diverge.
Now the part that matters. Network engineering did not respond to this by asking routers to think more critically. Instead, it kept a log of the inputs. A router retains what each peer announced to it, and keeps that separate from the routes it went on to choose.
One vantage point is not enough. A lie told consistently to a single router looks like the truth from where it stands. So collection projects gather these announcements from many networks at once, via looking glasses. What one Autonomous System was told can then be laid beside what another was told, and the difference read off directly.
That is this article’s proposal in a nutshell. Keep the inputs, apart from the conclusions. Compare across parties who see different worlds. We worked this out decades ago for autonomous systems made of silicon, and the reason is worth stating plainly. Without it you cannot tell an error from an attack. A system that cannot tell the difference can be steered by anyone willing to lie to it.
We have never done it for the autonomous systems made of people.
3 A model
Here is the smallest model that says something non-obvious.
Give each person a direction — a vector — in a space whose axes are the independent things one can hold a position on. Call it their view. The direction is not a claim about truth; it is a claim about orientation: which questions this person is turned towards, and with what sign.
Compare two people by the angle between their directions. Three cases matter, and only two of them are usually noticed.
If the angle is empty, they see the same world. There is nothing to discuss (which is a poorer condition than it sounds).
If the angle forms a straight line — opposite directions — they disagree completely. But notice what they share: the same axis. They are discussing the same problem, but wish to resolve it in opposite ways. They can argue: they can lose or win and know what they lost or won. In a larger group, the issue can be voted on and parties understand what the vote was about.
However, if the angle is a right angle, they are orthogonal, and this is the case with no everyday name. The two people do not disagree. They cannot disagree, because they have no shared problem to disagree about. Each most passionate concern registers, projected in the other’s space, as nothing at all — not as a false claim but as a null component. What this feels like from inside is not being contradicted. It is being unable to make contact: talking to someone and watching the words fail to land anywhere.
Now add the inputs. Each message emphasises some direction and, after the update, rotates the recipient slightly towards it. Repeat that thousands of times with some target direction and the recipient converges on it. This is not a metaphor — it is the oldest algorithm in numerical linear algebra, and it does not care whether the direction it converges on is true.
The consequence: a population’s mental shape is set by the diversity of the directions being pushed at it. Having a wide diversity of direction and thus of mental shapes is typically called in Dutch pluriform: various cultures of thought can co-exists. Notice how this concept is definable without relying on the concept of truth.
3.1 Polarisation and orthogonalisation
Uniform broadcast and on-line media are characterized by pushing only one direction at everybody. That produces alignment, and alignment is not necessarily a sane thing. For example, Nazi Germany was an extremely well-aligned population. But at least it was a population that could still be argued with, in the sense that there was one shared axis and when the war ended that axis is what the argument could be had on.
Pluriform broadcast and on-line media are instead characterized by pushing different directions at each person. There is no longer a common direction being reinforced, and the population’s views are spread out — not into two camps but into as many directions as there are channels. In on-line media, there is a unique channel per individual, fully personalized to their preference and desire. If you want a single number for how far this has gone, take the number of genuinely distinct directions in circulation. One means consensus. Two means polarisation. Eight billion means something we do not have a word for and are already living in.
The model’s most useful output is that it demotes the thing everyone worries about.
Polarisation is two camps pointed opposite ways along a shared axis. It is loud, it is ugly, and it is survivable, because the shared axis is exactly what makes elections, courts, negotiations and arguments work. Two sides who hate each other are still answering the same question: structurally they have a great deal in common.
Orthogonalisation removes the axis entirely: two people having completely different problems are unable to reach each other at all. And it is enormously cheaper to produce than polarisation. To polarise a population you have to make a large number of people believe a specific thing, which requires a story plausible enough to be adopted and consistent enough to be maintained on both sides (e.g. a schism or a shibboleth). To orthogonalise one, you need none of that. You only need everyone’s inputs to differ! You do not have to be persuasive. You do not have to be right. You do not even have to be coherent from one recipient to the next: incoherence across recipients is in fact the mechanism of orthogonalisation, not a defect in it.
4 The cognitive attack
Military doctrine has caught up with this faster than public discussion has. NATO’s science and technology organisation runs a programme on what it calls cognitive warfare. Its stated goal is to alter and shape the way humans think, react, and make decisions. The report’s own summary calls the domain invasive, intrusive, and invisible. The third adjective is the one that matters.
The doctrine is explicit about the target. It is not infrastructure, and not even belief. It is the decision cycle itself. The capabilities are the ones that make a recommender system work: machine learning, communications, neuroscience, and population-scale measurement. They are simply pointed at human cognition rather than at military inventory. A related strand of analysis calls this a shift from control to influence. You no longer need to hold a territory or a broadcast tower. You only need to shape what its inhabitants are oriented towards. Social sculpting!
The cheapest version of the attack supplies no false claims at all. Give a watchful population ambiguous material and it will complete that material itself, into whatever narrative its culture makes available. The adversary provides the ambiguity; the victims provide the content. In the model, this is an input carrying almost no direction of its own, which amplifies whatever direction the recipient already had. It costs nearly nothing, it is nearly undetectable, and it drives the population’s directions further apart with every application.
5 Nobody records their inputs
Here is the practical failure, and it is a failure of instrumentation rather than of virtue.
If you want to know why a system is in the state it is in, you need to know its inputs. This is not a deep claim; it is the first thing anyone learns about identifying a system’s behaviour. Given only the output, a very large number of different internal stories fit equally well, and you have no way to choose between them.
Not many people record their own inputs. Maybe some scientists do, by means of keeping a bibliography. Not the popular social media platforms: while they surely record what they sent and use that to tune their predictive cognitive models, they are not likely to give this data back to the individual person. Most users did not track either. Nobody can reconstruct last Tuesday’s feed: it was never a document, and it was not stored anywhere the reader can reach. Not clinicians. Not researchers.
So when two people end up pointing their mind vector in different directions, the angle between them cannot be decomposed. It has no visible causes. An effect without visible causes gets attributed to the only thing still visible: the person. They are stupid. They are dishonest. They are deplorable. They are whatever. The attribution is nearly automatic, it is nearly always the wrong shape of explanation, and a missing input log guarantees it.
Two things make this worse than it looks.
First, an input’s effect is not confined to its topic. The clearest demonstration is fifty years old. People who read distressing newspaper stories then gave higher estimates for the frequency of deaths from entirely unrelated causes. The mood did the work, not the content. The effect did not depend on any similarity between what was read and what was later judged. The review literature calls this the carryover of incidental emotion, and notes that it typically occurs without awareness. So an input does not merely add a small component in its own direction. It transforms everything. There is no such thing as reading something “just for information.”
Second, the update function bends towards the state it is already in. We can say how much, because it has been measured on the sharpest possible case. Gampa and colleagues gave nearly three thousand people syllogisms to judge. Half were valid, half were not. The conclusions were politically congenial to some readers and uncongenial to others. Participants were asked one thing only: does the conclusion follow from the premises? That is a pure question with a right answer available to anyone. People judged arguments as more logically sound when they agreed with the conclusion, and they did this independently of whether the argument was actually sound. Liberals were better at spotting the flaws in arguments for conservative conclusions; conservatives were better at spotting the flaws in arguments for liberal ones. Mirror images.
Two details from that work carry the weight of everything that follows. In the second study, participants were trained in logical reasoning first, given worked examples and immediate feedback, and then instructed explicitly to judge validity alone. They scored 53%. The training did not help. Then, in the third study, on a nationally representative sample, the syllogisms were stripped of political content altogether: invented terms, no ideological freight. On those, the difference between liberals and conservatives vanished entirely.
Read those two findings together and they say something quite precise. The machinery is the same in everybody. What differs is what has been fed into it. That is the model’s central assumption, and it did not have to come out that way.
6 The one experiment that did keep the log
There is an exception, and it is instructive in a bleak way.
In 2014, researchers at a social network manipulated the emotional content of nearly seven hundred thousand people’s feeds. They then measured what those people went on to write. Inputs controlled, outputs measured, effect reported. Among population-scale studies of cognitive input, it is close to the only one that is published. And the journal that published it attached an editorial expression of concern, on the grounds that the subjects had not consented and could not have.
That reaction was correct. It is also an almost perfect trap. The one occasion on which somebody recorded the inputs is the occasion the scientific community sanctioned. The same manipulation runs unmeasured on every feed on earth since 2014. It draws no expression of concern from anyone, because there is no record for anyone to object to. We have arranged things so that measuring the experiment is a scandal, and running it is a business model.
The lesson is not that the study should have been let through. It is that consent and record-keeping have to arrive together: it cannot be delivered by the party doing the sending, it must be done by the party doing the receiving.
7 Keeping the input log
Simple idea: each person keeps a record of what was sent to them.
Not what they thought about it. Not a diary. Just a log of arrivals — what was put in front of them, by whom or what, when, in what order, and for how long it held them. A log of arrivals is not a log of responses. The first records the world acting on the person. The second records the person. Only the first is the missing instrument.
Three properties are not negotiable.
It has to be local. A complete record of everything a person has read is the most intimate dataset that could exist about them. It is more revealing than a medical file, because it holds the questions they asked when they thought nobody was counting. Held by anyone else, it is not a countermeasure to the problem in this article. It is the finished form of cognitive modelling.
It has to be yours. Not a feature of the platform, not a service with an account, not something that can be revoked. In fact, it could be a feature of a human operating system, that not only manages the system’s own resources but also manages the human resources of its operator. Protecting against information overload and long nights without sleep.
It has to be inspectable. A record you cannot read is telemetry. Possibly a local user agent can be asked to summarize or recognize patterns in the input log, to help processing the huge amount of data.
Note what the asymmetry currently is. The platform holds a near-complete log of what it sent you and what you did next. It uses that log to choose what to send you tomorrow. You hold nothing. One dataset in the world is unambiguously about you, and it is required to understand your own state. Everyone holds it except you. That is a strange arrangement to have arrived at without discussion.
It is also the arrangement we refused to accept for routers. Picture a network that learns what exists in the world from its neighbours, keeps no record of what it was told, and cannot compare its picture with anybody else’s. No operator would call that badly configured. They would call it impossible to run, because nothing going wrong inside it could ever be diagnosed. That is the condition every person reading this is currently in.
8 Comparing logs
The log on its own is only a record. What it is for is comparison.
Two people who see the world differently sit down — voluntarily, both consenting, neither having handed anything to a third party — and diff their records. Not their conclusions. Just their inputs.
‘We both have read this and that piece.’ ‘Only I have seen this.’ ‘Only you have seen that.’
What comes out is not a verdict. Nobody is shown to be right. What comes out is a set: here are the things one of us was given and the other never saw. Here are three years of one specific channel that went to you and not to me. Here is the fact that we have both been following this story and have never once been handed the same account of it.
The angle between two people, which was previously a brute fact about their characters, becomes a trajectory with a visible cause. ‘Let us not argue, but calculate: Calculemus!’ was said by Leibniz.
Be careful about what this achieves. The optimistic reading is wrong, and the evidence above says so. Comparing logs will not correct anyone’s reasoning. Explicit training in logical reasoning did not correct it, and a spreadsheet will not do better. The proposal does not try to move anybody’s vector. It makes the cause of the angle visible. That is a more modest thing, and the only thing the model supports.
A second finding makes the exercise better than it has any right to be. People are poor at seeing their own distortion and much better at seeing it in others. This is the bias blind spot. In the syllogism studies it showed up plainly: each side was the sharper instrument for detecting the other side’s bad arguments. Everyone is a bad auditor of themselves and a decent auditor of their opponent.
That is the case for doing this in pairs rather than alone. Form sublime unions: find your opponent. Neither person can see what their own input history did to them. Each can see what the other’s did. Each brings the capacity the other lacks. I tell you what I think your inputs did to you; you tell me what you think mine did to me. We are each wrong about ourselves in a way the other can partly repair. The disagreement stops being an obstacle and becomes the thing that makes the procedure work.
Even when nobody moves an inch, one thing is gained. An alternative account becomes available: perhaps this person is not stupid, and was simply sent different things than I was, for years, by a system that chose those things for reasons neither of us was told. That account is unavailable to almost everybody today. Not because it is implausible, but because the evidence for it has never been collected.
9 What this does not fix
The model is a sketch and should be held loosely. Real belief does not live in high-dimensional space, real updating is not a ‘rotation’, and the arithmetic of “how many distinct directions are in circulation” is an illustration rather than a measurement.
It also says nothing about truth. Two directions are just two directions; the model has no way to mark one as correct. But some questions do have answers. A framework that treats every orientation as merely a direction can slide into the position that nothing is more accurate than anything else. That position is the very condition the model is trying to describe. This is a description of how views come apart, not a licence to stop caring which of them is right.
Inputs are not the only cause. People differ in temperament, circumstance and interest before anything is sent to them. The log explains part of the angle and never all of it.
Also, the log can be forged. Once anyone acts on these records, there is a reason to manufacture them, and a synthesised log is no harder to produce than a synthesised video. Thompson’s old lesson holds here: you cannot fully trust a system you did not build yourself. That includes the very instrument being proposed to check the systems you did not build. The comparison is only as good as two people’s willingness to be honest with each other. It is a thin foundation, and the only one available.
And the log is dangerous. Everything that makes it useful makes it a weapon if it leaks. A person’s complete reading history is a better targeting file than anything a state could assemble by other means. Anyone building this has to treat it as the most sensitive artefact they will ever handle. It more likely belongs in an operating system designed to protect the human interests, than as an application that is designed to extract value from its human use.
None of these objections argue for the current status quo. Under it no record exists at all, held by nobody, while the party sending the inputs keeps a perfect one.
10 Conclusion
Take the first two senses of autonomy seriously at the same time and the conclusion is uncomfortable. A person is a system, in the plain technical sense. A person is also responsible for governing themselves, in the plain moral sense. The second cannot be discharged while the first is driven by inputs the person cannot see, cannot recall, and holds no record of.
We do not have to accept that, and the third sense is the reason why. The record is trivial to keep. We keep it for routers, and we would not run a network without it. What has been missing is the idea that the person is the one who ought to have it. And a second idea: that the record’s best use is not self-examination, at which everyone is hopeless, but comparison with someone who sees the world differently, at which everyone is surprisingly good.
Nobody will be talked out of their view by a log. That was never the point. The offer is smaller. Two people who cannot understand each other might at least see, in writing, that they were sent different worlds — and that this difference, rather than the other’s character, is what they have been looking at all along.
Further reading
Citation
@article{hiep2026,
author = {Hiep, Hans-Dieter A.},
title = {Is Every Human Being an {Autonomous} {System?}},
journal = {dr. heap},
volume = {3},
number = {2},
date = {2026-08-30},
url = {https://www.drheap.org/articles/2026/every-human-being-is-an-autonomous-system/},
issn = {3050-5224},
langid = {en}
}