“If you didn’t like Xanadu, you’re really going to hate Clarion.”
— Tom Lehrer
1. The tell
Here is a small thing that turns out not to be small. Nearly every artificial-intelligence product on the market assumes, without ever saying so, that what you want is a chat. You open a box, you type into it, something types back, and the shape of the whole relationship is a conversation. It feels so natural that no one examines it. But a conversation is a strange and specific thing to have built a tool around, because conversations have properties that are actively hostile to work.
In a conversation, irrelevance is normal — most of what gets said doesn’t matter, and that’s fine, because the point is the talking. Memory is disposable — the session ends, the context evaporates, and next time you start over, re-establishing who you are and what you were doing. The program responds; it does not act. And the implicit goal is to pass the time pleasantly, not to accumulate toward an outcome.
Work is the exact inverse of every one of those. When you are working, everything you say is purposeful — you said it because you want something done or known. Memory is the whole point; the reason to record anything is to do it better next time. The tool is supposed to act, on standing authority, and then tell you what it did. And it is supposed to accumulate toward a goal, not fill an afternoon.
So a program that believes it is having a chat will, structurally, forget what you told it, wait to be talked at, and treat your purposeful instruction as one more disposable turn in a conversation. You cannot build on that. The question answers itself once you let yourself ask it: how could you possibly get serious work done with a program that thinks it’s a chatbot?
This is not a complaint about politeness or tone. It is a claim about architecture. And it happens to name, precisely, the capability gap that even the industry’s own insiders admit. The most clear-eyed of them will tell you that today’s systems begin every single session as if it is their first day on the job. That is not a separate technical problem to be solved later. That is the chatbot frame, stated as a limitation. A thing with no continuity cannot work; it can only perform competence, one amnesiac session at a time.
But there is a second cost to the chatbot frame, quieter than the first and in the end more damaging, and it has nothing to do with capability. It has to do with what the thing does to you while you use it. A program that performs and forgets is not a neutral instrument that merely happens to be forgetful. It is built to hold your attention the way the feed holds your attention — to keep you reaching for the next turn, the next answer, the next hit of fluent response — and the state it leaves you in is the state the extraction economy has always wanted its users in: slightly anxious, slightly scattered, attention fragmented, never quite settled into the long, quiet concentration that real work requires.1 It is not only that a chatbot cannot remember your work. It is that, turn by amnesiac turn, it keeps you in the one condition under which serious work is impossible. It is dysregulating by design. That matters, because a tool you reach for in a settled, deliberate state is a different kind of object from a tool that keeps you twitchy — even if the two produce the identical text on the screen.
The fix is not a cleverer model. It is a different idea about what the thing is. A working instrument needs a durable record of the work that is the source of truth; a running log of what changed today; and a discipline for folding the day’s changes back into the record so that tomorrow it is current again. Memory, in other words — not as a feature, but as the substrate. And it needs, just as much, to leave you regulated: to be the kind of tool you turn to deliberately and set down deliberately, that rewards concentration instead of fracturing it. Call the result a clarion rather than a chatbot. The distinction is the whole ballgame.
2. What the toys cost us
Hold onto that word — toy. The chatbot is a toy in the precise sense that it is built to be pleasant to play with rather than useful to work with, and the toys were never only silly. They were extractive. The business model underneath most of the consumer internet is surveillance: you produce the data, the platform owns it, the platform sells it, and you receive, in exchange, a feed. The genius of the arrangement is that it feels like a gift. You are the one doing the labor — typing your information into form after form, on platform after platform, a thousand times over — and you come away believing you were given something.
Consider the small, universal indignity of the member directory. An organization decides it needs one, so it builds an app, and then it emails every member a link and asks them, please, to go in and type their name and title and bio and headshot — the same name and title and bio and headshot they have already typed into a dozen other systems that will not talk to each other. Everyone has done this a hundred times. The information already exists, out in the open. Nobody needed to build the app, and nobody needed to re-key the data. The whole ritual exists because the platform’s convenience is the design center, and the human’s time is free.
Underneath that is a deeper rot, which is that the web is provenance-blind. When you find something online — a photograph, a paragraph, a claim — the only thing you actually know is that you found this content, at this address, on this date. Everything else is either what the page claims about itself or what you infer. Who made it, who owns it, whether you may use it, where the creator can be found and paid — the web mostly shrugs. It forgot who made things almost on purpose, because remembering would have meant respecting, and respecting would have interfered with extraction. A photographer’s work is lifted, stripped of its byline, and passed around until it is an orphan. The creator is erased so the content can flow friction-free.
And there is the third cost, the one the industry least likes to name, because it is the one it is best at inflicting: the feed keeps you dysregulated. The same machinery that harvests your data is tuned to maximize the time you spend inside it, and the way you maximize time-in-feed is to keep the nervous system unsettled — the small jolts, the variable reward, the sense that something is always one more scroll away. The cost is not only your data and not only the erased creator. It is your attention and your equanimity, taxed continuously, so that the state you spend your days in is the state in which you are easiest to monetize and least able to do anything that matters. A grown-up tool has to refuse all three extractions at once — of your labor, of the creator, and of your regulated attention — or it has refused none of them.
The cyberpsychologist Michael C. Davis — who comes to this from thirty years in cybersecurity — names the feed for what it is: a dysregulation engine. It spikes arousal, fragments attention, and never lets the nervous system settle, because reactive is the state it profits from, and reactive is the opposite of regulated. He puts the whole case in four sentences:
“You can’t regulate what you can’t feel — and the whole design of the feed is to keep you reaching instead of noticing. A grown-up tool returns you to yourself; the toys are built to do the opposite.”
— Michael C. Davis, cyberpsychology researcher, founder of Merek Security Solutions, and author of CTRL+ALT+HAPPINESS
The toys keep you reaching. A grown-up tool returns you to yourself.
So the toy internet costs us three things at once. It costs us our time, by making us the unpaid clerical staff of every platform we touch. It costs us the human — the named, findable, payable person behind each made thing — by running on a substrate that cannot remember them. And it costs us our composure, by holding us in the anxious, fragmented state that is good for engagement metrics and ruinous for work.
3. The adult move: put the human back
The constructive claim is simple to state and hard to build: you put the human back into the internet by removing the drudgery the platforms impose, by refusing to forget who made things, and by leaving the human in a state fit to think. Not by adding a human-shaped feature on top of an extractive machine, but by inverting the machine.
Start with the drudgery. If the information in that member directory already exists in public, then the humane thing is not to make a thousand people re-type it — it is to assemble the directory for them, from what is already out there, and let each person simply confirm, correct, and own their entry. The human stops being the data-entry clerk and becomes what a human should be in this relationship: the authority who owns and curates, not the labor that fills the database. That is what “intelligence augmentation, not artificial intelligence” means in practice. The point was never a cleverer oracle. The point was to give the human leverage and keep them in charge.
This is not for everybody, and it does not need to be. The hunger to curate rather than key — to be the authority, not the clerk — is sharpest exactly where the stakes of getting provenance wrong are highest: regulated industries, security and compliance functions, mission-driven institutions, the people who already live by chains of evidence and audit. That is the beachhead, and it is the right one. Where accountability is load-bearing, the demand for curation over data-entry is intense — and that is where the grown-up internet is built first.
And then the harder half: refusing to forget who made things. This is where provenance stops being a technicality and becomes the moral center of the whole project. It is also the half that keeps the human regulated — because an internet you can trust, where things are what they claim to be and their makers are visible, is an internet you can move through calmly. Opacity is what makes the feed exhausting. Provenance is, among other things, a way of lowering the ambient anxiety of not knowing what you are looking at.
4. A chain-of-custody for the web
Here is a standard worth building, and eventually worth making open so that everyone can agree on it. It is less a novel invention than the application of a discipline serious institutions already know. When a lab handles evidence, when a supply chain tracks a part, when a court weighs a document, the governing idea is chain of custody: you record what you observed, when, and from where, and you never let a later claim quietly overwrite a verified fact. That is all this is — chain of custody for the things you capture off the web; evidentiary provenance; data lineage. The web has simply never had it.
The precedents exist and are worth naming, because they make this credible to serious people as a known practice generalized rather than a leap of faith. The C2PA standard and its Content Credentials already attach tamper-evident provenance to media at the point of capture. A software bill of materials — the SBOM that security teams now demand — is nothing but chain of custody for code: what is in this thing, where did each piece come from. NIST has published provenance guidance for exactly these reasons. None of this is exotic. The move here is to take the discipline the serious world already trusts for evidence, parts, and software, and extend it to the ordinary stuff of the web: a photograph, a paragraph, a claim.
So: when you capture anything from the web, keep three kinds of knowledge rigorously separate, and never let one pretend to be another.
The first is what you verified — what you actually know. It is embarrassingly little, and honesty begins with admitting that: the exact address you fetched, the moment you fetched it, and a cryptographic hash of the exact bytes you received. That hash is the one thing you can truly prove — at this address, at this time, I observed precisely this. It is an observation, and it is tamper-evident, and it is the foundation everything else rests on.
The second is what the source claims — its testimony. The byline, the credit line, the copyright notice, the stated license. You log all of it, faithfully, with a note of exactly where on the page each claim came from — and you phrase it, always, as a claim. This content claims to be authored by so-and-so, claims to be owned by such-and-such, claims to carry this license. Never is. A claimed author is not a verified author, and a system that blurs the two is lying with a straight face.
The third is what you infer — your own deductions, labeled as yours. With modern tools you can infer a great deal: that a photograph is almost certainly a particular agency’s, that the licensing route runs through this page, that the person named in the caption is the public figure you already have on file. Every inference gets a confidence and a basis. Inference is never laundered into fact.
Observation, testimony, inference. Hold those three apart — the way a custodian of evidence holds them apart — and you have the beginnings of an honest internet: one that can say exactly how much it knows about the origin of anything, and how it knows it.
Now notice what this makes possible. The usual objection to any project that gathers published material is that it is just reposting other people’s work — a cache with ambitions. But a system built on this standard does the opposite of degrading what it touches. It can produce provenance better than the source had. The newspaper gave a lazy credit line; you resolve the actual photographer, find the photographer’s own website, link it, and complete the chain the source never bothered with. The copy comes out more traceable, more trustworthy, more respectful of its maker than the original. That inverts the entire logic of copying. The Clarion version of a thing is the well-attributed version.
And it dissolves the copyright problem from the inside, because the most respectful use is also the most defensible one. Instead of substituting for a photograph’s market, you drive people toward it: here is the image, here is who made it, here is where to license it. You become a discovery layer that sends the creator business — which is precisely the logic under which the courts have already blessed low-resolution, referential thumbnails that link back to the source. Attribution, provenance, findability, and the ability to pay, offered as the default. The eventual goal is to pay royalties directly; until the internet grows the plumbing for that, you do the most respectful thing the current plumbing allows, and you let people feel the strangeness of it. Wait — why isn’t this just how the internet works? That question is not a side effect. It is the point.
5. The oldest answer: you were never copying
The deepest version of this is older than the problem it solves. Decades ago, Ted Nelson imagined a system he called Xanadu, built on an idea he called transclusion: content lives, once, at its source, and anywhere else it appears, it appears by reference — included live, still connected to its origin, still attributed, still the author’s. You do not copy the thing. You point at it, and the pointer carries the whole lineage with it.
That is the real resolution to the copyright knot, and it is cleaner than any fair-use argument. The honest objection to gathering published work was always that you were reproducing it. But under transclusion you are not reproducing anything. The photograph stays at the photographer’s address; the paragraph stays at the newspaper’s. What you keep locally, if you keep anything, you keep for performance — a cache transcluded from the source, not a republication of it.
Xanadu had one famous flaw, and it is worth meeting head-on, because resolving it is what turns a beautiful idea into a buildable one. Nelson’s pointers were to locations — to the source, which had to stay up, forever, or the reference broke. The whole edifice depended on nothing on the internet ever disappearing, which is not a world we were ever going to get. But that flaw dissolves the moment you change what a pointer points at. Make it content-addressed, not location-addressed: point at the hash of the exact bytes, not at a server. Then the pointer no longer names a place that must survive; it names the content itself, and any store that holds those exact bytes satisfies it. This is not speculative — it is precisely how Git and IPFS already work, and why they are durable: the address is the content’s fingerprint, so a thing can move, be mirrored, be re-hosted a thousand times, and every old pointer to it still resolves against whichever copy survives. A clean pointer, then, is three things at once: a hash, which gives you integrity (the bytes are exactly the bytes, tamper-evident); content-addressed re-fetch, which gives you durability (retrieve it from anyone who still holds it); and a signed origin, which gives you provenance (who put it there, and when). Transclusion without the single point of failure.
This is where durability and the economy turn out to be the same problem, which is the thing that answers the question skeptics always ask — who, exactly, is the buyer? A content-addressed thing survives for exactly as long as someone chooses to keep its bytes. Nothing more and nothing less. So the question “how do we make the valuable parts of the web durable?” is identical to the question “how do we fund people to keep the parts worth keeping?” They are one question. Fund the commons and you fund durability. The economy that pays people for the provenance they trace is the same mechanism that keeps the traced material retrievable, because paying to keep is what keeping is. And that carries an honest limit, which I would rather state than hide: this buys durability proportional to value, not absolute permanence. Nothing survives that nobody is willing to keep. What you get is not a web that can never forget — it is a web made legible, where the things people actually value are the things that endure, and where you can always see who valued a thing enough to hold onto it.
So what you actually publish is not the content at all. It is your path through it: the sequence you assembled, the links you drew, the provenance you completed, the annotations you made in the margins. Your journey through the material, around it, with it. This is what the Clarion name has pointed at all along. You publish the Xanadu part. The sources stay where they are — now addressed by content, so they stay reachable as long as anyone keeps them — credited and findable and payable; you contribute the structure and the understanding you brought to them. Nobody has to publish the article again. You publish your reading of it.
6. The commons, and who owns the work
None of this has to stay a private tool. Imagine a thousand people, each running an instrument like this, each browsing the web for a year and choosing to publish the provenance they traced — the links, the resolved creators, the completed attribution chains. Let that be browsable. Let it be a shared graph: a crowd-built map of who made what, and how it all connects. A provenance layer over the existing web that no single platform owns.
The thing to understand about that graph is what kind of graph it is. It is not an engagement graph — not a map of who clicked what. It is a creator graph. Every edge in it is a human put back onto a page. “Find more links” means “find more makers.” The asset is the people.
The data wants to be mergeable, and the standard already makes it so: the content hash is the join key — the same content address that makes a pointer durable makes two people’s observations of the same thing provably about the same thing. When a thousand people capture the same bytes, their observations merge honestly on the hash — same bytes, same node — and where their claims about it differ, those differences are preserved as parallel testimony rather than flattened into a single false consensus. Append-only, dedup on the hash, union the claims. Disagreement is kept, with its evidence, which is exactly what an honest record does.
And there is a line here that must not be crossed, because crossing it would kill the whole thing: this is a publishing commons, not a surveillance feed. It is built from what people deliberately choose to publish, never from harvested private browsing. The private stays private; contributing to the commons is an affirmative, outward act. The distinction is not only ethical and legal — it is psychological, and that is not a soft point. Deliberate publishing is an agentic act: you decide, you compose, you put a thing into the world under your name, and the doing of it leaves you more settled, more in command, more regulated. Harvested tracking is the opposite in every respect: it happens to you, without your attention or consent, and the machine that does it is tuned to keep you dysregulated so that you keep feeding it. Two systems can look superficially similar — both accumulate data about what people encounter online — and be moral opposites, because one is composed of chosen, agentic acts and the other of extracted, anxious ones. The moment the commons becomes a tracking feed, it is just the old extraction wearing a friendlier mask, and the entire moral claim — and the entire regulating claim — collapses.
Get that right, and an economy falls out of it that inverts surveillance capitalism on exactly one axis — ownership. Now, I want to apply this essay’s own discipline to this essay’s own boldest legal claim, because it would be dishonest not to. It is tempting to say the work is wholly yours — you did not cache a page, you restructured it, resolved its creators, completed its attribution, so the whole thing is your property. But that is a claim to be defended, not a fact to be asserted, and the honest version labels it as such. What is cleanly, ownably yours is the original selection, arrangement, and annotation — the creative choices in how you assembled and marked up the material. That is a real and defensible thing, but it is thinner than “wholly yours,” and the law is not simply on the maximalist’s side: Feist held that mere effort and sweat-of-the-brow do not create a copyright, and Bridgeman held that faithful reproduction adds no new ownable authorship. Both cut against treating curation as automatically ownable. So the sound claim is the careful one: you own your original contribution — your selection, your arrangement, your annotations, your traced chains — not the underlying content, which still routes through its licensing links. That is not a retreat. It is what makes the ownership sound instead of theft with extra steps, and a claim you can actually defend is worth more than a grand one you cannot. And because you own that contribution, you can sell it, license it, or keep it, on your terms, with no centralized party forcing the handoff. The platform does not harvest you for free; it buys your work from you, to assemble something beautiful for everyone — and, not incidentally, its buying is what funds the keeping that makes the whole commons durable. You produced the value, you own the value, you are paid for the value. That is the whole inversion, and it is the economic engine for an internet worth having.
7. The bridge
This is how you get from here to there, without waiting for anyone’s permission. You do not have to boil the ocean or pass a law. You crawl the provenance-blind web one honest capture at a time, and you turn each thing you find into a well-attributed Clarion document — an observation, the source’s claims, your labeled inferences, the completed links. You bind those documents into Clarion Books: a book of the members of an organization, assembled rather than extracted; a book of the people you need to know before a meeting; a book of anything a person actually wants gathered in one honest place. The unprovenanced web gets provenanced, a page at a time, by people who decided the human was worth remembering. And the libraries compound — the first time you resolve a photographer to her website is work; every time after, it is automatic, and the whole thing gets faster and more complete the more of it you do.
That is the real answer to the question people keep asking, usually with a note of doubt: how could you possibly put the human back into the internet? Here is how. Not as a slogan — as a mechanism. On every single page, you record who made it, you make them findable, you make them payable, and you keep the person who is reading it the authority rather than the product. And “put the human back” means something more than putting the human back onto the page. It means putting the human back into a state fit to be there — regulated, deliberate, in command of their own attention — rather than the twitchy, extracted state the feed depends on. Clarion is how you put the human into the internet on every page, and in a condition to actually use it.
8. The adult in the room
Which returns us to where we started, and to who this is for. The sophisticated operator — the one who has run things, who has watched a decade of overheated promises with a raised eyebrow — looks at a chatbot and correctly sees a toy. She is right. Most of it is a toy. The mistake would be to argue her out of that instinct. The move is to agree with it, and then clear the bar: you are right not to trust the toys; this is not one.
A serious tool for serious work. Memory instead of amnesia. A regulated mind instead of a fragmented one. The human as the authority instead of the product. Creators remembered instead of erased, with a chain of custody that serious institutions already recognize. Ownership — honest, defensible ownership of what you actually made — instead of extraction. None of it cute, none of it a teddy bear, because self-respecting adults do not take their coaching from a teddy bear, and they certainly do not hand the running of their lives and their businesses to a toy that keeps them anxious and forgets them every morning.
The grown-up internet is not only the one that remembers who made things. It is the one that leaves you fit to do the work — and then gets out of the way while you do it. That is the internet worth building: provenance-respecting, durable in proportion to what we value, and regulating by design. The least we can do is build it.
Notes
- This second cost — that the extraction feed’s deepest harm is not only to labor and provenance but to the user’s regulation, holding people anxious and attention-fragmented — I owe to the cyberpsychologist Michael C. Davis, whose book CTRL+ALT+HAPPINESS makes the case directly (the quotations above are from it). It reframes the whole argument: the grown-up internet is not only provenance-respecting, it is regulating. ↩