An Economy That Does Not Sell the Original

So that memory does not lose its owner

Not long ago I wrote about my parents, in a piece called The Love That Survives Dislike.

It was about two people who have put up with each other for forty-seven years. As I wrote, there were passages I could not bring myself to put down, so I bleeped them out. What my mother said to my father, what my father did in return.

I did not have to reveal what happened in those spots to write about the time the two of them have lived together.

My mother read it and said: leave something long enough and it all gets good. That is miun jeong (미운 정), the bond that survives dislike.

Looking at the same years, I went hunting for words to explain them. She already remembered them another way.

I had not known it before I wrote, but afterward I saw that the bleeps were a surprisingly useful device. Because some things went unwritten, other things became writable. My parents’ lives are part of mine, but that does not mean everything I know about them is mine to put before the world on my judgment alone.

The more I think about recording people’s memories, the more this problem keeps coming back.

In an earlier essay I argued that we need to preserve citizens’ memories. The people who lived through war, industrialization, and democratization in their own bodies are still among us, and we now have the technology to take down their voices and organize them. There are stories that will never be heard again if they are not recorded now.

I still think so. But it turned out that being able to preserve something and being safe to entrust it with someone are two different questions.

Suppose a person has told you the story of their life. In it are things they never said even to their children, things about people who have since died, and passages where, even as they spoke, they were not sure they wanted them made public. Where should that story be kept? Who should be able to see it? What happens if they change their mind later?

And the person who listened, organized it, and keeps it safe for years: what are they supposed to live on?

The wish to preserve memory only becomes real work once it reaches these questions.

Nothing is preserved for free

A person’s story does not preserve itself.

Someone has to go and visit. You start with the weather, then the neighborhood, and little by little you enter that person’s time. You may not get it all in a day. They may say something different from what they said last time. The important part may come out only after the recorder is switched off.

Back home, you have to listen again. Check the names you misheard, look at where the speech broke off, ask how much may be kept. And once the record is made, it has to be stored, migrated, repaired when something breaks, and kept openable ten years from now.

AI can take a great deal of the labor out of this. What it cannot take over is the responsibility of having been entrusted with someone’s story. Someone has to carry that, and if that person is going to listen to the next story too, they have to be able to make a living at it.

So where does the money come from?

One easy answer presents itself. Why not sell the data we collect? AI companies need material, and here is human experience you cannot get anywhere else. Use the proceeds to pay the people who do the recording, and send a share back to the people who told their stories.

It sounds plausible. Companies have actually tried it. Datacoup started in 2012 and stopped in 2019. Wibson promised to let individuals sell their own data directly on a blockchain; today the company’s link leads to a service that deletes your data. The company that would help you sell your data became a company that erases it.

There is arithmetic behind the low prices.

Training a large language model from scratch takes tokens in the hundreds of billions. If the word is unfamiliar, think of a token as a fragment of a word. Now suppose a community works hard collecting people’s voices and amasses four thousand hours. That is more than a year of one person recording eight hours a day. Transcribed, it comes to roughly a hundred million tokens.

Hundreds of billions against a hundred million.

It is tempting to jump to a conclusion here: small, therefore worthless. That is wrong. Small datasets sell for a lot when they are used for evaluation or fine-tuning in a specific domain. In those uses, accuracy and rarity set the price, not size.

The problem lies elsewhere. The math fails when you try to sell that material as generic training feedstock and split the proceeds among tens of thousands of people. The unit price as raw material is low, while the cost of confirming whose data it is, getting consent, settling up, and distributing the money is high regardless. This is where the companies that promised dividends stalled.

Verifying bank accounts, handling taxes, and answering inquiries cost more than the money there was to hand out.

Nor is there any guarantee that the buyer shows up every year. The cost of collecting the material and confirming consent is incurred whether or not a sale happens.

Above all, once a record is handed over, the question of where it is used afterward remains. Some transactions are finished the moment the money changes hands. Being entrusted with a person’s story may be exactly when the hard part begins.

What is starting to carry a price

Over the past few years, though, something has changed in how AI companies acquire data. What sells is shifting from the material itself to the material’s history.

A few things set that off.

Lawsuits
On June 23, 2025, Judge William Alsup of the U.S. District Court for the Northern District of California ruled in a case three authors had brought against the AI company Anthropic. Training on books the company had lawfully bought was fair use, he held; the seven million books it had downloaded from pirate sites and stacked in a central library were not.

“It had no right to keep pirated copies in a central library.”

Same training. How the material was obtained decided the outcome.

Three months later the company settled for 1.5 billion dollars. Some 482,000 books, about 3,000 dollars per book. The largest settlement in the history of U.S. copyright.
Regulation
The European Union’s AI Act requires providers of general-purpose AI models to publish a summary of the data used in training. The European Commission issued the mandatory template for that summary on July 24, 2025. It applies to new models from August 2, and to existing models by August 2, 2027. Executive Vice-President Henna Virkkunen called it “another step toward trustworthy and transparent AI.”

Korea’s Basic Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust passed the National Assembly in December 2024 and took effect on January 22, 2026. Article 31 requires labels on generative AI output: either a label a person can recognize, or a machine-readable method such as a watermark or metadata. For deepfakes, only a visible label is allowed. After a one-year grace period, fines run up to 30 million won (roughly twenty thousand dollars).
Contamination
A paper by Oxford and Cambridge researchers, published in Nature on July 24, 2024, showed that if you train machines on machine-generated text for a few generations, the model collapses.

The rare things disappear first, until only the average is left. The title is the conclusion: “AI models collapse when trained on recursively generated data.” The paper added one more line: the value of material from real humans will rise as machine output spreads across the internet.

A rebuttal soon followed. If you keep the human data instead of discarding it, and add machine output on top, the collapse does not happen.

That changes the question. It is no longer whether to use machine output, but whether you can pick out which material came from humans.

And in the meantime the internet really did fill up. In a survey published in October 2025 by the search-optimization firm Graphite, across 65,000 English-language web documents, machine-written text first overtook human-written text in November 2024 and sat just above half by May 2025.

Another survey found that three of every four newly created web pages contain machine-written sentences. Which leaves the industry with one question. Where is the material that is certainly human?

All three changes point to the same place: provenance and history.

And the next question follows naturally.

Who should hold the authority to confirm that history, to set conditions, and to open and close the door?

Promises break when circumstances change

In 2025, during the bankruptcy of the genetic-testing company 23andMe, the question arose of whether its users’ genetic data might be sold or transferred to another party. It was data on more than fifteen million people. The chairman of the U.S. Federal Trade Commission sent the court a letter to the effect that any acquirer would have to honor the privacy promises made to users.

When the news spread, so many people rushed to delete their data at once that the login page stopped loading.

Someone who sent in a sample to learn about their family history was probably not thinking about how their data would be handled in the company’s bankruptcy proceedings.

Genetic data and oral records are not the same kind of material. Still, the question this episode left behind is a familiar one. If the circumstances of the place I trusted change, must what I entrusted follow those circumstances?

Companies get into trouble. Owners change. A line of business that did not exist at the start gets launched. If people’s memories acquire a new use every time, at the operator’s discretion, how far does the original consent reach?

That is why entrusting memory calls for durable mechanisms rather than good intentions. It is not only a matter of who may see what in ordinary times. It has to be settled in advance where the records go if the operator shuts down, and who inherits them, on what terms, when the person who gave them dies. The operator must not be able to open them alone.

It is not enough for rights to live in a contract.
They have to be built into the technology and the structures through which the data is managed and handled.

The one that survived

The Swiss data cooperative MIDATA was founded in 2015 and has been running for ten years. The method they chose was precisely the opposite.

Their articles of association say this:

Use of an individual’s data requires explicit consent (Article 5, paragraph 1).
The cooperative does not offer services that let individuals sell access to their own data for money (Article 5, paragraph 3).
No dividends are paid to members; any surplus goes to the continuity of the service and to purposes in the public interest (Article 6).
Fees may be charged for the platform and related services (Article 3, paragraph 2).

And two more things.

A data ethics committee, which reviews services and research projects, is elected separately by the general assembly, and its members may not also sit on the board (Articles 29 to 33).
If the cooperative is dissolved, all account holders’ data is deleted, but with a deadline to export it first (Article 35).

These are the articles adopted at the general assembly of June 2019. The chairman is Ernst Hafen of ETH Zurich.

The articles alone do not tell you how comfortable the cooperative’s finances are. Setting principles does not solve every operating problem either.

Still, one distinction is clear. They separated sharing sale proceeds with the people who provided the data from running a service through which people decide how their own data is used.

They did not gather goods to sell. They gathered things to protect.

The original.

Here I want to use the word “original.”

A digital file can be copied endlessly, so the word cannot mean the only file of its kind in the world.

Here, an original means

the reference record that keeps, together with the content, who said it, when, and in what process;
which passages were changed and how;
and how far its use was permitted.

There may be ten files, but there is one reference record.

So selling an original is not handing over a file. It is handing over control, or a permanent and unlimited right of use. Giving someone a copy and giving someone control are entirely different acts.

The same words carry different weight depending on who they were said to, and where.
A story meant for the family should not be treated as testimony released to the world.
Permission to use something for research should not be taken as permission to use it in advertising.

Protecting the original means never forgetting those differences.

What can be sold is concrete

The original is not goods for sale. It is the source and root from which everything that is sold grows.

If you keep the goose that lays the golden eggs safe instead of cutting it open, the economy that sustains it can take three forms.

First, making the record.

Going to visit, listening, listening again, organizing. Capturing one person’s life properly takes around twenty hours, and the traveling, preparing, and organizing around those twenty hours take several times that. This cost can be borne by the public. Every country has its corpus projects and its data-construction programs. But a program runs once. The money for making records stops when the program ends.

And the money that changes hands here is not a sale price for the material. Spending a budget on a road does not mean the road was sold. So a record made with public funds does not belong to whoever supplied the funds.

Second, protecting and proving.

Keeping a finished record openable ten years on, and confirming that this voice came from this person, on that day in that place, and has not been touched since. Preservation is continuous, not a one-time act, and this branch is what pays for it. Fees are charged for confirming provenance, managing consent, and opening and closing access, and those fees buy preservation.

Ten years ago, a recording was its own proof. Today a few seconds of a voice are enough for anyone to make a person say things they never said. So it now runs the other way: without confirmed provenance, a record becomes unusable. This proof has a peculiar property. It can be sold without handing over the material. The content stays locked; all that is confirmed is when it came from whom, and how. It is like certifying that the seal on a sealed box has not been broken.

One thing must be made clear, though. What is being proved is not that the content is true. It is that the person actually said it and that the record has not been altered since.

Memory can be wrong. Confirming that an old recording has not changed does not verify the memory inside it. Making a record trustworthy includes stating this limit precisely.

Third, what grows from the original.

A small book to leave for the children. A tool that retrieves the memories the person has allowed. The stories of people who lived through a region’s changes become an exhibition, a study, material for reviewing policy. A personal AI assistant that raises its owner’s productivity. An AI-assisted public forum for democratic deliberation, grounded in individual memories and relationships. Each time, the use is decided, as much is used as needed, and the people who do the work are paid. If there is revenue, what to return to the speakers and the community, and how, can be decided together.

What we can sell is this concrete. There is no need to decide what a person’s eighty years are worth. We only have to decide what it costs to visit that life ten times and listen, what it costs to organize it into one book, what it costs to support a permitted study.

And these three are paid by different parties.

The money for making records comes from the public: the data-construction projects of the Korean Studies Institute (국학진흥원), the culture-technology research programs of the Ministry of Culture, the oral history programs of local governments.

The money for protecting and proving comes from the institutions and industries that want to use the records: a museum paying a verification fee so it can attach source links to its docent script, a researcher paying an access fee under reading conditions, an AI company paying a fee to have the provenance of its training data certified.

The money for what grows from the original comes from individuals and institutions: a book for the children, a personal AI built from one’s own memories, a regional exhibition, a policy report.

Three revenue streams, each different, so if one is cut, two remain. In theory, after a public budget ends, verification fees and derivative works remain as other sources of income. The work does not hang on any one government’s budget, nor is it bound to any one company’s contract.

The three have one thing in common. The moment control of the original is handed over, all three branches drop at once.

If a single company monopolizes the records, the public’s case for funding their creation weakens, the power to trace who used what, to revoke, and to audit disappears, and the speakers and the community lose their seat at the table to negotiate over derivatives. So not selling the original is not a matter of being nice. It is how this work keeps its bargaining power.

Not selling the original is not a matter of being nice.
It is how this work keeps its bargaining power.

Somewhere already doing it

Kaituhi, the Māori-language transcription service built by Te Hiku Media in New Zealand, runs on a principle called kaitiakitanga. The word is close to guardianship: to protect, to care for, to hold in trust.

The service states that even after material is uploaded or processed, the authority to guard it stays with its original custodians. It also restricts using the service to mine Māori-language material or build a corpus from it.

The tool gets used, but the language and knowledge that come in through it do not become something anyone can take without conditions.

Looking at this case, I find myself thinking: if technology is to be built from our memories, shouldn’t our judgment go into the conditions under which that technology is used?

Even the word “our” needs care here. The name of the community must not override an individual’s refusal. One person’s story contains other people’s lives. That something happened to me does not make everyone who appears in it mine to expose.

So protecting the original cannot end with handing someone a key. There has to be a person who looks at where one party’s rights overlap another’s, covers the difficult passages, and asks again when needed.

Technology can help with that work. It can prevent an operator from opening a record or changing its terms of use alone. But it must never promise to recall everything once made public, or to undo books and studies already out in the world. Explaining in advance how far things can be reversed is also part of the job of holding a record in trust.

Before this economy can stand

There is reason for the public to share the cost of making records in the first place. The memories disappearing now cannot be remade later, however large the budget then. If preserving a region’s history and its citizens’ experience benefits the whole community, the cost can be shared the way we maintain libraries and archives.

Of course, public money is no reason to make everything public. From the start of any program, who retains which rights has to be settled.

Once a record exists, managing it remains: confirming provenance, managing consent, reviewing requests for use, logging who saw what. Researchers and institutions that need this service can pay for it.

On top of that, new uses grow: books, exhibitions, research. The people and institutions that want each of them can bear the cost. The cost of creating, the cost of long-term management, and the cost of new uses are carried in different places by different hands.

I think the structure is possible. But writing down three sources of income does not make the books balance on their own. You have to confirm who will pay, for what, and repeatedly. You have to be able to keep preserving after the budget that paid for the initial recording runs out. You have to calculate whether the revenue from popular stories can also protect the records almost no one asks for.

That Swiss cooperative wrote even this into the last of its articles: if the money runs out and the doors close, all entrusted data is deleted, but with a deadline to return it to its owners first. Not only the promise for when things go well, but a decision made from the start about how to return other people’s memories when they do not.

The market will not pay for the reason to protect records nobody asks for. The community has to supply that reason.

This is a business opportunity, and at the same time a question of what costs a community will bear together.

The right not to speak

Above all, there is something not to forget while doing this work.

People have the right not to speak.

The recorder wants to hear more good stories. The researcher wants to fill in the gaps, and the business wants to collect the material people will come looking for. When those desires pile up, at some point getting people to say more becomes the measure of success.

That is when you have to be able to stop.

“That’s enough for today” is also part of the record. “This part is for the family only” has to be respected. There may be stories that are never opened to anyone at all. That space has to exist for a person to tell you everything else.

Because of those decisions, there will be books that cannot be made and contracts that are not signed.

Protecting and earning do not always run in the same direction. Often enough, protecting well earns trust and lets you work longer, but there will be moments when immediate revenue has to be given up. There has to be a line held at those moments.

You cannot ask people to entrust you with their lives on the strength of rights that are honored only when they turn a profit.

The place of the memories that were not kept

In the last essay I wrote that when people’s memories come together, we can understand more deeply the society we have lived in.

I want to attach a condition to that now.

A large collection does not become the voice of the whole society just because it is large. The people who had no leisure to speak, who did not trust the recorder, for whom speaking was dangerous: they may be missing.

When reading the memories that remain, we have to account for the place of the memories that do not.

Where people remember the same event differently, the difference must not be erased. If a report is built on the records, the people it concerns need a place to read that interpretation and say it does not match what they meant.

For memory to inform a community’s decisions, there has to be a procedure for questioning the interpretation, as much as a procedure for collecting the memories.

That is how to avoid the accumulation of records becoming nothing more than the accumulation of the operator’s power. The point was never to build a structure in which whoever holds the most words decides the most.

Back to my parents.

The essay my mother read still had its bleeps in it. I did not expose those parts, and yet the room they left came across. The specific scenes behind the bleeps were never the point. I was able to say how the two of them had kept a place for each other, and my mother added her own reading to the piece.

That was enough.

I hope recording citizens’ memories can be like that. The person who leaves the record decides how much to leave; the person who listens remembers that intention; and whoever reads the record later knows the place that has been allowed to them.

Doing that work for a long time takes money. The person who goes to visit and listen has to be paid. There has to be somewhere to keep the records. And when someone knocks on the door ten years from now, there has to be someone to answer.

That cost is what we have to raise together.

I think of the moment when a person who has left their memory opens their own record again. The moment they can say: this much may be written, leave this part for now, that name is spelled wrong.

Keeping the place where they can say that, for a long time.

I want to build an economy that pays for that.

Continued from the earlier essay on legitimacy and collective memory.

Sources for the cases and figures

Datacoup, 2012 to 2019. Service shut down November 2019. Summary
Wibson, launched 2018. Its LinkedIn now points to illow, a data-deletion service. Observation
Bartz v. Anthropic, N.D. Cal. 3:24-cv-05417. Summary judgment 2025-06-23, 1.5 billion dollar settlement 2025-09-05, final approval 2026-07-20. The fair-use ruling covered only the three plaintiffs’ books; the class was certified on the pirated portion. Authors Guild explainer · Report on final approval
EU AI Act (Regulation 2024/1689), Article 53(1)(d). Commission template for the public summary of training content, 2025-07-24. Press release
Korea’s Basic Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust, Article 31. In force 2026-01-22. Summary
Shumailov et al., “AI models collapse when trained on recursively generated data,” Nature 631, 755-759, 2024-07-24. Paper. The paper’s last sentence: “the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet.”
Rebuttal: Gerstgrasser et al., 2024, arXiv 2404.01413. Accumulating human data instead of discarding it prevents collapse.
Graphite survey, 2025-10. 65,000 English-language documents. Judged by AI detectors, so the accuracy is disputed, and paywalled sites that block Common Crawl may have left human writing out of the sample. Summary
Ahrefs survey, 2025-04. 74.2% of new pages contain machine-written sentences; 2.5% purely machine-generated. Summary
23andMe bankruptcy, FTC chairman’s letter 2025-03-31. FTC
MIDATA articles of association, adopted at the general assembly 2019-06-26. Articles
Te Hiku Media, Kaituhi kaitiakitanga principle. Principle

Original Korean version written on September 15, 2026. Translated into English with AI in October 2026.