Where Things Come From

In the spring of 2008, a Burgundy producer named Laurent Ponsot flew to New York to stop an auction.

The sale included bottles of his family’s Clos Saint-Denis from vintages stretching back to 1945. The bottles were beautiful, the catalog was persuasive and the estimates were in the thousands of dollars. There was one problem. The Ponsot domaine had not made that wine until 1982. The older bottles could not exist and yet there they were, labeled, corked and offered for sale by one of the most respected auction houses in the country. Ponsot had the lots withdrawn. The consignor, a collector named Rudy Kurniawan, was later convicted of wine fraud and sent to federal prison.

I think about that story more than a person probably should. Wine is one of my real passions, and red Burgundy is the center of it, so I have a collector’s interest in the case. But what stays with me is the shape of it. Everything that made those bottles valuable was invisible. You could not taste a wine without opening it, and opening it destroys the thing you are evaluating. A buyer was paying for a story: who made this, when, where it had been kept, whose hands it had passed through. The liquid was almost secondary. The value lived in the provenance and the provenance was a lie.

The oldest question in commerce is “where did this come from?” It has suddenly become the most important question in technology.

An object with a history

I collect records too, mostly limited-run twelve-inch deep house and nu-disco from small European labels, along with some original jazz pressings. Record collectors develop a habit of turning a disc sideways under the light to read the dead wax, the smooth band between the last groove and the label. Etched there by hand or by machine are matrix numbers and sometimes small inscriptions from the mastering engineer. They tell you which plant pressed the record, from which master, in which run. Two copies of the same album can look identical and be worth wildly different amounts, because one of them can prove where it came from.

Anthropologists noticed long ago that objects carry their histories with them and that the history is often the point. In the Trobriand Islands, Bronisław Malinowski documented the kula, a ring of exchange in which shell necklaces travel in one direction around a circuit of islands and armshells travel in the other. The objects have little practical use. Their value comes from where they have been. A famous necklace is known by name and so are the people who have held it. Possessing it for a while makes you part of its story and the story is what everyone is trading.

Marcel Mauss built a theory of society on observations like these. In The Gift, published in 1925, he argued that in many cultures a gift is never fully separated from its giver. Something of the person travels with the object and creates an obligation to give, to receive and to give back. Modern markets were designed, in part, to cut that cord. The anthropologist Igor Kopytoff wrote about the “cultural biography of things,” and observed that commerce pushes objects toward becoming commodities, interchangeable units with no particular past, while culture keeps pulling some of them back toward being singular. A dollar is a dollar. A wedding ring is not a gold ring.

The simplest way I know to put it: an object without a history is a commodity. An object with a history is a relationship.

The largest act of borrowing in history

Now consider what a large language model is.

It is a system that has learned to write, reason and converse by studying an enormous portion of everything humans have written down. Books, articles, forum posts, legal opinions, recipes, poems, arguments, instructions, love letters made public. Whatever else these systems are, they are an act of collective inheritance on a scale without precedent. Every fluent sentence one of them produces is built on patterns that millions of people laid down, almost none of whom were asked.

It was inevitable that the question of provenance would arrive, and it has arrived through the courts.

In June 2025, Judge William Alsup in San Francisco drew a line through the copyright question that had been hanging over model developers. Training a model on books, he held, could be fair use where the books had been lawfully acquired. Building a library out of pirated copies was a different matter, and not protected. A separate decision involving Meta reached a favorable result for the developer on the record before that court, while making clear that the outcome depended heavily on the evidence and the arguments presented. The second half of Judge Alsup’s ruling is what moved money. It led Anthropic to a class settlement of $1.5 billion covering roughly half a million books, about $3,000 per work, the largest copyright recovery in American history. The court granted final approval in July 2026.

Then, on September 29, 2026, the first federal appeals court to rule squarely on fair use in AI training decided Thomson Reuters v. ROSS Intelligence. The Third Circuit held that ROSS’s use of Westlaw’s editorial headnotes to train a legal research tool was not fair use. The court emphasized that ROSS’s system purportedly did not generate new expression and that it allegedly competed in the same market as the material it learned from, including an emerging market for licensing that very material as training data. The court was explicit that it was not deciding the harder questions raised by generative models. Those are still working their way through other courts.

Read together, these decisions did something easy to miss. They did not resolve whether building models from copyrighted material is lawful in general. What they established is that how the material was obtained and what the model is used for carry decisive weight. That converts an open question of law into a set of questions of fact. Where did the data come from? On what terms? What does the system do with it? Who does it compete with?

Questions of fact can be investigated, documented, warranted and priced. Provenance has moved out of philosophy seminars and policy debates and into the part of a commercial agreement where people already know how to work.

An object without a history is a commodity. An object with a history is a relationship.

The lemon problem, again

In 1970 the economist George Akerlof published a short paper about used cars. A seller knows whether his car is good or a lemon. A buyer cannot tell. Because buyers cannot distinguish, they will only pay the price of an average car, which means owners of good cars withdraw from the market, which lowers the average, which drives out more good cars. Left alone, a market with hidden quality can unravel. Akerlof’s point was that the institutions we take for granted (warranties, brands, dealer reputations, licensing, inspection) exist largely to rescue markets from this spiral. They let information that one party holds become something the other party can rely on.

An AI transaction today is a lemon problem with unusually high stakes. A company licensing a model cannot audit what it was trained on. It has no practical way to verify what went in and often the vendor’s own visibility is partial, because data arrived through layers of upstream licensing where assurances thin out at every handoff. So the buyer asks for an indemnity, reasonably, because it cannot protect itself from a risk created upstream. The vendor resists, also reasonably, because it is being asked to underwrite an exposure no one can size, in an area where statutory damages can reach six figures per work.

Both positions are defensible, which is exactly why these negotiations take so long. In a mature market, precedent does most of the work and negotiation narrows to price and a handful of live points. Here, two sophisticated parties can hold opposite views on a central term and each be right, so every deal gets built from first principles.

Kenneth Arrow, another economist who thought hard about information, wrote in 1972 that “virtually every commercial transaction has within itself an element of trust, certainly any transaction conducted over a period of time.” Contracts are how strangers manufacture trust they have not had time to earn. In AI agreements, the substance now sits in a few places where that manufacturing happens:

Whether the indemnity covers the model itself or only its outputs and how claims arising from the customer’s own prompts are treated.

Who owns the outputs and whether the vendor keeps a right to learn from customer interactions. That right is often worth more than the license fee and it deserves to be priced as consideration rather than tucked into a data section.

What a performance warranty can even mean for a probabilistic system, where conventional acceptance testing and service levels need to be rethought rather than reused.

Whether the customer is told when the model changes, since the system evaluated in the spring may not be the system running in the fall.

What outside testing, audit access or third-party attestation the vendor will support. In my experience the answer to that question is among the most informative things a buyer can learn.

Each of these is a way of answering, in contract language, the question Laurent Ponsot answered by getting on a plane.

Rules arriving on their own clocks

All of this now sits alongside regulation that did not exist when most software agreements were templated and the regulation is itself moving. Europe passed a comprehensive AI statute and then, in 2026, voted to push its obligations for high-risk systems back to late 2027 and 2028, while keeping transparency duties for AI-generated content in place from August 2026. Texas’s Responsible Artificial Intelligence Governance Act took effect at the start of 2026. Under these regimes, a meaningful share of the obligation falls on the company that deploys a system, not only on the company that built it.

That changes the character of the negotiation. Allocating regulatory responsibility between the parties is no longer a boilerplate compliance clause. It is a commercial term with a number attached. The companies treating it that way are getting better agreements and the ones treating it as boilerplate are inheriting risks they have not priced.

Reading contracts after something went wrong

Most advisors who work on these deals came up through transactional practice. I came from the other direction. For about twenty-four years I was a litigator, much of that time in technology and intellectual property disputes, often in situations where the law had not yet caught up with the technology in front of it.

Litigation is where contract drafting gets graded. Two decades of reading agreements after something has gone wrong teaches you which provisions actually do the work under pressure, as distinct from the ones that consume the most hours in negotiation. The two lists are not the same. The clauses people fight hardest over are often the ones that never matter and the ones that decide the outcome are often a definition, a notice requirement or a sentence about records that nobody thought twice about. Once you have seen that enough times, you draft differently going in.

Provenance clauses will be graded the same way. Very little has yet tested how an output indemnity or a model performance warranty actually behaves when a deployment causes real damage. The companies negotiating these terms now are not applying a market standard. They are writing it and the terms they settle on will shape the next decade of agreements.

Provenance as a form of respect

I want to step back from the deal room, because I don’t think provenance is only a commercial problem. Underneath the indemnities and warranties is a much older human question about acknowledgment.

For more than fifteen years I helped the jazz vocalist Patty Waters with her recordings, royalties and touring. Her mid-1960s recordings for ESP-Disk influenced generations of singers and her music outlived most of the business arrangements built around it. Working with her taught me that great art tends to outlast the commerce attached to it and that the people who make it deserve careful stewardship of their rights and their legacy. It also taught me how much history can live in one person’s memory and how easily it is lost when nobody asks where something came from.

Medieval scribes sometimes added a colophon at the end of a manuscript: a few lines recording who copied it, where and when, occasionally with a complaint about cold hands or a request for the reader’s prayers. It was a small act of provenance and a small claim to be remembered. Our machines are now the largest copyists in history. The question of whether they will carry anything like a colophon, some trace of the people whose work they learned from, is partly legal and partly economic. It is also a question about what kind of relationship we want to have with the tools that are absorbing our collective voice.

The markets that work best over time are the ones where people can answer, honestly and with documents, the question of where things came from. They are the markets where good bottles can be distinguished from forgeries, where the record in your hands can prove which press it came from and where the people whose work created the value can see their part in it.

Provenance is how we make the invisible trustworthy. It is also, quietly, how we say thank you.