From the Desk of Ashok Mehan
I promised you this issue back in No. 7: a look at who owns the refinery, who supplied the crude, and what, if anything, we are owed. Here it is — a few days later than planned, for reasons that will be obvious to anyone who has tried to write honestly about their own complicity in something.
Nobody signed a contract that said "I will give you every search I have ever typed, every place I have ever driven, and every photograph of my children, in exchange for a map application and an email account that never runs out of storage." And yet that is, in substance, the trade thirty years of internet users made — one click of "I agree" at a time, on terms none of us read, for services that felt, and still feel, essentially free.
The word for this, coined by the Harvard scholar Shoshana Zuboff, is surveillance capitalism: the idea that human experience itself became the raw material for a new kind of market, extracted quietly as a byproduct of services we thought we were merely using.[1] We were never the customer of the free product. We were the deposit it was mined from.
I have been diplomatic for two paragraphs, and I want to stop being diplomatic. Somebody built this. It was not an accident of the internet's architecture, the way spam and pop-up ads were accidents nobody quite intended. In 1998, Sergey Brin and his partner, Larry Page, launched a search engine whose own founding research paper warned, in so many words, that an advertising-funded search engine would have a structural incentive to serve its advertisers rather than its users.[2] Two years later they built AdWords anyway. Every "free" product that followed — the email, the maps, the documents, the video, the phone operating system — was reverse-engineered from that one reversal: that the correct price for a search engine, it turned out, was your attention, mapped and auctioned. I do not think this was a conspiracy in the cartoonish sense. I think it was a decision made by two exceptionally smart people who correctly calculated that almost nobody would read the terms of service closely enough to object. But a decision made by identifiable people, for identifiable and profitable reasons, is not a force of nature. It has authors. I intend to keep naming them.
For most of the last twenty years, the business built on that extraction was advertising: know enough about a person to predict what they might buy, and sell that prediction. It was lucrative, and it was, in its way, honest about its own mechanics once you looked — the ad-targeting dashboards were right there if you went looking for them.
What has changed is not the extraction. It is the destination. The same trillions of words, images, and interaction patterns that once trained an advertising algorithm to guess your next purchase are now the training corpus for something that can write your emails, generate your images, and — per this issue's companion piece — increasingly write its own code. The con was never the free email account. The con is that we assumed the story ended there.
"We were never the customer of the free product. We were the deposit it was mined from."— Ashok Mehan
The training data behind today's large language models is, in the plainest possible terms, most of the public internet: forum posts, product reviews, news archives, personal blogs, code repositories, scanned books, and — where licensing deals now exist — the archives of publishers who once thought they were merely being indexed by a search engine, not fed wholesale into a model. Nobody who wrote a product review in 2009 imagined it would end up, in some statistically diffused form, inside a model worth hundreds of billions of dollars. That is the part of the con that is genuinely new: it is retroactive. The data was given away under one set of assumptions and repurposed under a completely different one, years later, with no renegotiation.
I want to be fair here, including to the company whose model I am dictating this issue to. In July 2026, a federal judge granted final approval of a $1.5 billion settlement in Bartz v. Anthropic — the largest copyright settlement in U.S. history — after Anthropic was found to have trained Claude on roughly seven million books downloaded from pirate libraries.[3] More than 440,000 authors and publishers filed claims. That is not a hypothetical about the future of AI and data. That is a check that has already been cut, for behavior that already happened, by the same industry writing this sentence for me. Google is fighting a similar suit from Hachette, Cengage, Elsevier, and novelist Scott Turow, filed in July 2026 over Gemini's training data.[4] As of September 2026, there were roughly 140 active U.S. copyright lawsuits against AI companies — up from 45 a year earlier.[5] The "just indexing" defense is not winning as often as it used to.
I like the refinery metaphor because I have actually run something adjacent to a refinery, and I know how the economics work. Crude oil is worthless in a barrel. It becomes valuable only after someone builds the infrastructure to crack it into something usable. The company that owns the refinery, not the country that owns the well, tends to capture most of the margin.
Data works the same way. Your individual search history is worth approximately nothing to you. Aggregated across a billion users, structured, and run through billions of dollars of compute, it becomes the substrate of a technology valued in the trillions. The people who supplied the crude — all of us, continuously, for two decades — hold no equity in the refinery. We get the free map. The refinery keeps the margin, and now, increasingly, the refinery's product is capable enough to compete for the jobs of the people who supplied its raw material in the first place. That is the closing of the loop this issue's title is really about.
I do not think the honest answer is a check. Individual data is nearly worthless; it is only valuable in aggregate, which is precisely what makes individual compensation schemes mostly symbolic. The more serious answers are structural: transparency requirements about what training data was used and under what terms; a real, enforceable right to have your data excluded from future training runs, not just from search results; and — the one I find most underrated — public investment in AI capability that is not owned by any single company, so that the returns on the collective raw material at least partly flow back to the collective that supplied it.
None of this is close to settled. The lawsuits are ongoing. The licensing deals being struck between AI companies and publishers are a real, if partial, acknowledgment that the original "it's just indexing" defense has quietly been abandoned by the industry itself — nobody licenses something they believe they were always entitled to for free.
The word "refinery" is not just my metaphor for this issue. It is the literal word my own site already uses, one essay over, about an entire country rather than a person: The Refinery Problem makes almost exactly the argument I am making here, just run at the scale of a nation instead of an individual. India generates roughly one-fifth of the world's AI training data and owns something like 3.5 percent of the compute infrastructure needed to process it — an asymmetry ratio of nearly six to one, data generated against refining capacity owned. The data leaves. The intelligence comes back. The value created stays abroad. That is the same con this whole issue has been describing in the personal case, except at the scale of 1.4 billion people instead of one inbox. If you want the version of this argument with hard numbers instead of my aggregate outrage, that essay is where they live.
From the Book · Memoir Passage
After SMAC, I built FedMine — a company whose entire business was the gap between the government's rules and its reality. Federal spending data came from more than twenty sources, each coded by different agencies under different regulations, all meant to describe the same dollars and almost none of them agreeing with each other. The rules said the data was consistent. The data said otherwise. I made a living in the space between the two.
What I learned building FedMine, in the basement, with no venture money and a great deal of stubbornness, is that data has almost no value in its raw state. Twenty scattered, contradictory federal databases were worth nothing to anyone. The value was entirely in the work of reconciling them — the unglamorous, unfunded labor of making disorganized information trustworthy.
That is the piece of this argument I keep returning to when people ask whether the platforms "stole" anything. Raw personal data, like raw federal spending records, is nearly worthless on its own. The theft, if that is the right word, was never really of the data itself. It was of the reconciliation — the enormous, collective labor of two billion people typing, searching, and correcting each other for twenty years, which is what actually made the corpus valuable, and which none of us were paid a cent for performing.
Looking Ahead · Three Trajectories for Who Owns the Corpus
Enforceable data-portability and exclusion rights become real, and some form of collective compensation — a public AI fund, a licensing pool, something — routes a share of AI's returns back toward the population that supplied the training corpus, the way sovereign wealth funds route resource revenue back to citizens.
Possible, not likely. The settlements so far — Anthropic's $1.5 billion to named authors and publishers — compensate rightsholders of record, not the billions of ordinary people whose searches, posts, and photos trained the ad-targeting systems that came before them. No mechanism yet exists for that broader class.
The lawsuits mostly settle into licensing arrangements between AI companies and large content holders — publishers, image libraries, stock-photo agencies. Individuals whose personal posts and photos trained the earliest models get nothing, since no legal mechanism exists to trace or compensate them specifically.
Most likely. This is already the visible trend — Anthropic's settlement, OpenAI's roughly two dozen publisher deals including a $250 million, five-year arrangement with News Corp, Disney's licensing deal with OpenAI.[6] The pattern is consistently institution-to-institution. No deal yet routes money to an individual whose blog post or product review was part of the corpus.
The technology becomes so normalized, and so useful, that the question of who supplied the original raw material stops being asked at all — the way almost nobody asks today what became of the oil fortunes built before modern royalty regimes existed. The con completes itself by becoming invisible.
Plausible over a long enough horizon, but not yet — 140 active U.S. copyright suits against AI companies as of September 2026, up from 45 a year earlier, says the question is currently getting louder, not quieter.
Sources & Further Reading