I still remember the first time I heard someone casually mention that a frontier model had “read” every book ever written. It sounded almost romantic, like the machine had spent quiet nights in a vast library. Then the numbers landed and the romance evaporated. One company just agreed to hand over 1.5 billion dollars because those books were never paid for. That figure is not a licensing fee. It is the price of treating a shadow library like a free warehouse.
The Real Cost of Skipping the Checkout Line
A federal judge in San Francisco recently signed off on the largest publicly reported copyright recovery anyone can recall. The settlement covers hundreds of thousands of titles. Roughly three thousand dollars per book before legal fees. A typical hardcover sells for around thirty. The rest is the surcharge for taking the digital copy instead of buying it.
The company won on the training question. The judge called the use of lawfully acquired books “exceedingly transformative.” That part of the ruling was clear. What sank the defense was the sourcing. Millions of titles had been pulled from known pirate repositories and kept in a permanent internal library. The court drew a bright line: downloading material you could have purchased is still ordinary infringement. Fair use does not wash the acquisition clean.
I’ve spent decades looking at how institutions handle assets. When a firm cannot show a clean chain of title, we do not call it a philosophical disagreement. We call it a diligence failure and we price it accordingly. For three years the industry debated epistemology while the real problem sat in the inventory logs.
Three Separate Questions That Keep Getting Tangled
Fair use is an affirmative defense. You admit the copying happened and then try to justify it. That is why a tool that ingested student papers for plagiarism detection survived, and why a search index that scanned millions of volumes also survived. A lawful downstream use does not erase an unlawful upstream acquisition. A stolen car still gets you to the office. It remains stolen.
Three distinct questions keep getting mashed together. How was the data obtained? What did the training process actually do? And what comes out the other end of the model? Separate those and the case law becomes less confusing, though still messy.
Within days of the first ruling, another judge in the same courthouse reached a different conclusion on similar facts involving another large lab. That second opinion treated the transformative nature of the ultimate use as enough to cover the downloading itself. The judge also flagged the risk that cheap machine-generated substitutes could dilute authors’ markets, then ruled against the authors because the arguments presented were not the strongest available. Two judges, one week, two different answers on the piracy question.
A few months later an appeals court affirmed a loss for a smaller company that had copied specialized legal headnotes to build a competing research product. The panel described the use as only minimally transformative and stressed that the system was not generative and that the defendant competed directly with the original publisher. The pattern is becoming familiar. When the output substitutes for the original work, the fair-use shield thins quickly.
Dirty Hands on Both Sides of the Table
Both camps have behaved in ways that undercut their public talking points. In public, AI companies speak of fair use with the fervor of recent converts. In private, once the download records surface, the checks get written. Settling the piracy claim while rightsholders claimed coverage of more than ninety percent of eligible titles was smart litigation. It does not make the public argument and the private settlement tell the same story.
Publishers are not pure either. Some of the same organizations that file lawsuits for “massive freeriding” have simultaneously signed licensing deals with other labs. One executive described the approach as a “woo and a sue” strategy. When identical conduct is labeled partnership in one conference room and theft in another, you are listening to an opening bid, not a settled principle.
I’ve watched this pattern in other industries. The party that can afford the lawyers gets to keep talking about principle while the settlement numbers quietly rearrange the incentives. That does not make the underlying legal questions disappear.
Who Actually Benefits From Strict Licensing Rules
Anyone who cares about competitive markets should feel a little uneasy here. Mandate fully licensed training sets and you create a fixed cost that favors the largest players. The money flows toward big publishers, not the freelance historian or the small regional paper that never had a licensing department. Similar compliance rules in other fields have repeatedly shown that smaller organizations pay a higher relative price.
The opposite extreme is worse. Let everything slide and the least careful operator obtains the cheapest data while the firm that tries to do it right absorbs costs its rival simply skipped. Bad practice drives out good. That story is older than any of us.
The strongest practical objection to a clean rule is infrastructure. Court documents showed one major lab pausing its own book-licensing efforts after very slow uptake. Fiction publishers sometimes admitted they did not even control the necessary rights. Demand universal licensing when the pipes do not exist and you have effectively written a ban dressed up as a market solution. Music faced the same fragmentation a century ago and eventually built collective systems with published rates. Fragmentation is an argument for building the pipes. It is not an argument for taking the water without asking.
A Simple Rule That Could Actually Fit on an Index Card
Start with a hard line against piracy. Copies taken from known shadow libraries should be treated as ordinary infringement with no fair-use defense available. Labs ought to document provenance the same way a bank documents collateral.
Regurgitation of protected expression remains infringement no matter how clean the training set. Verbatim or near-verbatim output should stay actionable, with a practical safe harbor for companies that design and audit output filters. One of the judges noted that the settling company already filtered outputs. That approach also answers the market-dilution concern raised in the second ruling: police substitution where it actually appears.
Lawfully acquired material should carry a default presumption of fair use for training. If the company bought the book, licensed the text, or collected it from the open web without breaching a contract, the training itself should be protected. For material that cannot realistically be licensed title by title, a collective licensing body with a published rate, an opt-out registry, and a small-claims track would be far more workable than endless case-by-case litigation.
Congress is the proper place to write those rules. Courts are already producing divergent results, and foreign jurisdictions are splitting as well. One European court found infringement in song lyrics stored inside a model; another found no storage of protected images. Leaving the entire field to individual judges guarantees years of uncertainty that neither authors nor engineers can plan around.
The Cheapest Path Was Always Available
The company that just wrote the 1.5 billion dollar check already knew the honest route. It purchased millions of used print books, removed the bindings, and scanned them. The same judge who condemned the pirate downloads called that process fair use. At roughly thirty dollars a copy, half a million books would have cost about fifteen million. On the scale of a frontier training run that number is a rounding error. The firm chose the other path and paid roughly one hundred times the sticker price.
I want the models. I want better protein folding, earlier cancer detection, and the tutor that can still answer questions at three in the morning. None of those benefits requires treating the largest technology firms on the planet as exempt from the checkout line. Analysis can be free. Expression remains owned. Theft is still theft even when the thief is building something impressive.
The Copyright Clause exists to promote progress by securing authors’ rights for limited times. Progress is the purpose. Ownership is the means. When the means are discarded because the product looks dazzling, we have simply chosen one set of owners over another and labeled the choice inevitable.
What Provenance Tracking Would Actually Look Like
In my experience the firms that survive long-term regulatory pressure are the ones that treat data lineage the same way they treat financial controls. Every major training run should be able to produce a verifiable map: where each corpus came from, whether a license or purchase receipt exists, and how the material was filtered before it entered the training mix.
That is not romantic work. It is inventory control. Banks do it for collateral. Pharmaceutical companies do it for clinical data. There is no technical reason AI labs cannot do it for text. The obstacle has been cultural. For a while the prevailing attitude treated the open internet and the shadow libraries as interchangeable free resources. That attitude just became expensive.
A practical system would combine cryptographic hashes of source files, timestamped purchase or license records, and automated exclusion lists for known pirate repositories. None of this requires new science. It requires the decision to treat data acquisition as a first-class risk rather than an afterthought.
Output Filters and the Substitution Problem
Even perfect training data does not eliminate the risk that a model will reproduce protected expression. The second judge in the pair of rulings worried openly about market dilution from machine-generated substitutes. The practical answer is already visible in the first case: the settling company had implemented output filters. Those filters can be audited, improved, and made part of any safe-harbor regime.
I have seen enough technical presentations to know that perfect prevention of regurgitation is difficult. Near-perfect prevention is achievable and already being done by responsible teams. The law does not need to demand perfection. It needs to demand that companies treat the problem as real and measure their own success against it.
When a model produces long stretches of copyrighted text, the right response is not a philosophical debate about transformation. The right response is to treat that output the same way any other unauthorized copy would be treated, while giving credit to the firms that invested in prevention.
Collective Licensing as the Missing Pipe
Music rights were once just as fragmented. Individual songwriters could not negotiate with every radio station or record label. Collective societies and statutory rates solved the coordination problem without eliminating ownership. Text can follow a similar path.
A collective body with a published per-token or per-title rate, an opt-out registry that actually works, and a small-claims process for disputes would remove the current all-or-nothing dynamic. Authors who want to stay out can stay out. Authors who want compensation can receive it without hiring a litigation team. Labs that want clean data can buy it at a known price.
Until something like that exists, courts will keep producing split decisions and companies will keep treating settlement payments as a cost of doing business rather than a signal to change acquisition practices.
Why the Fifteen-Million-Dollar Alternative Still Matters
The fact that the company already ran a clean scanning program is the detail that keeps bothering me. They proved they could acquire books lawfully at scale. They chose not to for the bulk of the corpus. The later settlement is therefore not a surprise. It is the predictable price of treating the lawful route as optional.
Fifteen million dollars for half a million physical books is not a rounding error to an individual author. It is a rounding error to a frontier training budget. When the largest players decide that rounding error is still too high, the rest of the market receives a clear message about how seriously ownership is taken.
I keep returning to the same simple observation. The technology is remarkable. The acquisition shortcut was not required for the technology to exist. Pretending otherwise only teaches the next generation of labs that the checkout line is for someone else.
Looking Ahead Without the Romantic Fog
The models will keep improving. Protein structures will get better, diagnostic tools will get sharper, and late-night study sessions will have more capable tutors. None of that progress depends on treating copyrighted books as free inputs. It depends on compute, algorithms, and carefully curated data. The data can be curated honestly.
Courts have begun to separate the training question from the sourcing question. That separation is healthy. Training on lawfully obtained material can remain protected. Downloading from pirate repositories should not. Output that substitutes for the original work should remain actionable. Those three principles fit on an index card and still leave room for genuine innovation.
Perhaps the most interesting aspect is how quickly the industry’s private behavior adjusted once the download logs became evidence. Public rhetoric about fair use continued, yet the settlement checks were written. That gap between speech and conduct is where the real policy work still needs to happen.
A country that lets its largest technology companies skip the ordinary rules of acquisition because the resulting product is impressive has not chosen progress. It has chosen one set of owners over another and called the choice inevitable. Analysis can stay free. Expression stays owned. The difference is not philosophical. It is the difference between a library card and a warehouse raid.
The 1.5 billion dollar figure will fade from the headlines. The principle underneath it should not. Provenance matters. Licensing infrastructure can be built. Output filters already work. The honest path was always cheaper than the settlement that just closed. The only remaining question is whether the next training run will remember that fact before the logs are subpoenaed.
