ThinkableWhat is this?

The Ethics of Technology

The Ghost in the Training Data: When AI Learns to Sound Like You

Copyright law has always protected what an artist makes — but never who they are, and AI has quietly turned that loophole into a business model worth hundreds of billions of dollars.

The Idea

Here is the central paradox. Copyright has never protected style. It protects specific expression — the particular words on the page, the exact melody, the specific brushstrokes — not the aesthetic sensibility behind them. A novelist can read every Raymond Carver story ever written and then produce minimalist, working-class fiction all their life. That is called influence, not infringement. Courts have upheld this for a century. Generative AI exploits this principle at industrial scale. A model trained on thousands of an artist's works can produce output that feels uncannily like that artist — same cadence, same harmonic choices, same compositional instincts — without technically reproducing a single protected expression. The style is learned. The specific outputs are new. The law, as written, struggles to call that theft. But the litigation machine has found a different lever. The legal fight in 2026 is not primarily about whether AI outputs violate copyright; it is about whether training on copyrighted material in the first place is lawful. These are two distinct questions, and courts are answering them differently depending on what the AI was built to do and how its training data was sourced. Training a general-purpose model on lawfully obtained books has been found, in some courts, to be transformative — a new use that does not substitute for the original. Training a competing legal research tool on a rival's proprietary database has been found to be infringement. The act is the same; the market context changes everything. Style, meanwhile, remains the great unprotected frontier — legally visible, economically real, and entirely up for grabs.

In the World

On September 29, 2026, the Third Circuit became the first federal appeals court to rule directly on whether AI training can constitute fair use — and it said no, in this case. The dispute was between legal publisher Thomson Reuters and a now-defunct AI startup called ROSS Intelligence. ROSS had used Westlaw's proprietary headnotes — short editorial summaries of legal holdings — to train a competing legal research tool. The court affirmed that those headnotes were original, protected expression, and that training a product designed to compete in the same market was not transformative use. Crucially, the ruling explicitly distinguished itself from the generative AI training cases still working their way through the courts. Meanwhile, in a parallel universe of litigation, Anthropic had already settled with book authors for one and a half billion dollars — described by plaintiff attorneys as the largest copyright recovery in US history — after a court found that Anthropic had illegally downloaded and stored millions of copyrighted books from pirate repositories to train its models. The training itself was found to be fair use; it was where the copies came from that proved fatal. Then there is the music industry. Universal Music Publishing Group, Concord, and ABKCO filed a new lawsuit against Anthropic in January 2026, alleging that the company had used BitTorrent to acquire more than 20,000 copyrighted songs from shadow libraries — a training-data piracy theory targeting how the material was obtained, not purely what the model produces. By early November 2026, one tracker counted 169 AI copyright cases in total, and no two are quite alike. What courts are slowly establishing is the narrower point: how that learning was paid for, and whose market it is eating into.

Why It Matters

The style question is not just a legal technicality — it is a genuinely unsettled ethical problem that cuts to what we believe creativity is for. If style cannot be owned, then every human artist who spent years developing a distinctive voice has built something that anyone — human or machine — can replicate freely. That has always been true in theory. What has changed is the scale and the speed. A model trained on a living artist's entire output can produce thousands of works in their style per hour, with no payment, no credit, and no recourse. The law's response so far has been to focus on the pipes rather than the principle — asking where the training data came from and whether the AI competes with the source, rather than asking what it means to learn from someone without their consent. That framing may be practically necessary, but it leaves something important unaddressed. As a reader and consumer, it is worth asking what you think the implicit deal was when artists published their work online. As an economy, we are only beginning to work out whether the right to learn from someone's creative output — at scale, automatically, for commercial gain — is the same kind of thing as a human apprentice studying under a master.

A Question to Ponder

If a machine can learn everything about how you think and create, and reproduce it indefinitely without paying you, does it matter that it technically never copied anything you made?

One reader's Thinkable for Tech Tuesday — a random draw from the library.

Start learning with Thinkable
Get your own draw, every day.Start free