Sometime in the next few months Andrew Stellman expects a check from Anthropic: roughly $9,000 for him, about the same for his longtime coauthor Jenny Greene, and about $18,000 for their publisher, O’Reilly. The payment stems from a court settlement after pirated copies of their books — among many other works — were included in the training data for Claude. The settlement, which counted every edition separately, works out to a little over $3,000 per book shared among rights‑holders. The settlement received final court approval in July 2026, but processing such payouts takes time, so it remains to be seen whether the checks actually arrive.
Stellman says he doesn’t primarily care about the money — though he acknowledges nine grand is significant — but about getting visible credit. He would prefer that chatbots like ChatGPT, Claude, or Gemini indicate when an answer partly draws on his books, with a line such as, “Some of this comes from Andrew Stellman’s work; see this book to go deeper.” That kind of acknowledgement — a ‘‘hat tip’’ — matters to him more than the settlement payment.
Why attribution matters now
Beneath the technical and legal debates is a simpler emotional reality: people are angry and often with good reason. Artists and writers have seen models learn from material scraped without permission. Many creators fear that the work they rely on will be performed by machines that learned parts of the task from them. Even when use might be legally defensible as fair use, authors are upset that their work could be included without permission and often without any way to find out.
That frustration frequently funnels into copyright arguments, but copyright is a blunt, often misunderstood tool. Stellman argues copyright litigation is unlikely to deliver the kind of visible credit creators want, and that creators may gain more by negotiating with AI labs than by fighting in court.
What attribution actually means
Attribution is the chain linking a fact or idea back to its source. Bibliographies, references, Wikipedia footnotes, and news items crediting other outlets are all forms of attribution. Academic publishing is built on it: papers show whose work a new result is built on.
AI changes the picture because it can ingest enormous amounts of material and return answers with the chain erased. Live retrieval‑based citations (links to pages the system fetched while answering) are straightforward: they credit the pages consulted in real time. The much harder problem is causal attribution: did a particular work shape the model parts that produced this answer? That is the honest sense of attribution and the one most difficult to compute.
People often conflate three different ideas: resemblance (does the answer look like my work?), retrieval citations (links to live pages fetched during answer generation), and causal influence (did this source materially shape the model’s output?). The first is noisy and misleading, the second is easy to implement, and the third is the core and hardest form.
Can attribution be done technically?
A common objection is that models aren’t databases of sources: learning is distributed across billions of parameters, and there’s no receipt saying “this weight came from Stellman.” The brute‑force ground truth check would be leave‑one‑out: remove a source, retrain the model, and measure the change. For frontier models that’s computationally infeasible.
Researchers have developed cheaper approximations. One attractive approach is TracIn (Google researchers, 2020, “Estimating Training Data Influence by Tracing Gradient Descent”), which saves training snapshots and measures how much each example pushed the model toward a given answer by comparing gradients—no retraining required. Influence functions (Pang Wei Koh & Percy Liang, 2017) are an older method; Anthropic scaled similar ideas up to tens of billions of parameters in a 2023 paper (“Studying Large Language Model Generalization with Influence Functions”). TRAK from MIT’s Madry Lab addresses accounting problems around near‑duplicate sources. None of these are settled for frontier‑scale models, but the research direction is active.
Stellman tested the idea himself by asking an AI to build a very small model he could retrain quickly. He created “tiny‑provenance,” a roughly 37,000‑parameter language model that can be retrained in seconds, allowing an actual leave‑one‑out ground truth comparison against TracIn. He posed a trick question—“Who designed the Analytical Engine?”—where a look‑alike line about Babbage and the Difference Engine sat next to the correct source in the training set. A simple resemblance check wrongly credited the look‑alike; both leave‑one‑out and TracIn pointed to the source the model actually used. The lazy resemblance method was confidently wrong, while the influence‑based approaches were correct, and the whole test ran in a couple of minutes on a laptop.
That demonstration shows attribution is not categorically impossible, but it relies on a model small enough to retrain at will. Attribution works best when a model nearly memorizes a passage or leans heavily on a single source; it remains hard in the “muddy middle” where countless works each contribute tiny amounts. The fair summary: attribution is feasible in principle, but currently expensive and impractical at frontier scale. Stellman published the demo code and README so others can inspect the method.
The legal obstacle
Stellman argues that copyright is the wrong tool for the attribution problem. Courts so far have sometimes called model use “exceedingly transformative,” meaning the model changes the work enough to be treated differently than a straight copy. In the Anthropic case the payout stemmed from the use of pirated downloads, not the abstract transformation that comes from model learning.
Still, the market‑effect argument matters. If an AI gives free answers that compete with the books it trained on, that could be weighed against fair use. Some courts have hinted that market dilution from cheap substitutes could sway outcomes when argued differently. A real complicating factor is statutory damages: up to $150,000 per work for willful infringement. That number creates incentives for plaintiffs and their lawyers to maximize the number of works at issue.
That legal environment discourages voluntary attribution. If an AI lab writes “This answer draws on Andrew Stellman’s book,” that statement can be used to argue the lab knew whose work it used, potentially supporting a willful‑infringement claim. Given the risk, lawyers may advise companies to say nothing, because any good‑faith admission could be turned into litigation ammo. The result is that even if traceability technology improves, legal risk often keeps labs from providing open attribution.
What Stellman actually wants
Stellman proposes a practical, lower‑risk step: even when causal attribution can’t be proven, models can point users to authoritative, standard works on a topic—books and resources worth reading to go deeper. That doesn’t prove those sources shaped the specific answer, but it’s useful for readers and is technically and legally safer for labs to implement today.
He reiterates that credit and discoverability matter more to him than the settlement money; he would gladly trade some of the payout for greater exposure for his books. The “hat tip” is a small intervention with outsized benefits for authors if it increases readership and opportunities.
Finally, Stellman’s view is that attribution is more tractable than many who call it “impossible” admit: researchers and labs are making progress, and his demo shows feasibility at small scale. The main remaining barriers are cost and legal incentives—especially the specter of statutory damages—which shape whether companies will risk voluntary disclosure. He concludes that the hard part may not be the math but the legal and incentive structure around giving creators visible credit.
