Zero. That’s how many contract clauses from the 1990s were written with AI training settlements in mind. Which is exactly why the Anthropic settlement has turned into a fight over who gets to cash the check.
Authors are disputing claims made by publishers and agents on settlement payments, with some alleging unfair claims on older works. Mystery and thriller author April Henry summed up the mood publicly with a blunt “WTF is Harpe—” before the quote runs out, and honestly, that fragment captures the confusion better than a press release could. Writers are posting on social media that publishers appear to be claiming more than their fair share of certain payments. The result is a growing worry about whether settlement funds get distributed fairly at all.
I build bots for a living. I write ingestion pipelines, scrapers, retrieval layers, and the glue code that turns a pile of documents into something a model can answer questions about. So I want to talk about this from the seat I actually sit in, because the authors’ dispute is not really a legal story. It’s a data story, and it’s one that every one of us shipping AI systems should be reading closely.
This is a provenance bug, not a payout bug
The reason nobody can agree on who owns what slice of a settlement is that the ownership record was never machine-readable in the first place. Rights were split across contracts signed decades apart, amended by letter, transferred through acquisitions, and stored in filing cabinets and PDFs. When money finally arrives at the end of that chain, everybody points at their own copy of the paperwork and gets a different answer.
If you’ve ever tried to reconcile two systems that both claim to be the source of truth, you know how this ends. Somebody wins the argument by being louder or better resourced, not by being right.
That’s the lesson worth carrying into your own build. Most of us treat provenance as a nice-to-have field in the metadata blob. Source URL, maybe a fetch timestamp, done. Then a year later someone asks which documents in your index came from where, under what license, and with what expiration, and the answer is a shrug and a weekend of forensics.
What I actually track now
- Origin, not just location. Where the content came from originally, separate from where I got my copy. Those are different facts and they drift apart fast.
- Rights status as a first-class field. Not a note in a README. A column, with a value, that a query can filter on.
- Ingestion lineage. Which pipeline run pulled it, which transformations touched it, which index versions contain it. If I need to remove something, I want a delete path, not an archaeology project.
- An immutable log of claims. When someone tells me they own something, I record who said it and when, without overwriting the previous answer.
That last one is the piece the publishing world is missing right now. There’s no shared ledger of claims, so competing assertions collapse into a single confusing pile.
The scam wave is the second half of the story
Alongside the settlement dispute, fraudulent emails impersonating copyright offices are on the rise. Writer Beware has flagged deceptive messages from someone posing as the United States Copyright Office, asking recipients to verify their copyright registrations. Direct solicitation has become one of the main ways scammers recruit victims, along with ads.
This is the predictable consequence of money moving through a process most participants don’t fully understand. Confusion is the attack surface. When authors aren’t sure who legitimately needs their information, a convincing email asking them to “verify” something has a real shot at working.
For builders, there’s a direct parallel and a direct responsibility. Any bot that touches identity, payments, or document verification needs to be designed on the assumption that a well-crafted impersonation is coming. Out-of-band confirmation, verified sender domains, and flows that never ask for sensitive data over email are basic hygiene, not paranoia. If you’re building anything author-facing right now, assume your users have already been targeted this month.
What I’d take away from this
The Anthropic settlement was supposed to resolve something. Instead it exposed that the underlying ownership data was never solid enough to divide cleanly. Money arrived and the records couldn’t answer the question the money was asking.
Our systems will face the same test. Attribution, licensing, and takedown requests are going to keep getting more specific, and “we scraped it and lost track” won’t be an acceptable answer for long. The cheapest time to build a real provenance layer is before anyone asks you to prove something. The most expensive time is after the check clears.
Write the metadata now. Your future self, and possibly your lawyer, will thank you.
🕒 Published: