For knowledge owners
Most of what large models learned from was scraped from the open internet, much of it without the author’s knowledge. No public record identifies which works were taken, from whom, or on what terms. You cannot check, and neither can anyone else.
In Bartz v. Anthropic, authors discovered their books in a training corpus. The settlement — approved by the court in July 2026 — paid roughly $3,000 per eligible work, across 482,460 works.
Eligibility turned on one thing: the record. Statutory damages of up to $150,000 require a timely copyright registration, and books without one were excluded from the class. A work without an owner of record cannot be paid for.
Most owners never find out at all.
Bartz v. Anthropic, final approval 20 July 2026, N.D. Cal. The settlement covered 482,460 works, with claims filed for 440,490 of them. Sources: Authors Guild, Penguin Random House, JURIST, and Kluwer Copyright Blog.
Models retrain, new systems launch, and works are fetched and cited in generated answers every day. Use that predates your record is beyond reach, because no registration is retroactive. Use that follows it occurs against your name and your stated terms. When the $1.5 billion was distributed, it reached only the owners the court could identify from a record.
Litigation is expensive and its outcome uncertain. Selling a catalogue outright is irreversible. Blocking crawlers prevents the use and produces no revenue.
A registration adds a fourth position. You fingerprint the work in your browser, and the file never leaves your device. You state the terms: no AI use permitted, or AI use permitted with attribution. Pooled licensing at a set price, and per-work arrangements, are marked on the form as not yet available. The record carries the work, the owner, and the terms. Anyone holding the file can look it up. Free, and it takes about a minute.
In the European Union the position is settled. Article 4(3) of Directive (EU) 2019/790 permits commercial text and data mining unless the rights holder has expressly reserved it, and for work published online the Directive specifies machine-readable means as the way to reserve. Nothing obliges you to reserve. If you do not, mining your work is lawful. If you do, the reservation has to be readable by a machine at the moment the work is collected, which a copyright office certificate is not, in any country.
A file on your own server can carry that reservation, and where you control the domain it will. Most authors do not. Work published on LinkedIn, on Substack, or in a newspaper sits on a domain you cannot place anything at. A registry record attaches the reservation to the work rather than to the site.
Your record page, at its own permanent web address — point anyone to it: an AI company, a lawyer, a licensing counterparty, a platform.
Your page embed — a small “Registered with Alltio” marker for your website or the end of an article, linking to the record.
Recognition — anyone holding the file you registered can scan it, get the same fingerprint, and find your record. Checking the registry is free for AI companies, permanently.
A versioned history — register a new version and the registry links it to the old one. The registry never edits a record silently; a chain of versions is evidence.
Today this serves written work: the file you registered is the file the record identifies, exactly. Photographs, video, and audio are different — every platform re-encodes them, so a circulating copy no longer matches an exact fingerprint. Serving them honestly needs a second kind of fingerprint, one that survives re-encoding. It has a reserved place in the registry’s design and is on the roadmap; until it ships, we do not claim it. Datasets, CAD and engineering files move as exact files between parties, so exact matching serves them today.
Settlement is not live. Registration and the record are. When the payments start, the record is what they’ll pay against.