What if your reader isn't a human?

For decades, the publishing industry has been organized around one assumption: the reader is human. One book, one copy, consumed at a roughly human pace.

That assumption is breaking down.

The new reader is a machine. AI systems already consume content at a scale humans can't match — and consumption is projected to grow by a factor of ten, or a hundred, over the next few years. The mechanics are different too. A machine doesn't read sequentially; it ingests, indexes, retrieves, and generates. The old industry was built for the human reader. The new infrastructure has to account for both.

AI companies are functioning as publishers

Look at what they do: acquire content (often by scraping), build models, distribute outputs, charge per token. The mechanics differ from traditional publishing — they don't always pay for inventory, and they sell at a level that would have seemed too cheap to meter only a few years ago — but the role they play is structurally similar.

This matters because the rights frameworks publishers have always relied on were built for a publisher-to-reader transaction. The new transaction is publisher-to-AI-to-reader, and most rights infrastructure stops at the first arrow.

Two ways AI uses publisher content

There are two distinct mechanics, and they require different rights treatments.

Input licensing governs training. A model is trained once, on a vast and mostly fixed dataset, then frozen. If your content was in that dataset, the knowledge from it is baked into the model — even when the model never reproduces your text verbatim. Once training is done, the relationship to your content ends, but the value extracted persists.

Output licensing governs retrieval. When a model uses RAG (retrieval-augmented generation), it goes out at query time, pulls in current content, grounds an answer in it, and discards. The content isn't baked in. It's a one-time use supporting a one-time output.

A rights declaration has to be specific enough to address both.

The regulatory reality

Article 4 of the EU Copyright in the Digital Single Market Directive is unambiguous: text and data mining is permitted unless the rights holder has explicitly reserved the work in a machine-readable form.

The implication cuts the other way from the language. If you don't opt out, you've effectively opted in.

August 2, 2026 — the deadline for the EU AI Act's enforcement provisions on General Purpose AI model providers — is when that asymmetry becomes consequential. AI developers serving the EU market will be required to detect and honor declared reservations. A declaration made in late 2026 cannot retroactively cover content ingested in 2024.

The gap between regulation and infrastructure

The law requires machine-readable opt-outs. The law also assumes those opt-outs are findable, verifiable, timestamped, and content-specific.

Current options don't get there. Robots.txt blocks crawlers, but doesn't travel with content. EPUB metadata or ONIX flags identify titles, but don't survive when a file is chunked, reformatted, or translated. None of these systems carry an auditable record of who declared what, when.

What's missing is the evidence layer.

The Amlet approach

Amlet is building that evidence layer on top of an open standard called ISCC (ISO 24138:2024). ISCC produces a content-derived fingerprint — calculated from the work itself — that survives reformatting, chunking, and even translation. It enables similarity matching: a passage from a German translation of an Italian original can be matched back to the source title.

On top of ISCC, the registry runs a three-step process:

Register. Publishers fingerprint their catalog. The work itself stays with the publisher; only the fingerprint and metadata go on the registry.

Declare. Publishers set machine-readable usage rights using a shared vocabulary — train, inference, derive, search, analyze.

Discover. AI companies query the registry through an API to verify rights at scale.

A monetization layer sits on top of registration and declaration as an optional layer two. The foundation is the rights claim itself.

The order of operations matters

Without a registered, declared, and discoverable record, there is no defensible position from which to negotiate. Compensation discussions assume ownership and consent can be demonstrated — and demonstrated in a form an AI company is required to recognize.

The first move is the claim. Everything else follows.

→ Join the waiting list at amlet.ai