Summary: Two new families of metadata have arrived in the EPUB 3 package document: text and data mining reservation (TDMRep) and the accessibility block from EPUB Accessibility 1.1. The first decides whether your book can be used to train AI models. The second decides who can read it. Neither is a technical decision.
For much of Brazilian publishing, metadata is still the thing you do at the end. Someone fills in title, author and ISBN, picks a BISAC category more or less at random, exports the file and sends it to the distributor. Job done, next.
For many years that worked, and it worked for a simple reason: what read the metadata at the other end was a dumb catalogue. It only needed to know where to file the book.
That is no longer the case.
When a reader looks for a book today, less and less often do they type into a search box and work through a list of results. They ask. They ask ChatGPT, Perplexity, the browser assistant, the Google that answers before it lists links. And what those tools read to decide whether your book is the right answer is not the cover you approved after four rounds with the designer. It is the metadata.
Describe badly and you disappear. The disappearance is silent, and that is what makes it dangerous: nobody calls the publisher to complain that they could not find the book.
What is new in the package document
The package document is the file inside the EPUB where the metadata lives. It holds the book's identity in a form a machine can read. In recent years two new families of properties have arrived in that file. Neither came from committee whim. Both came from concrete outside pressure.
Reserving rights against mining
This is the one fewest Brazilian publishers know about, and the one they most need to know about now.
The protocol is called TDMRep, for Text and Data Mining Reservation Protocol. It emerged as a technical answer to Article 4 of the European Copyright in the Digital Single Market Directive. That directive opened an exception allowing the mining of protected content for commercial purposes, including training AI models, on one condition: the exception falls away if the rights holder has reserved that use in a machine-readable way.
Read that last part again, because the legal detail in it is heavy. In the European logic, silence is consent. If you have not declared a reservation, you are presumed to have reserved nothing.
And the European AI Act tied off the loose end by requiring providers of general-purpose models to have a policy of respecting those reservations. Which is to say the declaration in the metadata has left the symbolic realm. It has become the mechanism by which your refusal is, or is not, legally recognised.
The implementation in EPUB 3 is almost disappointingly simple:
<package prefix="tdm: http://www.w3.org/ns/tdmrep#" ...>
<metadata ...>
<dc:title>Title of the work</dc:title>
<meta property="tdm:reservation">1</meta>
<meta property="tdm:policy">https://suaeditora.com.br/politicas/tdm.json</meta>
</metadata>
</package>
tdm:reservation with a value of 1 says mining rights are reserved. With a value of 0, they are released. tdm:policy is optional and points to a file where you set out the conditions under which someone can negotiate a licence with you. Two lines of XML for a decision that covers your entire catalogue.
Here I need to add an honest caveat, and it matters. TDMRep is not part of the EPUB 3.3 specification. It is a W3C Community Group final report, and one of the techniques it defines is precisely how to apply the reservation inside an EPUB 3. That means two things. First: it is not required to validate the file, and EPUBCheck will not complain about its absence. Second: its effectiveness depends on the mining agent respecting the protocol, which is a bet, not a guarantee.
I cannot say how this mechanism will hold up when it is tested in a real court. Nobody knows yet. But the alternative is to declare nothing, and declaring nothing, on the European reading, is saying yes.
The accessibility metadata
In 2023 the W3C consolidated EPUB 3.3 as the family's first formal Recommendation, and EPUB Accessibility 1.1 came with it. It is that second specification that governs what has to be declared about who can read the book.
Three properties are required in any conformant publication. accessMode says which senses the content requires in order to be perceived: textual, visual, auditory, tactile. accessibilityFeature lists what was done to widen access: alternative text, long descriptions, semantic structure, page navigation. accessibilityHazard warns if anything in the content poses a risk, such as flashing or simulated motion.
Two others are recommended. accessibilitySummary, the summary in human language, and accessModeSufficient, which declares the minimum set of modes sufficient to take in the work without losing information.
The conformance declaration has also changed its appearance, for the better. Version 1.0 of the specification used a URL nobody could interpret at a glance. Now it is a sentence:
<meta property="dcterms:conformsTo" id="conf">
EPUB Accessibility 1.1 - WCAG 2.2 Level AA
</meta>
You can audit it by eye. You can put it in a contract. You can explain it to a client without an interpreter. There is also a block of properties for recording who carried out the assessment: a11y:certifiedBy, a11y:certifierCredential and a11y:certifierReport, the last of these pointing to the report, inside or outside the file.
It is worth separating a common confusion before it takes hold. There is European law behind the accessibility metadata too, but it is a different law: the European Accessibility Act, which since June 2025 has required ebooks sold in the European Union to be accessible. Two different directives are pushing two different families of metadata. Confusing them in a conversation with a client costs credibility.
How we do it at Booknando
Here this metadata is not an optional line in the scope. We apply both families by default in every book we produce: the complete accessibility block, with the conformance declaration in the Accessibility 1.1 format, and the TDM reservation in the package document.
When a client asks whether this adds to the cost, the answer is no, because it sits inside the production pipeline. The cost of good metadata was never the filling in. It is the decision to treat it as part of the book from the start, rather than as a form at the end.
I learned that the tedious way. When metadata is filled in on the last click before exporting, it becomes a copy of the previous title. I have seen accessibilityFeature listing long image descriptions in a book that did not have a single long description. That is worse than an empty field, because it plants a false promise inside data a machine will read as truth.
This was never a production matter
Here is the point I wanted to leave you with.
Metadata looks like an agenda item for the technical department, something for whoever touches the XML. It is nothing of the kind. Metadata is where three editorial decisions become processable data: who can find your book, who can read your book, and who may use your book to train a machine.
None of the three is a technical question. One is about marketing, another about accessibility, the third about copyright. The XML is merely where the answer is written down.
As long as a publisher sees digital as the PDF that became an EPUB, metadata will go on looking like paperwork. Once they see digital as a book that has to explain itself, unaided, to systems nobody controls, metadata becomes the part of the file that most decides the work's commercial fate.
My question to you is this: can your catalogue introduce itself, unaided, to an AI that will never open your book, or does it depend on having someone nearby to explain?
Further reading
- W3C. EPUB Accessibility 1.1. Available at: https://www.w3.org/TR/epub-a11y-11/
- W3C. EPUB 3.3. Available at: https://www.w3.org/TR/epub-33/
- W3C Community Group. TDM Reservation Protocol (TDMRep). Available at: https://www.w3.org/community/reports/tdmrep/CG-FINAL-tdmrep-20240510/
- W3C. Schema.org Accessibility Properties for Discoverability Vocabulary. Available at: https://www.w3.org/2021/a11y-discov-vocab/latest/