Affero Press — Text and Data Mining Policy
Rights in works published by Affero Press are reserved from text and data mining, and from use in training, fine-tuning, or grounding artificial-intelligence or machine-learning systems for the purpose of building or improving those systems, under Article 4(3) of Directive (EU) 2019/790.
This does not restrict retrieval for a reader. An agent acting on behalf of someone who holds a lawful copy may fetch, read, quote, and summarise a work in order to answer that person. What is reserved is the use of these works as material for building or improving a system — not the act of reading one on a reader's behalf.
We make that distinction here, in words, because TDMRep cannot express it. TDMRep reserves or does not reserve; it has no term for this kind of use but not that one. The specification anticipates the gap and provides for it: a policy may point to a page that explains the publisher's position in human language. This is that page.
One vocabulary now expresses part of it. RSL 1.0
separates ai-train — training or fine-tuning a model — from ai-input,
which is retrieval, grounding, and generative answers. Our RSL licence
permits ai-input, ai-index and search, and by leaving
ai-train off that list prohibits it: under RSL a list of permitted uses is
exhaustive.
It remains an approximation, and we would rather say so than let a machine-readable file be
read as more exact than it is. What we reserve is grounding for the purpose of building or
improving a system, and neither TDMRep's binary field nor RSL's usage tokens carry a
purpose. RSL's ai-input is therefore broader than what we permit in words. Where
the two differ, this page governs.
To be precise about what is missing, because the distinction matters: the mechanism
exists. ODRL 2.2, a W3C
Recommendation, defines purpose as a standard left operand — "a defined purpose
for exercising the action of the Rule" — and TDMRep §7.1.5.4 already uses it. What is
missing is a value: no agreed identifier means for the purpose of building
or improving an AI system. TDMRep defines two, tdm:research and
tdm:non-research, and its own text calls them experimental and likely to be
replaced. Neither fits.
We could mint an identifier of our own. It would be standards-conformant and no one would recognise it, because the only software that understood it would be ours. That is how every decentralised vocabulary begins, and we would rather wait for one that consumers actually read than publish a term that looks precise and communicates nothing. Until then, the sentence above does the work, in words.
Permission is not refused outright — it is gated. Requests may be addressed to rights@affero.it. Tell us what you want to use and what for; we would rather license than litigate.
ai-input,
ai-index and search permitted, ai-train notLicense: directive points
a crawler at that licencePublished EPUB files carry the same reservation in their own metadata, so it travels with the file when the file travels away from this domain.