← Distillation From First Principles

Copyright and Training-Data Rights

When can copyrighted material become teaching data? The EU text-and-data-mining framework, Ireland's implementation, database and software rights, and a read-only audit of the E3 teaching-data lineage.

Chapter 36 asked what agreements permit. This chapter asks what the law permits when no agreement exists — because teaching data is where copyright enters the distillation pipeline, and the most expensive mistake in this territory is assuming that material which was lawful to read was therefore lawful to train on:

Which rights attach to the data and outputs used in a distillation pipeline, and which exceptions or permissions apply in the relevant jurisdiction?

The answer is built from four materials: the protected subject matter, the exception or licence asserted, the jurisdiction that decides, and the evidence that records all three. This chapter is educational, not legal advice. It states legal text with versions and dates, marks every interpretation as interpretation, and escalates the rest. An individual educational author is not assumed to be a research organisation; a TDM exception in one Member State is not assumed to hold in another; and a technical pipeline that logs everything is still not a legal conclusion.

Copyright protects original expression fixed in works — text, images, audio, audiovisual material, computer programs as literary works — together with related rights in performances, recordings, and broadcasts. It does not protect facts, ideas, or procedures, but the distinction is unreliable comfort for pipeline builders: training corpora reproduce expression at scale, and a dataset of facts about works can still infringe the database right examined below. Three separations carry the chapter.

First, lawful access is not permission to reproduce. Reading a public webpage, holding a subscription, or downloading a licensed dataset each establishes access; none establishes the right to copy the material into a training corpus. Access terms can also narrow what follows: a subscription that permits reading but forbids automated extraction, or an API whose terms prohibit model training on outputs, makes the access lawful and the mining unlawful at once.

Irish law sharpens this further — contractual provisions contrary to the TDM exceptions are unenforceable under Regulation 4(6) of S.I. No. 567 of 2021 — but unenforceability of a contract term is not a finding that the exception applies; the lawful-access and reservation checks still run first. Every TDM analysis in this chapter begins by recording the access basis and only then asks what reproduction right follows.

Second, exceptions are not licences. A statutory exception permits defined acts without permission under defined conditions; a licence permits what its terms say under its conditions. They fail differently: an exception fails when its conditions are unmet, a licence fails when its terms forbid the act. The E3 audit below rests on neither for third-party material, because E3 contains none — the finding is documented self-creation, not an exception claim.

Third, training, retention, and redistribution are different acts. An exception that covers reproduction for mining covers retention only as long as the statute allows, and redistribution not at all unless stated. Permission does not radiate across actions — the Chapter 36 rule, applied here to every dataset row.

The EU text-and-data-mining framework

Directive (EU) 2019/790 of 17 April 2019 (L-022; in force 7 June 2019) creates two mining exceptions with opposite shapes: one defined by who mines, one defined by what the rightholder did.

Article 3 permits reproductions and extractions by research organisations and cultural heritage institutions for scientific research, for material to which they have lawful access. Copies may be retained for research including verification of results. Rightholders and beneficiaries share security duties, and — decisively for pipeline design — any contractual provision contrary to the exception is unenforceable. The beneficiary class is the binding constraint: research organisations and cultural heritage institutions as defined, not individual authors, not startups, not this book’s programme standing alone. The chapter does not resolve whether any particular educational author qualifies; that is a specialist question with a jurisdiction attached, and it is recorded as one.

Article 4 permits reproduction and extraction for any purpose, including commercial mining, on two conditions: lawful access, and the absence of an express reservation by the rightholder in an appropriate manner. Article 4(3) defines appropriateness: for content made publicly available online, the reservation must be machine-readable — including metadata and website or service terms and conditions — and for other content it must be clearly communicated to all persons with lawful access.

Retention is allowed as long as necessary for the mining. The three-step test applies here too: the exception operates only in certain special cases, without conflicting with normal exploitation and without unreasonably prejudicing legitimate interests. The contract override does not, at Directive level: Article 7(1) makes contrary terms unenforceable for Articles 3, 5 and 6, not Article 4 — and Article 4(3) itself counts website terms and conditions as a valid reservation. Ireland goes further, as the next section shows.

The reservation rule is the operational heart of dataset work. A crawler that ignores robots.txt, metadata flags, or terms of service has not merely been impolite; where the reservation is in an appropriate manner, it has lost the exception. Conversely, a bare “all rights reserved” footer that is not machine-readable does not reserve online content under Article 4(3) — but relying on that reading without evidence of the actual reservation mechanism is exactly the kind of universal conclusion this chapter refuses. Record the reservation, its form, and its retrieval date, then decide.

Ireland’s implementation: S.I. No. 567 of 2021

The Irish copyright DSM transposing regulations, SI 567 of 2021 — the European Union (Copyright and Related Rights in the Digital Single Market) Regulations 2021 (S.I. No. 567 of 2021), made under section 3 of the European Communities Act 1972 with notice of making published in Iris Oifigiúil on 19 November 2021 (verified 16 November 2026 via the Irish Statute Book), transpose both exceptions with Irish particulars that matter for pipeline design.

Regulation 3 amends sections 53A and 225A of the Copyright and Related Rights Act 2000 (the Act of 2000, cited here through the Regulations; the consolidated Act text was not directly verified and any section-level reliance needs its own check). Article 3 copies must be stored securely with access limited to persons having lawful access — including IP validation or user authentication — under a named responsible person for network and database security. The author must be informed of the copying, may demand information on the security steps, and may request additional measures with a right to a prompt response. Retention covers research including verification. Contract terms to the contrary are unenforceable.

Regulation 4 inserts the Article 4 regime: section 53B for works, amendments to section 82 for computer programs, and section 225AA for performances. The reservation standard is transcribed faithfully — machine-readable for online content, clearly communicated otherwise — with retention as long as necessary and the same unenforceability of contrary contract terms. Two details deserve emphasis. First, computer programs get their own TDM provision (section 82 as amended): code corpora are mined under the program rules, not the general work rules, and any code-dataset pipeline must record which section it relies on. Second, press publications get parallel TDM exceptions (Regulations 14 and 15), so news corpora raise a second layer of rights beyond the author’s.

The implementation is IE law. Other Member States transpose the same Directive with national differences in procedure, beneficiary definitions, and enforcement practice; content hosted or mined elsewhere needs its own determination. The United States regime — fair use and its case law — is a different analysis entirely, mentioned here only to forbid silent import: no US finding in this chapter, no transatlantic inference from an EU exception. The contract’s US sources remain to-be-investigated, and the chapter leaves them there rather than filling the gap with assertion.

Databases, software, and the layers beneath the text

Directive 96/9/EC (L-026; in force 27 March 1996) creates the sui generis database right for substantial investment in obtaining, verifying, or presenting database contents. A scraped or compiled dataset can infringe three layers at once: copyright in the contents, the database right in the compilation, and contract in the terms of access. Exceptions exist but are narrow and jurisdiction-specific; the chapter asserts none for any real corpus and instead records the test: whose investment, what extraction, which exception, in which Member State.

Directive 2009/24/EC (L-027; in force 5 May 2009) protects computer programs as literary works and permits specified acts of observation, study, and testing. Those acts are not training licences: running a program to study it is permitted conduct; reproducing a million repositories to train on them is a different act requiring its own basis. Code datasets therefore need the section-82 TDM analysis above plus licence review of every contributing repository licence — permissive, copyleft, and proprietary terms do not blend, and the safe working assumption is that the mixture carries the strictest condition.

Teacher outputs and the resemblance problem

Outputs of a teacher model raise two distinct questions. The contract question — may outputs be used to train a student — was answered in Chapter 36 (for the book’s self-created teachers: yes, under the recorded actions; for API teachers under provider terms: frequently no). The copyright question is separate: does a generated output reproduce protected expression, and does training on it propagate that reproduction?

For the book’s own teachers the answer is documented rather than assumed: teachers trained exclusively on synthetic or self-created material generate outputs whose lineage terminates in repository code, and the fingerprints are in the manifests. For third-party teachers, two risks travel with the outputs. First, memorised reproduction — outputs that restate training material — makes the student’s corpus a copy of a copy, with the original rightholder still at the start of the chain. Second, derivative treatment by contract: Gemma Terms (L-004) expressly define distillation via intermediate representations or synthetic outputs as a Model Derivative, so the outputs arrive pre-encumbered regardless of their copyright status.

A generated output that resembles a protected work must be treated as the work for lineage purposes until shown otherwise; resemblance is a flag, not a verdict, and the flag routes to specialist review, never to silent inclusion. Style as such — a genre, a manner, a statistical texture — is not the protected subject matter; particular expression is. The pipeline cannot reliably tell the two apart, which is why the rule is conservative: resemblance flags the row, the reviewer decides the row, and the resemblance test is documented with the output preserved, not with the output deleted to hide the question.

Transformations, synthetic data, and mixed origins

Transformations do not launder rights. Filtering, reformatting, tokenising, or paraphrasing a corpus changes its bytes, not its provenance: every row retains the source asset, the basis, and the reservation status through the transformation, and the TeachingDataRights record carries the transformation as a field rather than a fresh start.

Synthetic data generated from protected sources inherits the source position — a paraphrase of a reserved article is still downstream of the reservation until a proper analysis says otherwise. Mixed-origin derivatives are treated as carrying the strictest condition of any contributor (Chapter 36 Case D), with row-level provenance or refusal to mix. The pipeline rule is mechanical: provenance flows forward through every transformation, never resets, and any gap blocks the rows it touches.

Retention, revocation, and the limits of technical compliance

Article 3 retention serves research including verification, under the security duties above; Article 4 retention lasts as long as necessary for the mining. Both end. A rightholder who reserves rights after mining began, a licence that expires, or a dataset withdrawn upstream each trigger the same downstream question the lineage graph answers in Chapter 39: which artifacts depend on the retired source, and what happens to each.

Concretely: the retired source’s dependents — datasets derived from it, students trained on those datasets, checkpoints, evaluation reports citing them — are enumerated from the graph, each re-evaluated under the new position, and either re-authorised on a fresh basis or retired with the retirement recorded. Nothing in the graph performs the re-evaluation; it performs the enumeration without which re-evaluation is guesswork. Technical compliance records — hashes, fingerprints, audit trails — prove the chain is intact; they do not decide whether the chain was lawful. That decision belongs to the reviewer with the documents, and the pipeline’s job is to make the review possible, fast, and complete.

The E3 teaching-data audit

The full record is planning/37-e3-rights-audit.md; the method is read-only manifest inspection plus the tested tdm_checklist. All ten E3 studies (RM-104 through RM-124) record VERIFIED_PERMITTED for their training-adjacent actions. The corpus families — reverse-complement symbols, chain arithmetic, structured JSON sums — are generated in-repository from recorded templates, tokenizers, and seed offsets, with per-seed source fingerprints in the structured, transformer, and preference manifests. Teachers are self-created models from the same runs — the E3-sequence-precursor teacher behind RM-104, the chain-arithmetic teacher of the E3R-20261115 run, the structured teacher of the E3S-20261116 run with recorded validation functional scores.

No third-party text, image, audio, code, or database content appears anywhere in E3; Articles 3–4 are therefore not engaged by the E3 teaching data, recorded as a positive finding with evidence, not as an assumption. Storage and retention follow the manifests; checkpoints are git-ignored. Redistribution and commercial use are blocked with no record (Chapter 36 rows 8–10). Per-seed fingerprints are absent in six E3 manifests — recorded as missing-link findings carried to Chapter 39, not repaired by re-opening frozen studies.

Worked examples

A. Self-created teaching data. The E3 JSON-sum corpus: repository generator, recorded template and tokenizer, fingerprinted seeds, self-created teacher, training covered by RM-117. Each of the sixteen rights fields is populated; the TDM checklist returns “proceed within recorded conditions” because no third-party reproduction occurs. Independently authorised: generation, filtering, training, retention as study artifacts.

B. Publicly accessible material with a TDM reservation. A web corpus whose host serves a machine-readable opt-out in page metadata. Lawful access is present; Article 3 is unavailable (no beneficiary status asserted, and none assumed); Article 4 is defeated by the appropriate reservation. Disposition: blocked for training. The tested control pins this: a reservation without machine-readable form still blocks in the checklist’s conservative reading, with appropriateness recorded for the reviewer.

C. Material under a restrictive licence. A code corpus mixing permissive, copyleft, and proprietary files. The section-82 TDM analysis runs per file, then the licences run over the whole: the mixture is treated as carrying the strictest condition, copyleft may, on some readings, require source disclosure of the derivative pipeline (whether a share-alike licence reaches trained weights, adapters, and training code is unsettled and goes to review), and proprietary files are excluded or separately licensed. Only the permissive subset proceeds, with its file list recorded. The share-alike case is the chapter’s warning against partial compliance: training on ninety clean files and ten copyleft ones without disclosure is not ninety percent compliant.

D. Synthetic outputs from a contracted API teacher. Outputs generated under terms prohibiting competing-model development (Chapter 36 Case C pattern). The TDM framework does not override the contract for this act: training a student is outside every asserted exception and inside an express prohibition. Disposition: VERIFIED_PROHIBITED in shape; do not begin.

E. Mixed-source derivatives with incomplete provenance. Rows of unknown origin blended into a training file. No basis is assertable for the unknown rows, the mixture rule contaminates the file, and the missing fingerprints block the batch. Disposition: quarantine the unknown rows or drop the file; unknown is never promoted into permission.

Implementation

TeachingDataRights in examples/common/provenance.py carries the sixteen fields (asset, type, origin, classification, jurisdiction, acquisition, access basis, TDM basis, reservation plus evidence and retrieval date, permitted action, retention and redistribution conditions, transformation, derived artifact, decision, reviewer, open conditions). The field set follows the dataset-datasheet and model-card documentation practice (Gebru et al.; Mitchell et al.). validate_teaching_rights enforces completeness and honesty — unresolved records require open conditions; permitted-without-basis is rejected. tdm_checklist encodes the ordered checks: lawful access, Article 3 beneficiary status (never assumed for individuals), Article 4 reservation with appropriateness, retention scope, jurisdiction (IE decided, others flagged). Seven focused tests pin the validation, both checklist directions, and a read-only spot check of the E3-structured manifest.

    flowchart TD
  L[lawful access?] -->|no| B[blocked: no basis]
  L -->|yes| A3[Article 3 beneficiary? never assumed]
  A3 -->|yes| K[check retention + security duties]
  A3 -->|no| A4[Article 4 reservation? check form + date]
  A4 -->|reserved| B
  A4 -->|not reserved| K
  K --> J[jurisdiction decided? IE or flagged]
  J --> R[record + reviewer; escalate ambiguity]
  

Figure 1. The ordered TDM checklist the code enforces: access first, exceptions only on their conditions, jurisdiction always stated. The chart is the procedure, not a legal conclusion.

The teaching-data rights matrix

The contract requires a rights matrix for the teaching data. Rows are sources, columns are the questions that decide each row:

Source Subject matter Access basis Exception / licence Reservation Retention Redistribution Decision
E3 JSON-sum corpus self-created rows created locally none needed none study artifacts blocked, no record PERMITTED to train
E3 chain-arithmetic traces self-created rows created locally none needed none study artifacts blocked, no record PERMITTED to train
Web corpus, machine-readable opt-out third-party text; possible database right public HTTP Art 4 asserted appropriate; defeats Art 4 n/a n/a blocked
Share-alike code corpus third-party programs (s82) licence grant licence + s82 TDM per-file check per licence share-alike binds derivatives permitted subset only
API-teacher synthetic outputs outputs; contract-encumbered contract none available n/a (prohibition, not reservation) n/a derivative terms attach PROHIBITED pattern
Mixed-origin file, unknown rows mixed, partly unknown partial none for unknown rows unknown n/a strictest condition governs blocked pending provenance

Each row maps to a TeachingDataRights record; the matrix is the human-readable index, the records are the evidence. The E3 rows are findings; the rest are reusable patterns with their escalations attached.

Decision framework and escalation

For each source: identify it with version; state the action and jurisdiction; locate the primary document with effective and retrieval dates; quote the operative clause; check lawful access, then Article 3, then Article 4 with the reservation evidence; record retention and redistribution separately; name the reviewer. Escalate to a qualified lawyer: any Article 3 beneficiary doubt, any reservation-form dispute, any database-right question, any non-IE determination, any API-output or mixed-origin case, any revocation or licence change, and every redistribution or commercial-use decision. The repository’s Apache-2.0 licence covers its code; it grants no rights in third-party material.

Limitations

IE determinations with one verified SI; the consolidated Act of 2000 text not directly verified. No case law examined. US and other non-EU regimes undescribed by design. TDM positions for real third-party corpora untested — the book holds none. Register entries used as dated; provider terms and model cards change. No independent or legal review; this chapter is not legal advice.

Preparing for privacy and regulation

Copyright asks who may copy; privacy asks who may be identified, and regulation asks who answers for the system. The pipeline built here — sixteen-field records, ordered TDM checks, provenance that survives transformation — is the machinery Chapter 38 reuses for personal data, with different fields and stricter duties, the E4 trajectories under audit and the AI Act’s consolidated timetable on the table, including the question this chapter leaves open: what happens to lawfully mined data when the data subjects inside it exercise their rights.

References

Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H. and Crawford, K. (2021). Datasheets for Datasets. Communications of the ACM 64(12). arXiv:1803.09010. https://arxiv.org/abs/1803.09010

Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D. and Gebru, T. (2019). Model Cards for Model Reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT 2019).* https://doi.org/10.1145/3287560.3287596

Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market. In force 7 June 2019. https://eur-lex.europa.eu/eli/dir/2019/790/oj (retrieved 12 November 2026). Cited as L-022 for the Article 3 and Article 4 text-and-data-mining exceptions including the Article 4(3) rights reservation; not legal advice.

Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases. In force 27 March 1996. https://eur-lex.europa.eu/eli/dir/96/9/oj (retrieved 12 November 2026). Cited as L-026 for the sui generis database right; not legal advice.

Directive 2009/24/EC of the European Parliament and of the Council of 23 April 2009 on the legal protection of computer programs. In force 5 May 2009. https://eur-lex.europa.eu/eli/dir/2009/24/oj (retrieved 12 November 2026). Cited as L-027 for program protection and the observation-study-testing acts; not legal advice.

Irish copyright DSM transposing regulations, SI 567 of 2021. Made under section 3 of the European Communities Act 1972; notice of making published in Iris Oifigiúil on 19 November 2021. https://www.irishstatutebook.ie/eli/2021/si/567/made/en/print (verified 16 November 2026). Cited for Regulations 3, 4, 14, and 15 transposing the Article 3 and Article 4 exceptions, the machine-readable reservation standard, retention rules, security duties, and contract-override provisions; not legal advice.

Gemma Terms of Use. Last modified 1 April 2026. https://ai.google.dev/gemma/terms (retrieved 12 November 2026). Cited as L-004 for the Model Derivative treatment of distillation via intermediate representations or synthetic outputs; not legal advice.