Licensing legal content for AI: case law, headnotes and secondary sources

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Which legal materials are public domain and which carry rights

Judicial opinions and enacted law sit in the public domain because officials acting with the force of law cannot be "authors" for copyright purposes. That is the government edicts doctrine. In Georgia v. Public.Resource.Org (2020), the Supreme Court extended it to annotations in the Official Code of Georgia Annotated. Those annotations were prepared for the state legislature's code commission by a commercial publisher under a work-for-hire arrangement [1]. Federal government works are separately excluded from copyright by 17 U.S.C. 105 [12].

The doctrine has limits. It turns on who authored the material and in what capacity [1]. A private publisher's own headnotes, summaries and classification systems, written on its own account, fall outside it. A treatise author's chapters belong to the author or, under a work-for-hire or assignment, to the publisher [10]. Court filings drafted by private lawyers are a third category: they are public records, but the government edicts doctrine does not cover them.

Three facts matter for any legal corpus you build:

• Who wrote each field. Opinion text by a judge differs from a headnote by a publisher's attorney-editor, even when both appear in the same record.
• Where you got it. Public-domain text copied out of a subscription database can still breach that database's terms of service, and acquisition method is a growing focus of copyright risk (lawful access and pirated sources).
• What the model does with it. RAG retrieval that displays text raises different questions from SFT on opinion-to-summary pairs or from pre-training.

What Thomson Reuters v. Ross means for headnotes

Publisher headnotes can be copyrightable editorial work, and using them to build a competing legal research tool was not fair use in the Third Circuit as of October 2026. The court affirmed the District of Delaware's partial summary judgment for Thomson Reuters on interlocutory appeal. It held that the Westlaw headnotes at issue were original enough to protect [2]. ROSS had obtained headnote-derived training material, so-called "Bulk Memos," through a third-party vendor, and the tool it built was non-generative.

Read the holding narrowly but take it seriously. Commentators note the opinion distinguishes generative systems, so it does not decide how a court would treat an LLM trained on headnotes. Do not predict that question either way. The practical rule is plain: if a training record contains a publisher's headnote, synopsis, key number or topic classification, treat it as licensed content. That holds even when the opinion text beside it is public domain.

Common contamination paths in legal datasets:

• Scraped "free" case law pages that embed a publisher's syllabus or headnotes above the opinion.
• Vendor-labeled question-answer pairs written by annotators who copied headnote language from a subscription product.
• Citator signals (negative treatment flags, depth-of-treatment markers) exported from a research platform.

Where to obtain public-domain case law lawfully

Open bulk sources now cover most US precedential case law, so you rarely need a publisher license for the opinion text itself. The Harvard Caselaw Access Project digitized historical US reporters. Its access limits, which came from its original agreement with Ravel Law, expired in March 2024. The Library Innovation Lab says the data can now be released without restriction on access or use [3]. Harvard continues to host the bulk data at case.law and points search and API users to CourtListener [3].

CourtListener, run by Free Law Project, includes the CAP cases plus opinions collected directly from court websites, and offers bulk downloads [4]. Its coverage page states what it holds by court and period. Check that page against your jurisdiction list before assuming completeness [4]. For pre-training scale, research corpora such as Pile of Law (about 256GB from 35 sources, including opinions, filings, agency materials and casebooks) [5] and MultiLegalPile [6] show what is assembled from open sources. Each component source in those corpora carries its own terms, so audit them source by source. Do not rely on the corpus-level description. For general open-license screening, see the open data license compatibility matrix.

Two gaps remain even with open sources. First, slip opinions and unpublished decisions vary by court in format and availability. Second, docket entries, briefs and transcripts sit behind PACER and state e-filing systems with fees and terms of use. Those filings can also contain personal data that courts' redaction rules did not catch.

Licensing secondary sources, practice guides and annotations

In-copyright secondary sources are authored expression, so plan to license them with terms that name AI use. Treatises, practice guides, form books, law review articles, CLE materials and jury instruction commentary are authored works. Rights may sit with the author, the publisher or a bar association, depending on contract [10]. A publisher that markets a treatise may not hold the AI training rights for every edition or contributor.

Ask each licensor to show the chain of title for the titles in scope, not just a catalog list. Treat non-official state code annotations as licensed content. Georgia turned on annotations prepared at the legislature's direction; a publisher's own unofficial annotated code is a different case [1]. The licensing questions that recur for legal publishers:

• Use scope. Does the grant cover pre-training, SFT, evaluation, and RAG display, each named separately? See the AI training rights grant clause.
• Output display. Can your product quote or summarize licensed text to end users, and with what attribution?
• Updates. Legal content goes stale with each pocket part and supplement. Is an update feed part of the deal?
• Termination. What happens to fine-tuned weights and vector indexes if the license ends? See model retention after license termination.
• Competition. Publishers may restrict use in products that compete with their research platforms, and a competing research product was central to Ross [2].

For retrieval-specific terms (caching, chunk display limits, index deletion), the RAG content licensing guide covers the mechanics.

Legal content rights matrix

Use this matrix to tag every source and field before ingestion; the tag decides whether you need a license, a terms-of-use review, or neither.

Illustrative example: invented to show structure; it does not describe an available dataset.

A record-level manifest for a mixed corpus should carry at least these fields: source_id, content_layer, author_type, acquisition_channel (bulk download, API, license feed), license_ref, terms_url_snapshot, stripped_fields (for example headnotes,synopsis), pii_screen_method, and allowed_uses. A manifest like this also feeds disclosure work: providers of general-purpose AI models placed on the EU market must keep a copyright policy [7][11] and publish a training-content summary using the Commission's July 2025 template [8].

Operational legal data: the layer publishers do not have

Publishers license what they publish, but much of what makes legal AI useful never appears in a reporter or a treatise. Contract redlines, clause negotiation histories, matter intake notes, review comments and billing records live inside companies and law departments. These records teach drafting and negotiation behavior that opinions cannot. They carry confidentiality and privilege questions rather than publisher copyright.

SourceX sources this kind of operational data from US companies on request, including legal workflows and documents, and manages licensing agreements and ongoing purchases. Every dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Names, emails, phone numbers and account numbers are removed or replaced before delivery; the method is recorded, a sample is checked, and no method is perfect. Every release is approved by the supplying company, and a request does not guarantee a match. You can describe the legal records you need to SourceX, or see legal AI training data from real contract work, contracts and legal documents and contract redline datasets.

SourceX does not source scraped web content, so it is not a channel for publisher headnotes or scraped case law. Use the open bulk sources above for primary law and license secondary sources directly from their rights holders.

Due diligence before you train on any legal corpus

The most expensive failure in legal AI data is discovering protected editorial content inside a model after release. The Bartz v. Anthropic class settlement received final approval in July 2026. A settlement is not a merits ruling [9]. Run these checks before training:

1. Diff a sample of opinion records against the official court slip opinion and flag any extra text blocks above or below the opinion body.
2. Search for publisher-specific markers: key number patterns, "Synopsis" and "Headnote" labels, citator symbols and copyright notices.
3. Confirm each vendor's acquisition channel in writing, including any sub-vendors, since Ross involved material obtained through a third party [2].
4. Screen filings and transcripts for names, Social Security numbers, minors' identities and sealed-matter content.
5. Map each source to the use you intend; a fine-tuning-only license does not cover pre-training or RAG display.

Building a legal AI corpus with licensed operational data

Pair open primary law with operational records licensed from the companies that hold them. SourceX sources legal workflow and document data from US companies on request and serves AI teams wherever they are based. Nothing is contracted until a supplier agrees, and terms are agreed per deal. Describe the legal data your AI needs at SourceX for buyers.

Related reading: the AI training data licensing hub and licensed text corpora for LLM pre-training.