Training AI on Copyrighted Books: Fair Use, Data Provenance, and a 90-Day Playbook

Is it legal to train AI models on copyrighted books? It’s complicated

Why this matters now: an AI product that summarizes, answers questions about, or drafts text based on books can be enormously valuable, and it can also step into the market authors and publishers serve. That tension is showing up in courtrooms, the U.S. Copyright Office, and commercial negotiations. For executives and product leaders the question is not an abstract legal puzzle. It’s a business risk you must design around today.

The legal framework, short and practical

The starting point is the Copyright Act and the statutory fair use defense (17 U.S.C. § 107). Fair use is a four-factor, fact-heavy balancing test. Recent cases have sharpened two issues courts press most often:

Transformative purpose: Does the new use add a different purpose, meaning or message (and not simply republish or repurpose the original)?
Market effect: Does the new use substitute for or harm the market for the original work or for its licensed derivatives?

Those two questions are front and center, but judges still consider the full four-factor test. Other concepts you’ll see discussed are:

Fair use: a statutory defense that permits limited uses of copyrighted material without permission under certain circumstances (17 U.S.C. § 107).
Transformative: a legal term meaning the use has a new purpose or character, not merely a reformatting or republishing of the original.
Intermediate copying: the technical copying that happens inside model training (embeddings, feature extraction). Courts differ on whether that internal copying counts as a “use” for copyright purposes.
Human authorship: the requirement the U.S. Copyright Office applies when deciding whether a work is eligible for registration; the Office currently refuses registration for works that are wholly the product of autonomous AI with no sufficient human creative input (U.S. Copyright Office report, Jan. 30, 2024).

Where courts have drawn lines so far

These outcomes are fact-specific, and the cases vary. A few court and administrative developments are important anchors:

Ross Intelligence v. Thomson Reuters (D. Del., cited at 2025 WL 458520): Judge Stephanos Bibas rejected Ross’s fair use defense after finding that the use of Thomson Reuters’ editorial material (headnotes and annotations) was not transformative and that the product competed with the publisher’s business. As the opinion put it, “Ross’s use is not transformative because it does not have a ‘further purpose or different character’ than Thomson Reuters’s.” That decision signals close judicial scrutiny when AI products use proprietary, curated content to build a competing research tool.
U.S. Copyright Office position on AI authorship (USCO report, Jan. 30, 2024): The Office has taken the position that works produced entirely by autonomous AI without sufficient human creative input are not registrable. Several lower-court practices and administrative refusals align with that approach; appellate outcomes are still developing (for example, the DC Circuit heard argument in Thaler v. Perlmutter on Sept. 19, 2024, a case that touches on authorship questions).

Press coverage has also reported large disputes and settlements involving book content and AI training datasets. Reporting and filings often differ, so read the actual court orders and settlement papers before treating a headline as dispositive. These matters make one practical reality clear: courts and plaintiffs will scrutinize how training data were acquired.

What judges and lawyers are zeroing in on, translated into business implications

Data provenance matters. If training material was taken from unauthorized repositories, plaintiffs and courts will treat that as a significant strike against a fair use defense. Business implication: keep acquisition records and favor licensing.
Intermediate copying is legally sticky. Models must internally represent text to learn from it. Some courts treat that internal copying as permissible learning, others focus on what the model produces and whether that output infringes. Business implication: be prepared to explain and, if necessary, demonstrate technical safeguards, filtering, and non-replicative behaviors.
Products that substitute for a publisher’s offering face higher risk. If your model reproduces editorialized content, like headnotes or summaries that replicate a publisher’s value, courts will view that as market substitution. Business implication: avoid product features that function as drop-in replacements for proprietary services unless you’ve licensed the content.
Human authorship affects ownership and enforcement. Outputs that are wholly autonomous AI creations may not qualify for copyright protection, which alters commercial bargaining power and remedies. Business implication: design workflows that preserve clear human creative contribution where copyright matters commercially.

Concrete steps for teams building or buying AI

Do not treat legal uncertainty as a reason to wait. Treat it as a design constraint. Here are practical, prioritized actions:

30-day: run a provenance audit (legal + engineering). Inventory all textual datasets used for training and fine-tuning. For each dataset, record source, license or access terms, acquisition receipts, and an initial risk score (low/medium/high). This is your baseline for negotiating and defending practices.
60-90 days: legal risk assessment and licensing review. For datasets rated medium/high risk or for features that could substitute for existing markets, engage counsel to map licensing options: direct licenses, collective licensing schemes, revenue-sharing, or feature redesigns to avoid infringement risk.
Design and product checklist (ongoing). Build guardrails that reduce direct market substitution. Prioritize models trained on licensed or public-domain corpora for market-facing features. Require human-in-the-loop steps where copyrightable outputs are claimed. Add provenance metadata and attribution where feasible.
Operationalize documentation and retention. Keep an auditable trail for dataset acquisition and curation decisions. Log model training runs, versions, and datasets used for each deployed model to support compliance and responses to claims.

Industry examples that clarify risk (what to avoid)

Use these as heuristics rather than bright-line rules:

If your product uses proprietary editorialized content (publisher headnotes, paywalled summaries) to deliver the same function a publisher offers, expect strong legal challenge (see Ross Intelligence).
If your product summarizes public-domain books or generates novel content inspired by licensed works while adding new purpose and expression, the risk is lower, but it still depends on acquisition provenance and whether the output is too similar to the source.
If a model’s output is claimed as a copyrighted work but it was produced purely by AI with no meaningful human creative input, the U.S. Copyright Office’s current stance is that the work will not be registrable (USCO report, Jan. 30, 2024).

Top 3 recommended actions for executives (quick)

Within 30 days: Complete a provenance audit and assign a dataset risk owner (legal + data engineering).
Within 90 days: Build a licensing plan for any datasets rated medium or high risk; prioritize rewrite or redesign of features that could substitute protected products.
Ongoing: Make human contribution and attribution part of product design where copyright ownership matters; document everything.

Key takeaways, questions you should be asking (and short answers)

Is it categorically illegal to train AI on copyrighted books?
No. Courts have reached mixed results. Outcome depends on how the books were obtained, how the model uses what it learned, whether outputs supplant the original market, and the totality of the four fair-use factors (17 U.S.C. § 107).
Does it matter if the text came from an unauthorized “shadow library”?
Yes. Unlawful acquisition increases liability risk and will weigh heavily against a fair use defense in litigation. Maintain acquisition records and avoid dubious sources.
What does “transformative” mean here?
It means the AI’s use adds a different purpose, meaning, or message, not just reformatting or repackaging the original. Courts follow Supreme Court guidance (see Andy Warhol Found. v. Goldsmith, 598 U.S. 508 (2023)) in applying that test.
Are AI-only outputs copyrightable?
The U.S. Copyright Office currently says works created entirely by autonomous AI without sufficient human authorship are not registrable (USCO report, Jan. 30, 2024); lower-court practice follows this in some contexts, but appellate law is still evolving.
If you could do only one thing in the next 30 days, what should it be?
Run a provenance audit of your text datasets and flag any sources that lack clear, lawful acquisition evidence. That single step materially reduces downstream risk and gives you leverage in licensing talks.

Courts and agencies are sending signals but not a single, nationwide rule. That patchwork will matter for business strategy. Expect litigation to shape incentives for licensing, dataset curation, and product design long before Congress or the Supreme Court steps in. Treat legal uncertainty as a product constraint you can manage with provenance, licensing, and design choices that keep your offering on the right side of both law and market relationships.

Notable judicial line: As Judge Stephanos Bibas wrote in the Ross decision, “Ross’s use is not transformative because it does not have a ‘further purpose or different character’ than Thomson Reuters’s.” (Ross decision, cited at 2025 WL 458520.)