Copyright in the Age of AI

Chapter 8 · Copyright in the Age of AI

imgai1-9 named training-data copyright and style mimicry as real issues without covering the legal mechanics underneath either one. This chapter provides that depth — and immediately splits what's often treated as one single question into three genuinely separate legal questions, each with its own, differently unsettled answer.

Three Separate Questions, Often Conflated Into One

Question	What it actually asks
A — The input question	Does training an AI model on copyrighted works itself infringe copyright?
B — The output-ownership question	Can AI-generated output itself be copyrighted?
C — The output-infringement question	Can AI-generated output itself infringe an existing copyrighted work?

These have different answers, different levels of legal settledness, and different reasoning behind each one — treating "AI and copyright" as a single question is itself a source of real confusion.

Question A — Is Training on Copyrighted Works Infringement?

This is genuinely, actively unsettled, and is the subject of multiple real, ongoing lawsuits at the time of writing. The core tension maps directly onto Chapter 4's own four-factor test:

Argument for fair use	Argument against
Training extracts statistical patterns rather than storing retrievable copies — arguably a genuinely new, transformative purpose (Factor 1) distinct from the works' own original purpose	Training typically uses the entirety of each work (Factor 3), often for commercial purposes, and can plausibly compete with the market for the original works or their licensing (Factor 4)

A real, notable data point — not a final answer

In 2025, a U.S. federal court ruled in Thomson Reuters v. Ross Intelligence that using Westlaw's copyrighted legal headnotes to train a directly competing legal-research AI was not fair use — one of the first substantive rulings to directly address AI training data. This involved a fairly specific fact pattern (training a direct market competitor on the plaintiff's own proprietary content), and shouldn't be read as settling the much broader question of AI training generally, especially against the backdrop of several other major cases (including disputes involving news publishers and stock-image libraries) still working through the courts.

Question B — Can AI-Generated Output Be Copyrighted?

Per current U.S. Copyright Office guidance, and confirmed by the courts in a case known as Thaler v. Perlmutter (which upheld the Office's refusal to register a fully AI-generated image with no claimed human author), the answer under existing U.S. law is: not without sufficient human creative authorship.

This connects directly back to Chapter 1

U.S. copyright law has long required a human author — a requirement that predates AI entirely but now sits squarely at the center of this question. Content generated purely by an AI system, with no meaningful human creative contribution, doesn't meet that requirement and isn't copyrightable at all — leaving it, functionally, without an owner in the way a human-authored work automatically has one per Chapter 1's own automatic-copyright rule.

This isn't absolute, though: where a human contributes sufficient, genuinely creative selection, arrangement, or editing on top of AI-generated raw material, that specific human contribution can itself be copyrightable — the Copyright Office has drawn a real, if genuinely fuzzy, distinction between merely typing a prompt (generally not enough on its own) and substantially selecting, arranging, or modifying AI output in a way that reflects real human creative judgment.

Question C — Can AI Output Itself Infringe an Existing Work?

Yes, in principle — and this is where Chapter 3's ordinary infringement analysis applies regardless of how the allegedly infringing material was actually produced. If a specific AI-generated output is substantially similar to a specific existing copyrighted work, the same analysis from Chapter 3 can apply exactly as if a human had copied it directly.

Documented, real near-reproductions

Some AI image generators have been documented producing outputs strikingly, sometimes near-exactly, similar to specific images that appeared in their own training data — undercutting the "it only learns statistical patterns, never reproduces anything specific" defense sometimes offered for Question A, at least for those specific documented cases.

Style mimicry is legally different from reproducing a specific work

Per Chapter 1's own idea/expression dichotomy, a general artistic style functions much more like an unprotectable idea or method than protectable expression — nobody owns "painting in a loose impressionistic style" the way they own a specific painting. An AI generating new images "in the style of" a specific living artist, without reproducing any specific protected expression from any of that artist's actual works, sits on genuinely uncertain legal ground under a straightforward Question C analysis — a real, serious ethical concern (exactly as imgai1-9 named it) that doesn't automatically translate into a strong copyright infringement claim the way reproducing a specific existing work more directly would.

The Honest Bottom Line

This is a live, moving target

AI and copyright is currently one of the most actively litigated, least settled areas of copyright law. Court rulings are arriving, but often address narrow fact patterns rather than resolving the broader questions definitively, and the law in this area will likely keep developing significantly. Treat everything in this chapter as a snapshot of where things currently stand, not a permanent, final answer — the same case-by-case caution Chapter 4 applied to fair use generally applies here with even less settled ground underneath it.

Hands-On Exercises

Exercise 1

Explain why treating "AI and copyright" as one single legal question is a mistake, using this chapter's own three-question breakdown.

📄 View solution

Exercise 2

Explain why a fully AI-generated image with no human creative input isn't copyrightable under current U.S. law, connecting your answer back to Chapter 1's own material.

📄 View solution

Exercise 3

Explain why "generating images in a living artist's style" is a real ethical concern per imgai1-9 but a genuinely uncertain copyright infringement claim, using this chapter's own idea/expression reasoning.

📄 View solution

Chapter 8 Quick Reference

Three separate legal questions: (A) is training infringement, (B) can AI output be copyrighted, (C) can AI output infringe an existing work
Question A is genuinely unsettled — a real Factor 1 (transformative) vs. Factor 3/4 (full use, market effect) tension, with Thomson Reuters v. Ross Intelligence (2025) as a real but narrow data point
Question B — under current U.S. law, AI output without sufficient human creative authorship isn't copyrightable at all (Thaler v. Perlmutter); meaningful human selection/arrangement/editing can change this
Question C — ordinary infringement analysis applies regardless of how output was produced; documented near-exact reproductions of training images have occurred
Style mimicry is a real ethical concern but legally different from reproducing a specific work — style itself is generally not protectable expression
This entire area is actively litigated and unsettled — treat this chapter as a snapshot, not a final answer
Next chapter: Web Scraping, Data Collection & Copyright