Jon Ippolito

Wendy Seltzer's Fair Use in the Age of Instagram (2017) is a good foundation on how fair use works in the 21st century, but AI has changed the landscape and the law is still catching up. All of the rulings below apply only to US law.

Who's winning?

Detailed verdicts

1. "Transformative" got narrower

In Warhol v. Goldsmith (2023), the Supreme Court held that whether a use is "transformative" depends less on whether the new work adds new meaning and more on whether it serves the same commercial purpose as the original. Warhol's silkscreen of Goldsmith's Prince photo wasn't fair use because both were featured in magazines as portraits of the same musician.

It's not an AI case, but it shows adding style isn't enough; ask whether your use competes in the same market with the original. (Creative Commons' read on what this means for remixers.)

2. The AI-training fights produced their first real answers

• Bartz v. Anthropic: a federal judge ruled in 2025 that training an AI model on lawfully acquired books can be fair use, but that downloading and storing pirated copies is not, even if the end use is transformative. Anthropic settled for $1.5 billion, the largest copyright settlement in U.S. history (480,000 works at about $3k each).
• Thomson Reuters v. Ross: a court found that training an AI legal-research tool on Westlaw's headnotes was not fair use, because the tool competed directly with Westlaw. Same lesson as Warhol: competing with your source is the danger zone.
• The docket keeps growing — well over a hundred active suits by 2026, from newspapers, record labels, and studios — and AI companies increasingly settle or sign licensing deals rather than risk trial. The big open question (can training on copyrighted work proceed without a license?) is still working through the appeals courts.

3. Machines can't hold copyright — and opting out got real

• In 2026 the Supreme Court declined to hear Thaler v. Perlmutter, leaving settled the rule that no one can copyright a work that has no meaningful human authorship. (Computer scientist Stephen Thaler had created an AI that then created a picture on its own.) The Copyright Office's 2025 Copyrightability report adds that prompts alone are not authorship: what you can copyright is your own creative contribution/selection/arrangement/edits.
• Publishers and platforms now routinely signal whether AI crawlers may index their work via robots.txt notices on their servers, Cloudflare's pay-per-crawl, and as of September 15, 2026, default blocking of mixed-use AI crawlers on ad-supported pages. Publishers are starting to insist on licenses to train on their works.

What this means for makers

Can I use it?

As a creator, you're still subject to the limitations of fair use described by Wendy Seltzer when using others' images, fonts, and code in your project, but now the Warhol case means competing with the source weighs heavily against you.

Safer paths you already know include using open-licensed works like Creative Commons or open-source software; see the "How to find shareable media" badge.

Can I copyright it?

If AI generated an asset in your project, that asset by itself is likely uncopyrightable. What you can protect is your human contribution: the layout you designed, the code you wrote (or substantially rewrote), the selection and arrangement of the whole. That's a practical reason your instructor may ask for an AI ledger: documenting your prompts, edits, and choices is the evidence of authorship that determines what you own.

Disclaimer: This is just a survey of the legal landscape; consult a lawyer for actual legal advice.

Exercise

Fair Use Referee

1. Instructor reads one scenario below aloud.
2. Raise your hand for the side you think won, creator versus re-mixer (human or AI).
3. Instructor tallies the vote on the whiteboard.
4. Click the scenario to reveal verdict and case.
5. Discuss surprises for one minute, then next scenario.

Scenarios (click each to reveal the verdict)

• 📸 You sell prints of a restyled celebrity photo you found online.
  🎨 Creators. Restyling for the same market isn't transformative — Warhol v. Goldsmith (2023).
• 📖 A lab legally buys ebooks, then trains a model on them.
  🤖 AI labs. Training on lawfully acquired books can be fair use — Bartz v. Anthropic training ruling (2025).
• 📚 A lab downloads millions of pirated books for its training library.
  🎨 Creators. Piracy isn't excused by a transformative end use; it cost $1.5 billion — Bartz v. Anthropic piracy ruling and settlement (2026).
• 🖼 You try to copyright an image made from a single prompt.
  🧑‍🎨 Nobody. No human authorship, no copyright; prompts alone aren't authorship — Thaler v. Perlmutter (2026) and the Copyright Office's Copyrightability report.
• ⚖️ A startup trains its legal-research tool on a rival's legal summaries.
  🎨 Creators. Training a tool that competes with its source isn't fair use — Thomson Reuters v. Ross (2025, on appeal).
• 🌐 A crawler scrapes ad-supported pages that never opted in.
  🎨 Creators (by policy, not court): major networks now block AI crawlers by default and charge for access — Cloudflare's pay-per-crawl and its September 2026 default block.
• 🎨 You heavily edit, rearrange, and combine AI output in your poster.
  🧑‍🎨 You — partly. Your creative selection, arrangement, and edits are protectable even when raw AI output isn't — Copyright Office Copyrightability report (2025).

Credits

AI disclosure

AI use in building this exercise: "Cyborg"

This page was outlined by Jon Ippolito; the research and exercise was generated by AI, then edited and fact-checked by Ippolito.