A6, Lesson 5: Who owns what AI makes
Tags: Sourced stories / Our view / Illustrative examples. Self-paced reading and practice. No math or code. Every case or news claim carries its source and the date we read it; invented figures are labeled invented.
1. The story
In February 2023, Getty Images sued Stability AI, the maker of the image generator Stable Diffusion, claiming more than twelve million of its photographs were copied to train the model without permission, reported Reuters. There were two cases. In the UK, the trial ruling came on November 4, 2025, largely in Stability's favor: Getty dropped its central claim that the training itself was infringing after it could not show the training happened in the UK, and the court rejected Getty's remaining copyright theory on the record before it - Getty's own case did not claim the model stores or reproduces its photos. Getty won two narrow, historic trademark points over watermark-like images from early versions, and in December 2025 the same court granted Getty permission to appeal the copyright ruling, so the UK case continues. In the US, the Delaware case from the Reuters story was dismissed and refiled in California in August 2025, where it continues as of November 2026. Be precise about what this teaches: the UK case did not decide whether training on the photographs infringed copyright, because that claim was withdrawn. It does not establish a general rule for merchants or other countries.
2. The one idea
The rules for what goes into AI, and who owns what comes out, are being decided right now, and they differ by country. This lesson is orientation, not legal advice; every case claim in it was checked against the court decisions and the US docket on November 3, 2026.
3. The kitchen-table version
A bakery lifts a handsome photo of sourdough from the internet for its boxes because "everyone does it" - until the photographer's letter arrives. Owning the recipe is one thing; owning the photo of the bread is another. AI adds a third layer: if the photo was never taken at all, but generated, whose is it?
4. Worked example
Our invented candle shop wants 200 new lifestyle images. Three routes, with invented, script-checked figures. Route one: a local photographer, two days, $1,800, with the rights put in writing - about $9 per image. Route two: an AI tool at $30 a month - about $0.15 per image in fees, plus review time: ten minutes per image across 200 images is 2,000 minutes, roughly 33 hours of somebody's week. Route three, the common one: AI drafts for volume, a photographer for the hero shots that carry the brand. The fee difference is enormous; the ownership ground is different too, and the cheap route's true price includes reading the tool's terms and checking what rights you actually get. Excluded on purpose: usage-rights negotiations, model releases, and the unpriced risk of building a catalog on contested material.
Instructor: "So the $30 route is obviously right for a small shop?" Learner: "For volume drafts, maybe. But the hero image on the packaging is where a dispute would hurt - that is where a written assignment from the photographer buys clarity, not just pixels."
5. Why this matters to a shop
Your product photos, listings and brand assets may already be training data somewhere, and the images and copy your tools generate sit on ground that is still moving. A shop that builds its whole catalog on contested material is building during an earthquake, however unlikely any single tremor is to reach it.
6. What it costs and what it breaks
Reading tool terms takes an hour most owners never spend. Keeping original files and drafts takes disk space and habit. A lawyer's hour costs real money - and is cheap against a rebrand or a dispute. Using AI for drafts while keeping humans in the creative loop is slower than full automation and safer for ownership questions.
7. Our habits, not rules
Read the terms of the tools you use, specifically what they say about training on your input and about who owns the output. Keep your original files, shoots and drafts, so you can always show what is yours and how it was made. And when a decision carries real money - a rebrand, a licensed catalog, a big campaign - that is when the lawyer's hour is cheaper than the mistake.
8. Monday checklist
List which of your catalog assets are photographed, which are AI-made, and which you are unsure about. Open the terms of one AI tool you use and find the two passages on your inputs and your outputs. Make sure originals and drafts for your ten most valuable assets are stored where you can produce them. Flag the assets that carry real money - packaging, hero images, licensed designs - for a lawyer's read. Note which tools say they train on your uploads.
9. Try it yourself (five minutes)
Open the terms of one AI tool you use and find the two passages about your input and your outputs. Most owners have never read them. If you cannot find them or cannot parse them, that is itself a finding.
10. Self-check
- What did the UK trial court rule in November 2025, and what did it leave undecided? Answer: It ruled largely for Stability on the record before it; it did not decide whether training on the photographs infringed copyright, because Getty withdrew that claim.
- What happened in December 2025? Answer: The same court granted Getty permission to appeal the copyright ruling, so the UK case continues.
- Where is Getty's US case now? Answer: Refiled in California in August 2025 and ongoing as of November 2026.
- Why keep original files, shoots and drafts? Answer: So you can always prove what is yours and how it was made.
- Which of your current AI uses would you want a lawyer to bless first?
Sources
- Reuters, "Getty Images lawsuit says Stability AI misused photos to train AI," February 2023 (the US suit).
- Getty Images v Stability AI [2025] EWHC 2863 (Ch), High Court of England and Wales, November 4, 2025: training and outputs claims withdrawn, secondary copyright claim rejected on the record before the court, two narrow historic trademark infringements upheld; permission to appeal granted December 16, 2025 ([2025] EWHC 3343 (Ch)), so the case continues.
- The US case was refiled in the Northern District of California (3:25-cv-06891) in August 2025 and is ongoing as of November 2026.
- Judgments, AP and law-firm summaries read 3 Nov 2026.
- The shop-owner scenario and the catalog-shoot figures ($1,800, $30, 200 images, $9, $0.15, 33 hours) are invented, illustrative and script-checked, not a reported story.
- This lesson is legal-adjacent: it teaches sourced case history, not legal advice, and the case claims were checked against the judgments and the US docket on 3 Nov 2026.
- Review flags: "building during an earthquake" is our framing; no outcome of the pending appeal or US case is predicted.
- Go deeper, free: Harvard's session 9, Intellectual Property, works through fair use factor by factor and the big open cases. Watch it here.
Review flags
- "building during an earthquake" is our framing; no outcome of the pending appeal or US case is predicted.
Claim ledger
| Claim or example | Basis | Checked |
|---|---|---|
| UK training and outputs claims withdrawn; secondary-copyright theory rejected on the record before the court. | UK trial judgment; UK permission to appeal; US California case | 3 Nov 2026 |
| Permission to appeal granted 16 Dec 2025; no general merchant copyright rule decided. | UK trial judgment; UK permission to appeal; US California case | 3 Nov 2026 |
| US dismissed without prejudice and refiled in California 14 Aug 2025; ongoing at check date. | UK trial judgment; UK permission to appeal; US California case | 3 Nov 2026 |
Independence
Independently developed by Commerce Stories, not affiliated with Harvard. We link to the archived 2024 course; we do not reproduce its videos, slides or exercises. The linked videos were not watched for this written-material review.