Who Owns the Training Data? Copyright, Consent and the Coming Data Wall
Two forces now shape the data layer from outside: the law and scarcity. Courts in the United States, Britain and Germany have started ruling on whether training on copyrighted work is allowed; Europe now requires model makers to publish what they trained on; and the supply of fresh, high-quality human text is running short. This final part of the series maps the rulings and rules as of October 2026, explains the opt-out signals sites use, and models when the "data wall" arrives and what labs are doing about it.
1. Case file
2. Timeline
Law and limits, 2018 โ 2026
Tap a milestone. Colour: legislation, US cases, other courts, scarcity.
3. United States: fair use, case by case
US copyright law has no specific rule for AI training; it turns on fair use, a four-factor test weighing the purpose of the use, the nature of the work, how much was taken, and the effect on the market for the original. Three 2025 decisions set the early tone:
Dozens of other cases remain open, including The New York Times v. OpenAI and Microsoft and suits by music publishers, artists and newspapers. The pattern so far: how data was acquired matters as much as how it was used, and harm to the market for the original is the battleground.
4. Britain, Germany, Japan
5. The EU AI Act: disclose what you trained on
Since 2 August 2025 providers of general-purpose AI models in the EU must keep technical documentation, adopt a copyright policy that respects rights-holders' machine-readable opt-outs under the EU's text-and-data-mining rules, and publish a summary of training content using a template the European Commission released on 24 July 2025. The template balances transparency with trade secrets; a voluntary Code of Practice offers a route to compliance.
The training-data summary template
Tap a section. Our summary of the Commission's template.
6. Rulings at a glance
Simplified; many cases have further claims, appeals or settlements. โ ๏ธ = reported, not checked against the judgment.
7. Is it allowed? A scenario explorer
Pick an activity and a place
Our editorial reading of the direction of rulings and rules as of October 2026. Not legal advice; outcomes depend on facts.
8. Opt-out signals
| Signal | How it works | Legal weight |
|---|---|---|
| robots.txt | Disallow specific AI crawlers (GPTBot, ClaudeBot, CCBot, Google-Extendedโฆ) โ see Part 2 | Voluntary in the US; can count as a machine-readable reservation in the EU |
| Terms of service | Site terms forbidding scraping or AI training | Contract claims; enforceability varies |
| TDM reservation | EU Article 4(3) opt-out, e.g. in metadata or protocols such as the TDM Reservation Protocol โ ๏ธ | Binding in the EU for commercial text and data mining |
| Network blocking | Cloudflare's default AI-crawler block and pay-per-crawl (Jul 2025) | Technical, not legal: crawlers simply cannot fetch |
| Dataset opt-outs | "Am I in The Stack?", artist opt-out registries | Voluntary, per dataset |
9. The data wall
Epoch AI estimated in 2024 that the effective stock of public human text, adjusted for quality and repetition, is about 300 trillion tokens (90% range 100Tโ1,000T), and that if training datasets kept growing at their historical pace, frontier models would use all of it between 2026 and 2032. Restrictions (Part 2) shrink the usable share further; filtering (Part 3) shrinks it again.
When do we hit it?
Our model: the largest training dataset starts at ~15T tokens in 2024 and grows by a chosen factor per year; compare with the stock estimate and the share you assume is usable. Change the inputs.
10. Ways over the wall
11. The series, complete
Inside the Data Layer: eight parts
12. Reading the numbers
- Primary / widely confirmed
- Bartz v. Anthropic: fair use for lawfully acquired books, not pirated copies (Jun 2025); $1.5B settlement, ~$3,000 ร ~500K books, final approval. Kadrey v. Meta: summary judgment for Meta, judge's warning on market harm (Jun 2025). Thomson Reuters v. Ross: fair use rejected (Feb 2025). Getty v. Stability AI: High Court, 4 Nov 2025, model not an infringing copy, narrow trademark findings. EU AI Act GPAI obligations from 2 Aug 2025; template published 24 Jul 2025. Epoch AI: ~300T tokens (100Tโ1,000T), 2026โ2032.
- Reported โ ๏ธ
- Ross appeal status; GEMA v. OpenAI (Munich, Nov 2025); Japan Article 30-4 scope; UK consultation; TDM Reservation Protocol details.
- Our illustrations
- Scenario explorer verdicts, data-wall model (15T start, chosen growth, usable share and epochs), strategy notes.
- Cases and rules
- AP via TPR: Anthropic $1.5B settlement
- McKool Smith: Kadrey v. Meta ยท Thomson Reuters v. Ross
- ABA: Thomson Reuters v. Ross one year later
- Cleary Gottlieb: Getty Images v. Stability AI
- Mayer Brown: EU AI Act GPAI rules and training-data template
- The data wall
- Epoch AI: Will we run out of data?
- Part 7: data infrastructure ยท Part 1: the map