New · 29 Sep 2026 · United States
Thomson Reuters v. Ross (appeal)
First appellate ruling: training a competing tool on headnotes not fair use
Moves “Training” · see on timeline
Data as of 21 Oct 2026 · v2026.10 · 94 decisions and laws · 59 policy papers · 19 jurisdictions
The law and the proposals disagree on two of the six questions that will decide IP law in the age of AI.
Every mark below is one source: a square for each jurisdiction’s controlling law, a ring for each policy paper. Select a question to trace how it got here.
Law, one per jurisdiction Proposal, one per paper Protective of rights holders Middle ground Open toward AI
01
Can AI be trained on protected works?
8 jurisdictions · 30 papers
Opposed
Law and proposals point opposite ways
02
Is a trained model a copy of its training works?
2 jurisdictions · 3 papers
Untested
Fewer than three sources on one side
03
Must AI developers disclose what they trained on?
2 jurisdictions · 1 paper
Untested
Fewer than three sources on one side
04
Can AI-generated work be protected?
8 jurisdictions · 13 papers
Aligned
Law and proposals point the same way
05
Is an AI provider liable when outputs copy protected works?
4 jurisdictions · 2 papers
Untested
Fewer than three sources on one side
06
Can AI be an inventor?
10 jurisdictions · 4 papers
Opposed
Law and proposals point opposite ways
New · 16 Sep 2026 · United States
Doe v. GitHub (Copilot appeal)
Generating new code is not removing copyright information under the DMCA; infringement left open
Moves “Output liability” · see on timeline
New · 31 Jul 2026 · Germany
GEMA v. Suno
Training an AI music generator on GEMA works without consent infringed; outputs did too
Moves “Training” · see on timeline
Timeline
How we got here, decision by decision
Drag the year to replay the law. Circles are courts, squares are legislation, diamonds are office decisions. Hollow marks are no longer good law.
2026-09-29 · United States · Court
Thomson Reuters v. Ross (appeal)
First appellate ruling: training a competing tool on headnotes not fair use
Answers “Can AI be trained on protected works?”: Only with permission or pay
Jurisdictions
The whole map on one screen
Each cell is the controlling source in that jurisdiction. Select any two jurisdictions to compare them side by side.
| Jurisdiction | Training | Model as copy | Disclosure | Authorship | Output liability | Inventorship |
|---|---|---|---|---|---|---|
| – | ||||||
| – | – | |||||
| – | – | |||||
| – | – | – | ||||
| – | – | – | – | |||
| – | – | – | – | |||
| – | – | – | – | |||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – | ||
| – | – | – | – | – |
Each cell is the most authoritative good-law source in that jurisdiction; a dash means none in our data yet. Select any two to compare.
Germany vs United Kingdom
Training
Freely allowedKneschke v. LAION (appeal)
Only with permission or payCDPA section 29A (text and data analysis for non-commercial research)
Model as copy
A copy of its training worksGEMA v. OpenAI
Not a copyGetty Images v. Stability AI
Disclosure
No controlling source yet
No controlling source yet
Authorship
Human authors onlyAI-generated logos case (Local Court Munich)
AI output protectedCDPA 1988, section 9(3): computer-generated works
Output liability
Provider liableGEMA v. OpenAI
No controlling source yet
Inventorship
Human named, AI help allowedDABUS (Federal Court of Justice)
Human inventors onlyThaler v. Comptroller-General (DABUS)
What's next
Pending matters that could redraw the map
Decisions still under appeal or awaiting judgment, each linked to the questions it could settle. Follow one to see it in your alerts.
Court of Justice of the EU (reference from the Budapest-környéki Törvényszék) · European Union
Like Company v. Google (C-250/25, reference to the CJEU)
Pending at the Court of Justice.
Could settle: Output liability
Oberlandesgericht Hamburg (Higher Regional Court) · Germany
Kneschke v. LAION (appeal)
Further appeal to the Federal Court of Justice allowed.
Could settle: Training
Landgericht München I (Regional Court) · Germany
GEMA v. Suno
Not final; open to appeal.
Could settle: Training
Delhi High Court · India
ANI v. OpenAI (interim order)
Interim order; the main proceedings continue.
Could settle: Training
For researchers and builders
Open, versioned, citable
Download it, query it, cite it. Every row carries its primary source, authority level and good-law status.
Download
Version 2026.10, updated with the data.
Method in four rules
- Law is read one jurisdiction at a time, by its most authoritative source.
- Proposals count one per paper, tagged by the kind of institution.
- Overruled or narrowed decisions leave the current count but stay in the history.
- Nothing is averaged. Every count opens its sources.
Query the API
GET /api/v1/sources ?question=training &jurisdiction=Germany &good_law=true
{
"id": "kneschke-laion-2024",
"title": "Kneschke v. LAION",
"body": "Landgericht Hamburg (Regional Court)",
"jurisdiction": "Germany",
"date": "2024-09-27",
"kind": "court",
"question": "training",
"position": "open",
"level": "first-instance",
"good_law": true,
"citation": "310 O 227/23",
"primary_url": "https://www.wipo.int/wipolex/en/judgments/details/2381",
"source_url": "https://www.dlapiper.com/en/insights/blogs/mse-today/2025/robert-kneschke-v-laion",
"url": "https://www.havenmarkip.com/pro/decisions/kneschke-laion-2024"
}Run this queryCite the dataset
@dataset{havenmark_observatory_2026,
author = {Havenmark IP},
title = {The AI \& IP Observatory},
publisher = {Havenmark IP},
version = {2026.10},
year = {2026},
url = {https://www.havenmarkip.com/pro}
}Follow the changes
New decisions and versions go to the RSS feed, and to your alerts for anything you follow.
Positions come from a curated selection, not every source worldwide. Narrowed or overruled decisions are excluded from the counts. Foundational pre-AI cases are kept in the database as background and are not counted in the six questions.