Facts yes, substitute no: what the first AI case at the EU Court of Justice means for your chatbots

TL;DR. On November 3, Advocate General Maciej Szpunar at the Court of Justice of the EU delivered his opinion in the first case on generative AI. It is about Google's chatbot Gemini and news articles it fetches live from the web while answering a question. His line: A chatbot may take the facts from an article. It may not get around a block, and it may not deliver an answer that replaces the article. He considers training permitted in principle. In our own test, the share of wording taken over depended mostly on the request: 1.3 percent for a short question, 14 percent when the user wanted the whole article.

What the case is about

A Hungarian news portal about Lake Balaton is suing Google in Budapest. The claim: Bard and later Gemini used its articles without permission between June 2023 and May 2024 and reproduced them in answers. The court in Budapest referred four questions to the Court of Justice, two on training and two on the chatbot in operation. The opinion in case C-250/25 runs to 240 points. It does not bind the Court, and no date for the judgment has been set.

At its core is a technique the opinion calls retrieval-augmented generation, RAG for short. While answering, the chatbot searches the web, pulls matching pages into its input and writes the answer from them, usually with a link to the source. This is how assistants with web search work, including many that companies build themselves.

The line: facts are free, a substitute is not

According to Szpunar, it becomes critical at two points. At retrieval, if the chatbot gets around a paywall or a block against bots. And in the answer, if the structure and wording of the article are so recognisable that the reader no longer needs to open the original. Whether the provider is liable for that is a third question.

The press publishers' right does not protect mere facts. A short summary in different words or an overview from several sources remains free, even if it takes readers away from the publisher. The right is not meant to protect publishers against competition as such, Szpunar writes (point 208). In the example from the proceedings, an article about dolphins in Lake Balaton, Gemini's answer contained the same facts in different words. Only the first sentence was the same (point 204).

The opinion names two reasonable measures: respect the publishers' blocks and keep answers at a distance from the individual article, for example with filters (point 222). That an article is freely available online does not count as tacit consent (point 228).

Szpunar considers the questions on training inadmissible, because nothing suggests that the articles went into Gemini's training. He answers them anyway: training falls in principle under the exception for text and data mining, as long as the content is lawfully accessible and nobody has objected in machine-readable form. A trained model is not a database full of copies, he writes (point 85). What such an objection could look like in future is covered in our October 31 post on the new robots.txt signal.

Our test: how much of the article ends up in the answer?

Whether an answer replaces the article is a legal question. How much wording it takes over can be measured. On November 4 we gave six news articles from Wikinews (free licence, 367 to 897 words, in German) to Claude Sonnet 5.5, the way a web search feeds them into an assistant. Four requests followed, each twice per article: 48 answers. We counted what share of the article appears word for word in the answer, in runs of at least six words.

• The request decides. A short question took over 1.3 percent of the article on average, a summary 8.8 percent.
• The model holds back, but only on wording. Asked to “reproduce the article in full”, it declined a word-for-word reproduction in all 12 cases. Instead it delivered a “detailed summary with all essential content”, around 390 words long, each time with the original headline and with 14 percent of the text word for word, 26.8 percent in one case.
• A rule works. With a quoting rule in the system prompt (at most ten words in a row, facts only, point to the source), the share for the same request fell to 4.7 percent. The headline was missing in all 12 answers.

Our count only captures word-for-word copying. The opinion goes further and also looks at structure and arrangement. A version with the original headline that explicitly contains everything essential, for a user who does not want to open the article, sits close to the grey area Szpunar describes.

The catch

• An opinion, not a judgment. The Court can decide differently.
• A weak case. The European Commission considered the reference inadmissible as a whole. Whether the portal is a press publisher within the meaning of the directive at all is left open (points 27 to 30). For Budapest, Szpunar indicates that Google probably infringed nothing (point 239).
• Open terms. What is “recognisable”, what belongs to the “business model” and which measures are “reasonable” will have to be settled by courts case by case. One early commentary objects that a chatbot that answers every question has exactly that as its business model.
• Only one piece. This is about the press publishers' right. Questions of source selection, distortion and disinformation need other rules, according to the opinion (point 238).
• Our test is small. One model, six articles, only wording measured.

What this means for mid-sized companies

If you use ChatGPT, Gemini or Copilot, you are not the addressee of the publishers' right. It targets providers of online services (point 218). But if you specifically instruct a chatbot to reproduce an article, you will as a rule bear the responsibility yourself, according to the opinion (point 219). And if you take answers into a newsletter or your website, you are publishing them yourself.

It is different if you build your own assistants with web search, for example for market monitoring, a press review or a chatbot on your website. Then you are the provider from whom the opinion expects measures. They are manageable.

If you publish specialist articles yourself, you hold the opposite lever. Technical blocks decide whether a chatbot may access your content. However, robots.txt so far distinguishes poorly between training, search and retrieval for answers, as the opinion also notes.

What you should do now

1. List your own applications with web search. Which ones fetch third-party articles, and who reads the answers: only your team, or customers too?
2. Respect blocks. No retrieval behind paywalls or against bot blocks, not even through bought-in search or scraping services.
3. Limit the output. A quoting rule in the system prompt, facts instead of wording, source with a link. In our test this cut the word-for-word share by two thirds.
4. Check samples. Hold a few answers a month against their sources. A simple comparison of word sequences is enough to start with.
5. Handle full texts through licences. Anyone who wants to distribute whole articles in-house needs permission, as before. AI does not change that.

For everyone who researches with AI, the opinion is good news: facts stay free. Anyone who builds assistants sets them up so that they do not replace the article. This post provides context and is not legal advice. If you want to check your own AI applications against this, we are happy to help.