EU Court's First AI Copyright Opinion: Training OK, Chatbot Answers May Not Be

Introduction: The Case Everyone in AI Was Waiting For

The first generative AI copyright case to reach the Court of Justice of the European Union just got its roadmap. In an opinion delivered today in Luxembourg, Advocate General Maciej Szpunar sketched where European law will likely land on AI and copyright — and the answer splits neatly in two. Training large language models on copyrighted material? As a rule, "freely permitted" under the EU's text-and-data-mining exception. But an AI chatbot that retrieves a news article in real time and reuses it to answer a user's question may infringe the rights of press publishers — "in certain circumstances."

The opinion isn't binding; the Court's judges follow their Advocates General in the majority of cases, and a final judgment is expected in roughly six months. But with Google Gemini, ChatGPT, and Perplexity increasingly replacing the link-based search results that publishers live on, this is the opinion that will shape how every AI chatbot and AI search engine operates in Europe — and how publishers fight back. Here's what it actually says.

The Case: A Hungarian Publisher, a Dolphin Story, and Gemini

The dispute started small. Like Company, a Hungarian operator of news portals (including one covering the Lake Balaton region), sued Google Ireland in Budapest. It claimed that between June 2023 and May 2024, Google's Bard — later Gemini — both trained on its articles and reproduced them in chatbot answers, violating the "press publishers' right" created by the EU's 2019 copyright reform.

The centerpiece was almost comic: a July 2023 article about Hungarian singer Kozsó's ambition to introduce dolphins to Lake Balaton. Asked to summarize it, Gemini produced an account with details — pool size, location, environmental considerations — that the publisher believed could only have come from its journalism. Szpunar found that more plausibly explained by retrieval-augmented generation (RAG), where the chatbot fetches live web content to ground its answers, than by training-time memorization. The Budapest court referred the questions to the CJEU, and the case became the EU's first word on generative AI and copyright.

The Opinion: A Sharp Line Between Training and Answers

Szpunar's split will be quoted for years:

• Training is (probably) fine. The use of copyrighted works to train an LLM is, as a rule, freely permitted under the EU's text-and-data-mining exception — provided developers respect its conditions, including lawful access and rights reservations. Notably, he advised the Court to declare the training questions inadmissible in this case, since no evidence showed Like Company's articles were actually used to train Gemini. The training analysis is guidance, not a holding.
• Answers are where the risk lives. RAG retrieval is different in kind: "When an LLM-driven chatbot retrieves a publication from a press publisher's website, reuses that publication to generate an answer, and provides that answer to the user, the press publisher's right may be infringed in certain circumstances."

He also drew a crucial boundary: copyright protects only the "form of expression" of a publication — never the "mere facts." An AI tool that reports the news a story contains is not infringing. One that reproduces the writing so recognizably that the reader "is entirely dispensed from accessing the original publication" may be.

The Two-Factor Infringement Test

For anyone building or buying AI tools, the practical core is the test Szpunar proposes for when a chatbot answer infringes:

On the facts of this case, Szpunar doubted the test is even met: Like Company apparently lacked effective anti-scraping barriers, and Gemini's dolphin answer didn't seem to reproduce protected expression. That nuance matters — it signals the opinion is a framework, not a guillotine for AI search.

Provider Liability and the Opt-Out Era

The opinion's sharpest teeth are on liability. A chatbot provider is responsible for infringements committed through its service when they are inherent to its business model — or when the provider "failed to put in place appropriate measures" to counter them credibly and effectively. For AI companies, that reads as a duty to build publisher controls, respect technical barriers, and police scraping by default.

The timing is not accidental. Days earlier, on November 1, OpenAI, Google, Microsoft, and Anthropic announced they will support a common robots.txt opt-out signal for AI training and retrieval — one declaration a publisher can make instead of negotiating with each lab. Google's response to today's opinion leaned on exactly that: Gemini "learns linguistic patterns to create original outputs rather than replicating content, supported by extensive publisher controls that give rightsholders clear choice and the ability to opt out." The legal pressure and the technical plumbing are converging on the same place: publisher choice, enforced by law.

What It Means for Tools and Publishers

For users, the likely outcome is actually good: AI search tools that summarize the news — whether Gemini, ChatGPT, or Perplexity — will face stronger incentives to cite sources, link out, and cut licensing deals, while no-publisher-tracking options like Kagi keep their appeal. For coding and research agents such as Claude Code or Gemini CLI, which work on your own code and documents, today's opinion changes little: the press publishers' right targets news reuse, not private workflows.

The Bottom Line

Europe's first serious answer to "can AI use my content?" is a split verdict: training gets a pass, answers get scrutiny. The Court usually follows its Advocate General, and a final judgment in six months or so could obligate the biggest AI chatbots to pay for news content — or redesign how they answer from it. Publishers gain leverage they haven't had since AI summaries started eating their traffic; AI companies get a workable playbook (respect barriers, don't substitute for the original, honor opt-outs). For everyone picking tools in 2026, the signal is clear: prefer AI tools that cite their sources — and if you publish, lock down your opt-outs now.

Frequently Asked Questions