AI Chatbot Copyright: When Gemini Rewrites the News

You ask a chatbot what happened at last night's city council vote. It answers in four tidy sentences, and you close the app satisfied. You never see the local reporter who sat through the three-hour meeting, and their paper never sees your visit.

That everyday exchange is now at the center of the biggest AI chatbot copyright question in Europe. On November 3, an adviser to the EU's top court said a chatbot that fetches a news article and reuses it to answer you can infringe the publisher's rights. Below is what the opinion actually says, why it treats AI training so differently, and what it could change about the answers you get.

What the EU Adviser Said About AI Chatbot Copyright

The case is Like Company v Google Ireland (C-250/25). Like Company runs several Hungarian news portals and sued Google before the Budapest Regional Court over how its Gemini chatbot used those articles over almost a year of answers ending in May 2024. The Hungarian court paused the case and sent four questions to the Court of Justice in Luxembourg.

Advocate General Maciej Szpunar delivered his answer on November 3. According to the Court's official press release, it is the first generative AI case to reach the Court of Justice. He did not say Google broke the law. He set out the test a court should use to decide whether it did.

The mechanism at issue is called retrieval-augmented generation, or grounding. Instead of relying only on what a large language model learned in training, the chatbot pulls fresh pages from the web while you wait and writes its answer from them. That is how a chatbot can tell you about news from an hour ago. It is also why a specific article, not just general knowledge, ends up feeding the reply.

Facts Are Free, Wording Is Not

The opinion turns on one distinction. The EU's press publishers' right does not protect the facts in a story. It protects the specific way the story is written, because its purpose is to reward the investment in producing journalism, not to hand publishers a monopoly on the news.

From there, Szpunar set out two questions a court should ask about any chatbot answer:

• How did the chatbot get the article? If it got around an "effective technical measure" such as a paywall or an anti-scraping block, the copying is unlawful from the start.
• What does the answer reuse? An answer that simply reports facts is fine. One that reuses the article's form of expression in a recognisable way, so closely that you no longer need to open the original, can infringe.

A third point covers the right to make content available to the public. That right is, as a rule, infringed when the article's expression is systematically recognisable across the chatbot's answers to related questions. One lucky match is not the issue. A pattern is.

Liability sits with the provider when the copying is inherent to its business model, or when it has not put "appropriate measures" in place to stop it credibly and effectively. Google's position, reported by Reuters, is that Gemini "learns linguistic patterns to create original outputs rather than replicating content."

Why AI Training Got a Pass

Like Company also argued that Google trained Gemini on its articles. The adviser told the Court to throw out those two questions as hypothetical. There was no evidence the articles had been used for training, and training a model is not an "online use" of a press publication, which is the only thing this right covers.

He went further anyway. For completeness, he suggested that training a model on copyrighted material is, as a rule, freely permitted under the EU's text and data mining exception. That exception comes with conditions: the AI company needs lawful access to the material, and rightsholders can opt out.

So the opinion splits AI into two acts. Learning from articles in bulk, months before you ask anything, is treated as mining. Fetching one article at the moment you ask and rewriting it for you is treated as using it.

That's the overlooked part of the opinion, and it runs opposite to where most of the American legal fight has focused. Our explainer on the lawsuits reshaping AI covers the US cases, many of which center on training data instead.

What a Zero-Click Answer Costs a Newsroom

Publishers care about the second test because of traffic. When the answer is good enough, you stop there. A Pew Research Center study tracked a month of browsing by 900 US adults and measured the effect of Google's AI summaries directly.

• Users clicked a traditional result on 8% of visits when an AI summary appeared, against 15% when none did.
• They clicked a link inside the summary itself on just 1% of visits.
• They ended their browsing session on 26% of pages with a summary, against 16% without one.
• AI summaries appeared on 18% of the searches in the study.

A worked example: one local news site

Picture a regional news site whose stories show up in 100,000 searches a month. Apply Pew's rates to see what an answer box does to its visits.

• Without a summary: 15% of 100,000 = 15,000 visits.
• With a summary: 8% clicking ordinary results plus, at most, 1% clicking the summary's links = 9,000 visits.
• The difference: 6,000 fewer readers, a 40% drop on the same stories.

Every one of those lost visits is an ad impression or a subscription pitch the newsroom never gets to make. The European Publishers Council's executive director, Angela Mills Wade, put the bind this way to Reuters: "Publishers should not have to choose between search visibility and protecting their rights."

What Changes for You

Nothing changes on your phone tomorrow. An adviser's opinion is not a ruling, and the dispute itself covers a period that ended in May 2024. Still, the test he proposed would reshape how chatbots answer news questions in Europe if the Court adopts it.

If you read the news through chatbots

A provider worried about the "systematically recognisable" test has two cheap fixes: answer with facts, not phrasing, and send you to the source. You could see shorter news summaries, more prominent links, or more licensing deals with publishers. That matters more as Gemini spreads beyond Google's own apps, a shift covered in our piece on Apple tapping Google Gemini to rebuild Siri.

Your own habits matter too. If a chatbot answer on a news event gives you everything, click the cited source anyway when the story affects your money, health or vote. Summaries drop the caveats first.

If you publish anything online

The opinion gives real weight to technical barriers. A blocked scraper or a paywall turns copying into unlawful access, while an open page leaves you arguing only about what the answer reused.

• Check your robots.txt. Google's crawler documentation says the Google-Extended token controls whether your content is used for Gemini training and for grounding, and that it does not affect your inclusion or ranking in Google Search.
• Read what each token covers. Google's page describes Google-Extended in terms of Gemini Apps and Vertex AI grounding. Confirm separately how your content is treated in AI Overviews before assuming one switch covers everything.
• Keep evidence. The test asks whether answers are systematically recognisable, so screenshots of chatbot answers that echo your wording are the kind of record that matters.

What Happens Next

The Court's judges have now begun deliberating, and the press release says only that judgment will come "at a later date." Reuters reported that the Court follows most of its advisers' opinions, though it is not bound to. The Budapest court then applies the answer to Like Company's claim, and the ruling binds every other national court in the EU facing the same question.

A second track runs alongside it. The European Commission opened an antitrust investigation last December into whether Google uses publishers' content for AI Overviews and AI Mode without payment or a real opt-out, the Associated Press reported. A breach of EU competition rules can cost up to 10% of a company's annual global revenue, and the Commission set no deadline.

Put together, the copyright case decides what a chatbot may lawfully do with an article. The competition case decides whether publishers can realistically say no.

Frequently Asked Questions

Is it illegal for AI chatbots to summarize news articles in the EU?

Not as such. The adviser said reporting the facts from an article is lawful. An answer risks infringing the press publishers' right only if the chatbot got the article unlawfully, for example past a paywall, or reuses its wording so recognisably that you no longer need the original.

Is the Gemini copyright opinion legally binding?

No. An Advocate General's opinion advises the Court of Justice, which rules later and can disagree. Reuters reported that the Court follows the majority of these opinions, so it is a strong signal rather than a final answer.

Can AI companies train models on news articles in Europe?

The adviser suggested that training is, as a rule, permitted under the EU's text and data mining exception, as long as the company has lawful access and the rightsholder has not opted out. He also said the specific training questions in this case were hypothetical and should not be answered.

How do I stop Gemini from using my website content?

Google's documentation says adding a Google-Extended rule to your robots.txt controls whether your content is used to train Gemini models and to ground Gemini answers. It says this does not change how your site ranks in Google Search. Check the current documentation for what it covers before relying on it.

The Line Between a Summary and a Substitute

This opinion does not ban chatbots from talking about the news. It draws a line that most readers would recognise: telling you what happened is fine, while rewriting someone's story so you never need to read it is not. If the Court agrees, the chatbots you use in Europe will be pushed toward shorter answers and clearer sources.

Next time an answer feels complete, look at where it came from. The reporter behind it is part of the price of that answer.