
When Does a Chatbot’s Summary Count as Reproducing an Article?
A Hungarian publication ran an article about dolphins. A user asked Google Gemini a question, and the chatbot replied with the same facts and one sentence from the original. A comparison of the article and the response became part of a dispute between the publisher, Like Company, and Google. The case raises a question for any AI service with web search: when does a chatbot’s answer amount to reproducing someone else’s publication?
On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Case C‑250/25 before the Court of Justice of the European Union. He proposed looking at how the chatbot obtained the material and how much of the original is recognizable in its response. The Opinion is advisory. The Court will issue the final ruling.
Can a summary substitute for the original?
Imagine a journalist who gathers information from several sources, interviews experts, and builds an article around three findings. A chatbot uses the same article, follows the same line of argument, and keeps its distinctive examples while rewriting every sentence. The reader gets practically the entire piece in the chat.
Under Szpunar’s proposed approach, a summary of this kind may amount to reproduction if the publication’s protected form of expression is recognizable and the similarity is substantial enough for the response to serve as a substitute for the original. The structure, order of presentation, and distinctive wording may all matter. If a chatbot takes the facts and presents them in its own way, overlap in the facts alone is insufficient.
The dolphin article illustrates the distinction. Szpunar observed that most of the overlap between Gemini’s response and the article concerned facts, although the response included one sentence from the original. In his assessment, the example did not clearly show that the article’s form of expression, or that of any part of it, was recognizable in the answer. For publishers, comparing topics and facts alone may be insufficient. For developers, checking for verbatim matches addresses only part of the risk.
What changes if the article is behind a paywall?
The user sees a finished answer. Behind it, the service may have found the page, retrieved its contents, and passed the text to the model. Szpunar proposed examining how the service obtained the material in the first place. If a chatbot unlawfully bypassed a paywall or other effective technical measures restricting access or automated collection, that may be relevant in its own right when assessing whether reproduction occurred. Even if the user receives a short summary, how the service obtained the full text still matters.
What if the article was used to train the model?
The Opinion considers this scenario too. When a chatbot finds an article in response to a query, the article serves as a source for that particular answer. During model training, the text is processed at a different stage and for a different purpose. In Szpunar’s view, the EU’s text and data mining exception may apply to training. Its conditions include lawful access to the material and the absence of a rightsholder’s express reservation of rights in an appropriate manner. For publicly accessible web pages, the reservation can be expressed in machine-readable form.
What should teams check when adding web search to AI?
* Access to sources. Which pages does the service read in full? What content does it obtain through a licensed API? How does it respond to access restrictions? Review the full process, including any third-party search and content providers. * Responses to real queries. What happens when a user asks for a detailed summary of a specific investigation or analysis? Compare the answer with the original to see whether it preserves the article’s structure, line of argument, and distinctive passages. * Ability to reconstruct what happened. If a publisher raises a claim, can the team establish where the service obtained the article, what it passed to the model, and what the user saw?
The most revealing test for a product team is to ask the chatbot about one specific article, then compare the response and the original side by side. If the response preserves the article’s distinctive presentation so fully that readers have little reason to open it, that is a scenario to review with legal counsel before launch.
Official source
SYNCO
This material was prepared by SYNCO lawyers based on the legislation of the Republic of Uzbekistan and is for information only. It is not legal advice — please consult a specialist regarding your situation.

