A Guth Labs publication ↗
Back to News

EU court adviser says grounded chatbot answers can infringe press publishers' rights

AI-generated conceptual editorial illustration: Newspaper speech bubble and balanced scales representing AI answers and European press rights.
guthlabs.ai · AI-generated conceptual editorial illustration

An adviser to the European Union's highest court has recommended that judges recognize potential infringement when AI chatbots retrieve and reuse news publications to answer users. Advocate General Maciej Szpunar delivered his opinion on November 3 in Like Company v Google Ireland. It is not binding, and a judgment will follow. Court announcement

The dispute involves a Hungarian publisher's allegations about Gemini. Szpunar's approach separates model training from real-time retrieval, then asks what the resulting answer actually reproduces. That distinction is central: a system's ability to search for a source does not settle whether its subsequent use of that source is lawful.

The answer matters, not just the retrieval

The adviser distinguishes bare facts from protected expression. His reproduction analysis considers unlawful access, including bypassing effective technical restrictions, and recognizable reuse extensive enough to substitute for the original publication. Systematically providing such expression can also implicate the right of making content available to the public. He links provider liability to its business model or failure to implement effective countermeasures. Opinion summary

For product teams, these are different questions from whether an answer is accurate. A factually correct response can still raise concerns about how it was obtained and how much protected material it reproduces. Conversely, a link to a news article and a substitute for reading that article should not be treated as interchangeable product behaviors.

Existing law contains boundaries and exceptions

Article 15 of the EU's 2019 copyright directive establishes publishers' rights concerning online uses of press publications. Its exclusions include hyperlinking, individual words and very short extracts. The directive also preserves authors' and other rights holders' separate rights in material incorporated into publications. These are features of the existing legislation, not new rules announced by the adviser. Directive 2019/790

The same directive provides a text-and-data-mining exception subject to conditions concerning lawful access and appropriately reserved rights. For publicly available online material, it addresses machine-readable reservations. This framework cautions against interpreting any one AI dispute as unconditional permission to ingest everything found online. Directive text

Szpunar recommended finding the training questions inadmissible because the alleged use was unproven and, in his view, not the relevant online use. His additional comments favored training's general eligibility for the mining exception. Court announcement

A reason to audit the entire workflow

The practical preparation is to examine retrieval and answer generation together. A provider could document which sources are accessed, how restrictions are handled, what passages reach the model and whether responses reproduce distinctive wording. Testing should include repeated prompts, since an acceptable answer to one question does not establish consistent behavior.

Publishers, meanwhile, need evidence about the actual use they challenge. Training, temporary retrieval and displayed answers are separate factual claims. The pending judgment may clarify their legal treatment, but the immediate lesson is already useful: accuracy, attribution and authorization are different checks, and a news-answering system needs to consider each.

Sources and citations

Revision history

Published .