Almost every AI product in professional services rests on an assumption that nobody has tested in court: that reading other people's published work at industrial scale, and then handing parts of it back to a user, sits within the law without a licence.

That assumption is now in front of the Court of Justice of the European Union, in the first reference to reach it on generative AI and copyright. A Hungarian press publisher has asked whether a chatbot reproducing its material infringes, whether training on that material is a reproduction, and whether the text and data mining exception covers it.

For a UK practice the case is not directly binding, and its outcome is genuinely unknown. What it does is give a buyer a precise, well-drafted list of the questions a vendor's licensing story has to survive.

• Case C-250/25, Like Company v Google Ireland, is the first CJEU preliminary reference on generative AI and copyright, referred by a Hungarian court and heard by the Grand Chamber on 10 March 2026.
• Four questions were referred, covering communication to the public, whether training is a reproduction, whether the text and data mining exception applies, and whether output matching a source is a reproduction by the service provider.
• The Advocate General's Opinion was scheduled for 3 September 2026. As at 29 September 2026 we had located no published Opinion and no judgment. Anyone quoting what the Advocate General or the Court "said" should be able to point to a dated document on the case page.
• The UK left the EU copyright framework, so the ruling will not bind UK courts. It will still reshape vendor economics, since the corpora and the vendors are largely the same on both sides of the Channel.
• The question closest to a tax research tool is the fourth. Retrieval, quoting a source back to a user, is nearer to what a publisher complains about than training is.
• A buyer's practical protection is contractual: know what the corpus contains, on what basis, and who carries the risk if the basis changes.

The proceedings began in Hungary, where a press publisher complained that a chatbot reproduced and made available its protected publications without consent. The national court referred questions on the interpretation of the Digital Single Market Directive and the InfoSoc Directive to the CJEU, and the reference was published in the Official Journal.

The Grand Chamber, a fifteen-judge formation reserved for the Court's most important cases, heard argument for six hours on 10 March 2026. Allocation to that formation is itself a signal about how consequential the Court considers the questions. The Advocate General's Opinion was listed for 3 September 2026, with judgment to follow some months later.

1. Whether a chatbot displaying text partially identical to a press publisher's material amounts to a communication to the public, and whether it matters that the system was predicting the next word rather than copying a document.
2. Whether training a model on those works is a reproduction.
3. If it is, whether the text and data mining exception covers it.
4. Whether an answer returned to a user, containing part or all of an article in response to a matching prompt, is a reproduction by the service provider.

⚠️ Watch: The Advocate General's Opinion was scheduled for 3 September 2026. As at 29 September 2026 we had located no published Opinion and no judgment in C-250/25, and several summaries in circulation describe reasoning that does not yet exist. An Opinion, when it arrives, is advice to the Court and does not bind it; the judgment is what counts. Before relying on any account of "what the Court said", check the case page for a dated document and treat undated commentary as speculation.

The United Kingdom is outside the EU copyright framework, and a CJEU ruling delivered now has no binding force here, although UK courts may have regard to it. UK law has its own text and data mining exception, confined to non-commercial research, and in March 2026 the government dropped its proposed opt-out model for AI training in favour of gathering further evidence before legislating. The domestic position, in other words, is narrower than the EU's and not about to widen quickly.

Vendor economics do not respect jurisdiction. The large model providers, the corpora, and most of the professional-tools layer built on them operate across both territories. A ruling that requires licensing for training, or for retrieval, changes the cost base of products sold everywhere, including into UK practices, and it changes them mid-contract.

Litigation of this kind rarely produces a switch. It produces repricing, renegotiated supplier agreements, and content that becomes unavailable without notice. A tool that today answers from a broad corpus may answer from a narrower one next year, without any announcement.

"A cost base resting on an open question is still a cost base."

Of the four, the fourth is the one a tax practice should read twice.

Training is an argument about a corpus: what was ingested, when, and under what basis. It is largely invisible to the user and largely historic by the time anyone litigates it.

Retrieval is different. A tool that shows you where an answer came from is, by design, putting source material in front of a reader at the moment of use. That is closer to what a publisher actually complains about, and it is the architecture most professional research tools have deliberately chosen, because citation is the thing that makes an answer checkable. It is also live in the case itself: by one observer's account of the hearing recording, the Advocate General asked whether infringement should be assessed across the whole chain of acts, training through to output, rather than act by act.

The design feature that makes a research tool trustworthy, showing its sources, is the feature that sits nearest the legal question. A buyer should expect a vendor to have thought about that, and should be unimpressed by one who has not.

A vendor with a defensible position can describe, without hesitation: what material the tool reads, on what legal basis each category sits, what it is licensed to display as against merely consult, and what happens to the product if a category becomes unavailable. Legislation and statutory instruments occupy a very different position from a commercial publisher's commentary, court judgments sit somewhere between the two, and a vendor who treats them as one undifferentiated "corpus" has not done the work. Our own sources and update policy sets out where our material comes from and how currency is maintained.

Ask these in writing, and file the answers with your supplier records.

1. What is in the corpus, by category? Primary legislation, HMRC guidance, tribunal and court decisions, and third-party commentary carry different risk. A vendor should be able to separate them without a pause.
2. On what basis is each category used? The Open Government Licence covers legislation and HMRC guidance, but court judgments on Find Case Law sit under the Open Justice Licence, which does not permit computational analysis without a separate licence from The National Archives. Commentary generally requires a licence. "Publicly available" is a description of access, never a licence.
3. What is displayed as against consulted? Quoting a provision back to a user is a different act from having read it during training. Ask which the product does, for which categories.
4. Who carries the risk if the basis changes? If a source becomes unavailable or a licence is renegotiated, does the product degrade, does the price move, and which of you absorbs it? This belongs in the contract rather than in a reassurance.

Nothing in the pending case changes a practice's professional obligations, and it is worth saying plainly because the topic attracts overstatement.

Using an AI research tool does not transfer responsibility for an answer. The adviser remains responsible for the treatment, the citation and the advice, and a tool that shows its sources makes that responsibility easier to discharge rather than lighter. Our limitations and responsible use page is explicit about the boundary, and the reasoning behind it is the same as the reasoning for building your own test file for an AI tax tool: verification stays with the professional.

If you want to see what a cited, source-linked answer looks like before putting these questions to anybody, create a free account.

Four questions sit with the EU's most senior court, no answers have arrived, and a substantial part of the AI industry has priced all four at zero. That is the honest state of play, and the useful response is neither alarm nor indifference.

For a practice, the exposure is commercial rather than legal: a supplier whose corpus rests on an untested assumption is a supplier whose product may change shape during your contract. That risk is managed the ordinary way, by asking what the product is built from, getting the answer in writing, and agreeing in advance who carries the consequences of a change.

The firms that will be comfortable whatever the Court decides are the ones who asked the four questions while the answer was still unknown.

For the broader assessment criteria, see how to choose AI tax research software in the UK. Firms conducting a formal supplier review across several seats can book a call.

Like Company v Google Ireland is the first preliminary reference to reach the Court of Justice of the European Union on generative AI and copyright. A Hungarian press publisher alleges that a chatbot reproduced and made available its protected publications without consent, and the referring court asked the CJEU four questions on the Digital Single Market and InfoSoc Directives.

No. The Grand Chamber heard argument on 10 March 2026 and the Advocate General's Opinion was scheduled for 3 September 2026. As at 29 September 2026 we had located no published Opinion and no judgment in C-250/25. Commentary describing what the Court or its Advocate General has "found" should be checked against a dated document on the case page before it is relied on.

No. The United Kingdom sits outside the EU copyright framework and has its own, narrower text and data mining exception. UK courts may have regard to the ruling but are not bound by it. The practical effect on UK practices is commercial: the vendors, models and corpora behind most professional AI tools operate across both territories, so a ruling that changes licensing costs changes products sold here.

Training concerns what material a model read while it was built. Retrieval concerns material shown to a user at the moment of use, which is closer to the reproduction a publisher typically complains about. Since citation is what makes a professional research answer checkable, the design feature that creates trust also sits nearest the legal question.

Ask what the corpus contains by category, what legal basis applies to each category, which material is displayed to users as against merely consulted during training, and who bears the risk if a licence or source changes. Record the answers with your supplier due-diligence file and repeat the exercise annually.

No. Responsibility for the treatment, the citation and the advice remains with the adviser. A tool that shows its sources makes verification practical rather than removing the need for it, which is why testing a tool on your own closed files and checking that citations resolve remain part of competent use.