How a law firm can write with AI and stay on the right side of copyright

A colleague reads a useful article about a change in the law. They want clients to hear about it from the firm today. So they ask an AI tool to rewrite it. The tool does a good job. The new version reads smoothly in fresh words.

Yet it still follows the original writer's argument and examples in the same order. New words do not make the writing the firm's own. A new opinion at the EU Court of Justice shows where the line sits.

What happened at the EU Court

Advocate General Maciej Szpunar gave his opinion in Case C-250/25, Like Company v Google Ireland on 3 November 2026. The Court's press release calls it the first generative AI case to reach the Court of Justice. A Hungarian publisher says Google used its articles without consent to train its AI models and in its chatbot's answers. The chatbot was Bard and then Gemini (points 14 to 16). The example before the Hungarian court is an ordinary one. A user asked the chatbot to summarise one of the publisher's articles about a singer's plan to bring dolphins to Lake Balaton (points 17 and 18).

He would have the Court leave the training questions aside. Nothing showed that Gemini had been trained on the publisher's articles. Training is also not the kind of online use the publisher's right covers (points 46 to 58). He still adds a provisional view for completeness. The EU's text and data mining exception in principle allows training on lawfully accessible content unless the rightholder has opted out (points 3, 58, 99, 141, 143 and 144).

His focus is the chatbot that fetches a press article to answer a question. He looks at it under the press publishers' right. EU law created that right in 2019 and Irish law adopted it in 2021 (Directive (EU) 2019/790, Article 15, and S.I. No. 567 of 2021, reg 13). He sees two cases where a chatbot's use of an article can fall within it. In the first the chatbot reached the article unlawfully. Getting round a paywall would be one way. In the second its answer keeps so much of the article's way of writing that the reader "can dispense with accessing the original publication entirely" (points 188, 202, 207 and 240).

Publishers have no monopoly over the facts. Their right covers the way an article reports them (points 199 and 208). Rewording is no way round that. The question is whether the writing is still recognisable "despite any rephrasing" (point 205).

Even then the chatbot's provider answers for it only where that use is built into its business model or where it lacks the measures expected of "a reasonably diligent operator". He would generally put the "primary" liability on a user who directs the chatbot to copy (points 219 to 221 and 240).

The dolphin answer did not seem to him to keep the article's way of writing. Google also seemed to have appropriate measures in place. The Hungarian court will decide both points (point 239).

An opinion is not a judgment. It does not bind the Court. The judges rule later (the press release, page 2).

Why it matters to a firm's website

The case is about chatbots and press publishers. It does not decide how a firm writes its website. A page on the firm's site is still the firm's own publication whatever a tool's maker may answer for. Copying a work without the owner's licence infringes copyright unless an exception applies. So does making it available to the public (Copyright and Related Rights Act 2000, s.37(1)(a) and (b) and (2)).

The opinion asks whether an answer could replace the article. That test belongs to the press publishers' right. The writer's own copyright sits beside it untouched (S.I. No. 567 of 2021, reg 13(2)(a)). That copyright covers "any substantial part" of a work (Copyright and Related Rights Act 2000, s.37(3)(a)(ii)). A part is substantial because of what is taken and not only how much. A short passage can still matter. A draft need not replace the whole article to cause a problem.

How short? The Court of Justice has held that eleven words in a row can be a copy of part of a newspaper article if they express the author's own intellectual creation (Infopaq, C-5/08, paragraphs 44 to 48). Even part of a sentence may be protected if it conveys the writer's originality (paragraph 47). What counts is whether the words carry the writer's own expression. There is no safe word count.

So it is not enough to ask whether the tool changed enough words. Does the article give the firm's own explanation or does it still lean on another writer's?

Five habits that help

Most firms already keep these habits for anything with their name on it.

Choose who can publish

Let anyone draft. Keep the right to publish with a few people who know the firm's standards and will stand behind the page. A finished draft should never be enough to put a page live.

The Law Society's December 2025 guidance on generative AI asks for office policy that sets out "permitted uses, accountability, safeguards and guidelines" (page 15). It leaves copyright itself for further guidance from subject-matter experts (page 2). Who may publish is a natural first line of that policy.

Use a story for its facts

Start with what happened. Check it at source in the judgment or the Act itself where you can. Then decide what the firm's readers need. What changed? Who is affected? What is still unclear?

AI can help gather that and shape a draft. Asking it to redo another writer's article is a different job. No amount of rewording changes that. A draft that keeps the original's structure and distinctive phrasing can still borrow too much.

Check the firm may use another writer's text that way before giving it to an AI tool. The firm may be free to read it and still not be free to copy it. Copying any substantial part needs the owner's licence or an exception (Copyright and Related Rights Act 2000, s.37(1)(a), (2) and (3)(a)(ii)). The text and data mining exception is unlikely to help. The Advocate General reads it as covering analysis that looks for patterns and trends. Taking one article to use directly is something else (point 179).

Quote briefly and give credit

Sometimes the exact words matter. A short quote lets the firm explain them or test them.

Irish law allows quotations from a work lawfully made available to the public. The use must not "prejudice the interests of the owner" and must name the work and its author (Copyright and Related Rights Act 2000, s.52(4) and s.51(3)). The matching EU rule is Article 5(3)(d) of Directive 2001/29/EC. Under it the Court of Justice looks for a direct and close link between the quote and the writer's own thinking. The quote should play a secondary role (Spiegel Online, C-516/17, paragraphs 78 and 79).

In practice that means four things.

• Quote only what the point needs.
• Make the quoted words unmistakable.
• Name the work and its author, and link to the original.
• Add the firm's own explanation.

Credit matters. It still does not make any amount of copying fine.

Read every draft

Someone who knows the subject should read the whole article. That includes the headline, the questions and the summary. Check the law and the dates, names and links. Put the draft beside any article it drew on. Has it kept the same argument or examples or order even where no sentence matches?

Read for the advertising rules too. They cover any communication meant to promote a legal practitioner "whether produced by electronic or other means" (S.I. No. 644 of 2020, reg 2(a)). An AI draft can slip in something they ban. The firm's success rate is one example (reg 8). "No win no fee" on a personal injuries page is another (reg 6(b)).

An article meant mainly to inform on the law can fall outside them (reg 2(a)). The regulations go further for an article on a legal topic written by a legal practitioner when no part of its space is paid for by or for them. Such an article is deemed not to be an advertisement. The Legal Services Regulatory Authority may still treat repeats of a substantially similar article as promotion (reg 10(a)(ii) and 10(c)).

A draft that reads well still needs checking.

Keep a record

Note who approved the article and when. List the sources it used. Keep the version that went live.

The regulations ask for more when the article is an advertisement. The firm must keep a copy with any written instructions for it and any consents it needed. It keeps them for at least twelve calendar months after the article last appeared so the Authority can see them on request (reg 11(f)). A note of who approved it does not replace that copy.

A record beats anyone's memory of what the tool produced when a question comes later.

How PageMax builds the habits in

They sit in the steps a firm already takes to publish an article.

• The firm chooses who can publish. An article PageMax drafts waits for one of those people to approve it. Nobody at PageMax can sign it off for them.
• A shared story is a source of facts. PageMax reads it for what happened. It then drafts a new article about the law the story raises for the firm to check. Its rule is to credit the outlet and link to it whenever it has the address.
• The story sits beside the draft. The person approving sees it with any repeated wording marked. Drafts that follow the story too closely, quote it at length or use its facts without credit are flagged too.
• Repeated wording can stop approval. Ten or more words in a row from the story hold the approval until the passage changes. That applies outside quote marks in the body and in its questions and answers. Ten is the product's own threshold. The law sets no number and a shorter passage can still matter. Changing a few words does not make a passage safe to publish. Nor do quote marks round a long passage. The fix is to write the passage afresh from the facts or to quote it briefly with credit.
• PageMax says who it is when it reads a link a firm shares. It never gets round a login or a paywall. It skips any page whose robots.txt turns away PageMax or a common AI tool. A firm can paste a story's text instead. Pasting does not change who owns the words. Paste only text the firm is free to use that way.
• The advertising rules get a first read. PageMax flags the plainest wording the regulations ban in an Irish firm's article. The firm's own success rate is one example. The check helps the person approving and decides nothing for them.
• Approval is recorded. PageMax records who approved each article and when. It keeps every version that went live until twelve months after it came down.

These checks compare a draft with the story it was given. They do not search the whole web and cannot prove every line original. The firm's own read matters most.

A separate decision about AI crawlers

Writing with AI is one decision. Letting AI tools read the firm's own website is another.

The big providers give sites separate controls for training and for search. Our piece on AI search explains how. OpenAI says blocking GPTBot tells it a site's content should not be used for training. A site that blocks OAI-SearchBot "will not be shown in ChatGPT search answers" but can still appear as a navigational link (OpenAI).

Google says Google-Extended covers training future Gemini models and grounding in Gemini Apps and Vertex AI. It "does not impact a site's inclusion in Google Search" (Google). AI Overviews and AI Mode are part of Search. Google points sites to Search's own controls for them. A firm can therefore turn away training crawlers by name and keep its place in ChatGPT search and Google Search. Blocking Google-Extended has one cost. Gemini can no longer ground its answers in the site.

The usual way to say no is robots.txt. It is a file that tells crawlers what they may read. The Advocate General gives a provisional view on it as part of the training questions. He reads it as a valid way to opt out of text and data mining where today's crawlers reliably understand it (Directive (EU) 2019/790, Article 4(3), and points 3, 58, 148 and 149). Irish law asks for that reservation to be machine-readable for content online (Copyright and Related Rights Act 2000, s.53B(3)(a)). Think of it as a sign on the gate with legal weight. It is not a lock.

The robots.txt PageMax writes for the firm websites it runs lets training and search crawlers in. The one exception is Amazon's Amazonbot. The file turns Amazonbot away and otherwise does not reserve a firm's pages from text and data mining (Copyright and Related Rights Act 2000, s.53B(1) and (3)). The rule is the same for every live firm today. Allowing training is our current policy choice. We would rather the models learn a firm's subjects partly from its own pages. Training is not needed to appear in ChatGPT search or Google Search. Allowing it promises no extra visibility.

Five questions before the next article goes live

1. Who has the authority to publish it?
2. Where did its facts and quotes come from?
3. Does it give the firm's own explanation?
4. Who checked it against its sources and the advertising rules?
5. Where is the approved version kept?

Those questions will serve the firm well whatever the Court decides.

This is not legal advice. We build the system, we do not advise on the law. If copyright matters to how your firm publishes, take your own advice on it.

PageMax is the AI platform for law firms. Your firm edits in plain words, approves and publishes, with a record of who signed off on what. If you would like to see how that works on your own site, book a demo. No obligation.