Checking AI-generated text: what can businesses learn from a ruling set aside?

Picture a finance director who asks an AI tool for a board paper on a supplier dispute, based on the supplier’s letters and the company’s own emails. The paper is tidy and has a section on the supplier’s claims. Three of those claims were never made. The board turns down a settlement on that basis, and the error only surfaces when the lawyer reads the correspondence. The example is invented, but the pattern comes from a Norwegian case decided this autumn.

What happened in the case from Hardanger og Voss District Court?

On 24 September 2026, Gulating Court of Appeal set aside a ruling of Hardanger og Voss District Court in a property tax dispute between Ineos Tyssedal AS and Ullensvang municipality (LG-2026-143154). Under the heading for Ineos’s submissions, the ruling set out nine paragraphs of arguments the company had never advanced. The one submission Ineos had actually made was neither mentioned nor dealt with. The municipality’s submissions were also misstated, which the municipality itself accepted.

In its appeal, Ineos said it was reasonable to assume that this part of the ruling was AI-generated. The court of appeal did not decide the point. It held that the reasons were deficient under the Dispute Act (Dispute Act § 19-6) whatever the cause. The chief judge of the district court has since confirmed to Rett24 that AI was used when the ruling was written, and that responsibility for the content lies with the judge. The district court’s ruling has not been published, so the errors are known only as the court of appeal describes them.

Why did the error have such serious consequences?

Under Dispute Act § 29-21(1), a procedural error is to be given effect where it is reasonably likely that it may have affected the decision. The court of appeal stressed that the parties’ right to argue their case and have it decided on that basis is central to civil procedure, and described the error as serious.

The municipality asked the court of appeal to decide the dismissal issue itself. It declined. The district court’s handling was so deficient that a decision on the merits at appeal level would have meant the issue was tried at only one instance, so the ruling was set aside under Dispute Act § 29-22(2)(c). The parties have spent an appeal getting back to where they started.

What do the Norwegian courts say about AI in judgment writing?

The Norwegian Courts Administration is preparing guidance on the use of AI in judicial work. According to its board paper, the draft went out for consultation to the Norwegian Bar Association and members of the courts’ cooperation council on 7 September 2026, and is due to be finalised by the end of October at the latest. The guidance is clear that AI is not to assess evidence, apply the law or carry out the concrete application of law to facts. Rett24 reports that the draft allows AI as support for language, structure, flow and clarity.

That is a sensible line, but the case shows where it is fragile. Summarising the parties’ submissions looks like a task of language and structure. It is not. A summary is a statement about what someone said, and it can only be checked against what they actually wrote.

AI tools rarely leave a box empty. They fill it with something that sounds right.

Checking AI-generated text in the business: where is the risk greatest?

The risk is greatest where the tool restates what others have said or agreed and the reader does not have the source to hand. The model fills the gaps with what is probable, and the result reads as smoothly as the rest.

Under the Companies Act, the board must ensure the business is properly organised (Companies Act § 6-12) and supervise day-to-day management (Companies Act § 6-13). Where AI is used to prepare matters for the board, the checking procedures form part of that organisation. Read more about board responsibility for AI. For contracts, the risk is covered in the article on vibe lawyering, and for customer-facing answers in the article on AI hallucinations.

For lawyers the bar is higher. Under the Norwegian Code of Conduct for Lawyers, item 4.2, a lawyer must not knowingly give the court false or misleading information, and under item 4.3 a lawyer must become properly acquainted with the case. A pleading citing invented cases is hard to reconcile with the latter, even where nobody meant to mislead.

How should the checks be organised?

Tie the checks to the legal effect of the document, not to how much AI was involved. An internal draft can tolerate more than a letter to the Data Protection Authority. Several measures belong in the company’s AI policy for employees, and the topic page on artificial intelligence brings the rules together.

1. Check every restatement against the primary source, whether it is a quotation, a submission, a figure, a date or a reference to legislation or case law.
2. Record who used AI for which parts of the document, so that whoever checks it knows where errors may sit.
3. Have someone other than the person who used the tool read documents with legal effect before they are sent or signed.
4. Separate editing from substance. AI can tighten a finished text, but summaries of what others have said must be checked against the original.

The simplest test is a practical one. Whoever delivers a document containing restatements should be able to point to where in the source each one appears, and if nobody can find the place, the paragraph comes out.

Questions and answers

Next legal review: 15 November 2026