legal247

Copyright and AI: who owns the text, code and images that AI creates?

In brief

Content that AI creates entirely on its own probably has no copyright under Norwegian law, because the Copyright Act requires original and individual creative effort by a human. If a human reworks the output creatively, the human contribution may be protected. A vendor's promise that you own the output gives no more protection than the law itself provides.

The marketing department creates campaign images with an image generator. The developers let a coding assistant write large parts of a new service. An agency delivers copy that has mainly been written by a language model. In all three cases the question is the same. Who owns the result, and can the business stop others from using it?

The answer depends on how much a human has actually created. That has a direct bearing on the value of the content, on what competitors can copy, and on what you should require in contracts with employees, consultants and agencies.

Only where a human has made a creative contribution. Under Copyright Act § 2 of the Norwegian Copyright Act (åndsverkloven), the person who creates a work has copyright in it. A work means a literary or artistic work that expresses original and individual creative intellectual effort. Texts, photographic works, visual art and computer programs can all be works.

The requirement of original and individual creative effort is generally understood to presuppose a human being. A language model is not a person and cannot be an author. Content that AI produces without a human making the creative choices therefore probably has no copyright.

In Case C-5/08 Infopaq, the Court of Justice of the EU held that protection requires the work to be the author’s own intellectual creation. The requirement has since been followed up in a series of decisions and is also relevant to Norwegian law, since the Copyright Act implements the EU copyright directives through the EEA Agreement.

There is, as yet, no Norwegian judgment on copyright in AI-generated content. The conclusion above is based on the wording of the Act and the originality requirement of the Court of Justice, and it is consistent with legal developments in other countries.

US law requires human authorship, and the question is now finally settled for works created entirely autonomously. In Thaler v. Perlmutter, the Court of Appeals for the District of Columbia Circuit held on 18 March 2025 that an image created by an AI model on its own could not be registered for copyright. The US Supreme Court declined to hear the case on 2 March 2026.

The court made clear that the requirement does not preclude protection for works made with the help of AI. What matters is that the author is a human. In its report on copyright and AI from January 2025, the US Copyright Office concluded that prompts alone do not give sufficient control over the expression, but that a human may obtain protection for their own contributions, such as creative modification or the selection and arrangement of AI output.

US law is not binding in Norway, but these decisions indicate where the line is likely to be drawn here too.

Is it lawful to train AI on other people’s works?

The first appellate judgment on AI training in the United States went in favour of the rights holder. In Thomson Reuters v. Ross Intelligence, the Court of Appeals for the Third Circuit affirmed on 29 September 2026 that Ross Intelligence could not rely on fair use when it used headnotes from Thomson Reuters’ Westlaw to train a competing legal search tool. The case did not concern generative AI. The opinion was initially issued under seal, but was made public without redactions after the parties had stated that no redactions were needed.

The court held that the 2,243 headnotes at issue were sufficiently original to be protected, and that three of the four statutory factors in the US fair use provision (17 U.S.C. § 107) weighed against fair use. The use was commercial and added little, if any, new purpose, since Ross used the headnotes for the same purpose as Thomson Reuters, namely to build a legal search service. In assessing the effect on the market, the court gave weight to the fact that the market for licensing headnotes as AI training data was developing rapidly. By copying the headnotes without consent, Ross had deprived Thomson Reuters of the opportunity to license them for that purpose, and this applied even though Thomson Reuters had not licensed the headnotes to others. The court expressly distinguished the case from the actions against providers of large language models, in which it has been argued that training gives the use a new purpose because the model can produce new expression.

The decision is not final. It concerned an interlocutory appeal against a partial decision of the district court before trial, and Ross has been granted an extension until 3 November 2026 of the deadline for petitioning for rehearing by the full court (en banc). An application may thereafter be made to bring the case before the US Supreme Court.

For Norwegian businesses, US fair use case law is of limited legal significance, because the question of infringement in respect of use in Norway is determined under the Copyright Act. The case law is nevertheless relevant in two respects. Agreements with US AI vendors often contain a US choice of law, and the vendor’s litigation risk in the United States affects which warranties and what liability for infringement the vendor is willing to accept. The decision also supports the view that a market exists for licensing training data, which may strengthen the negotiating position of Norwegian rights holders who own texts or databases that are attractive for AI training.

There is no general fair use rule in Norway. Use of protected works without consent must fall within one of the listed limitations in Chapter 3 of the Copyright Act. According to Lovdata’s consolidated text (last amended by Act of 13 December 2024), the Act contains no specific provision on text and data mining. A proposal to implement the EU Digital Single Market Directive, which contains such rules, was put before the Storting in Prop. 41 LS (2025–2026) on 27 March 2026. The European Commission has also opened a targeted consultation on copyright and generative AI, including on licensing. Anyone training their own models on other people’s texts should therefore expect a licence to be required.

Can websites stop scraping for AI training through terms of use and technical barriers?

The collection of online content for AI training, commonly referred to as scraping, is now the subject of several actions brought by news publishers. On 6 October 2026, Postmedia, Torstar, The Globe and Mail and The Canadian Press commenced proceedings against Anthropic before a court in Ontario. According to the statement of claim as reported by The Globe and Mail, Anthropic copied large volumes of text from the publishers’ websites in order to develop Claude, and continues to do so when users submit prompts. The publishers allege copyright infringement, breach of the websites’ terms of use, circumvention of technological protection measures intended to prevent copying, and unjust enrichment. The Canadian trade press has reported that the publishers seek, among other remedies, statutory damages of up to 20,000 Canadian dollars per work, although that figure has not been verified against the statement of claim. Anthropic had not commented when the action became public, and the allegations have not been tested in court.

In the United States, USA Today Co. and affiliated newspapers brought an action on 8 October 2026 against OpenAI Foundation and several other OpenAI entities in the US District Court for the Southern District of New York (No. 1:26-cv-08892). The case is docketed as a copyright infringement claim and has been assigned to the same judge who is handling the consolidated copyright cases against OpenAI. According to Forbes, the plaintiffs seek more than 250 million dollars in damages, including for the removal of copyright management information from their articles, together with an order requiring the destruction of models and training datasets containing their journalism. These claims have not been verified against the complaint itself.

The Canadian action is notable because it targets the act of collection itself, and not only the subsequent use of the material in the model. Under Norwegian law, the circumvention of effective technological protection measures is prohibited by Copyright Act § 99, and the removal of electronic rights management information is governed by Copyright Act § 102. Whether a paywall, a login requirement or the blocking of known crawlers amounts to an effective protection measure within the meaning of the Act depends on how the measure is designed. Terms of use may also provide a contractual basis, provided that the party collecting the content is bound by them. In our assessment it is uncertain whether an automated collector that has never actively accepted the terms becomes a party to the contract, and the question has not been settled in Norwegian case law.

Under EEA law, the relationship between text and data mining and the rights holder’s reservation is governed by the Digital Single Market Directive, which the EEA Joint Committee has decided to incorporate into the EEA Agreement. Under DSM Directive Art. 4(1), Member States must permit reproductions of lawfully accessible works for the purposes of text and data mining, including for commercial purposes. The exception applies only where the rights holder has not expressly reserved its rights in an appropriate manner, and for content made publicly available online DSM Directive Art. 4(3) gives machine-readable means as an example. The requirement of lawful access means that the exception does not cover content obtained by circumventing a paywall or a login requirement. In addition, providers of general-purpose AI models must under AI Act Art. 53(1)(c) put in place a policy to comply with EU copyright law, including identifying and complying with such reservations. The AI Act has not been incorporated into the EEA Agreement, but the obligation applies to providers placing such models on the market in the EU.

Norway has not yet implemented the Directive’s rules on text and data mining. On 27 March 2026 the Government submitted Prop. 41 LS (2025–2026) to the Storting. The bill proposes a general text and data mining exception subject to reservation in a new Copyright Act § 50 e and a separate exception for scientific research in a new Copyright Act § 50 f. According to the Storting’s case page, the Standing Committee on Family and Cultural Affairs is due to deliver its recommendation by 3 November 2026, and the plenary debate is scheduled for 12 November 2026. Until the rules have been adopted and have entered into force, reproduction for AI training in Norway as a rule requires the rights holder’s consent or must fall within one of the existing limitations in Chapter 3 of the Copyright Act.

For Norwegian publishers and others who make content available online, these actions are a reason to treat reservations, terms of use and technical barriers as a whole. A machine-readable reservation, for example in robots.txt or in metadata, is the form the Directive points to for online content, and it should be supplemented by an express reservation against text and data mining in the terms of use. Businesses that themselves collect online content for their own models or analyses should establish whether the sources have reserved their rights, and should not circumvent login requirements, paywalls or blocking measures.

Where is the line between AI-assisted and AI-generated content?

The line is whether a human has made the creative choices that shape the final expression. The more AI has decided, the weaker the protection.

Situation Likely protection
AI creates text, an image or code from a short prompt Probably no copyright
Many detailed prompts, but AI determines the expression Probably no copyright, unsettled
A human substantially reworks the AI output Protection for the human contributions
A human selects and arranges AI elements in an original way Possible protection for the arrangement
A human writes the work and uses AI for editing or proofreading Ordinary copyright for the human

The risk is greatest for images, because image generators make most of the visual choices themselves. For text and code there is often a continuous interplay between human and machine, and protection must be assessed for the individual parts. In practice it is hard to prove after the event who did what. Anyone who wants protection should therefore document the human contribution.

What does it mean when the vendor says you own the output?

Less than it may seem. In their terms, the major AI vendors assign whatever rights they may have in the output. OpenAI’s EU terms, for example, state that the user owns the output and that OpenAI assigns its rights “if any”, meaning if there are any.

A vendor cannot assign a copyright that never came into existence. The term only ensures that the vendor itself makes no claim to the content. It gives the business no right to stop others from copying it. The terms also often restrict how the output may be used, and they usually offer little protection if the output infringes the rights of others.

Who owns what employees, consultants and agencies create with AI?

For employees, the starting point favours the employer. Copyright in computer programs created by an employee in the course of employment passes to the employer under Copyright Act § 71, unless otherwise agreed. For other works, rights pass under general non-statutory rules to the extent necessary for the employer’s ordinary business. The employment contract should nevertheless deal with this expressly.

For consultants and agencies, the starting point is the reverse. The rights remain with the author unless they have been assigned by contract. Under Copyright Act § 67, the author is not deemed to have assigned more extensive rights than the contract clearly expresses. Under Copyright Act § 68, an assignment does not include the right to alter the work unless otherwise agreed.

AI adds a new risk to such contracts. If the agency has largely let AI create the deliverable, there may be no copyright to assign. The business has then paid for exclusivity it does not get. The contract should therefore require the supplier to disclose its use of AI and to warrant that the deliverable does not infringe the rights of others. See also the article on vibe lawyering about the risk of contracts drafted with AI without legal review.

What does this mean for the business’s risk?

The greatest risk is that content you regard as an asset is not protected. Logos, campaign material and product designs mainly created by AI can probably be copied by others without the Copyright Act giving you any remedy. Protection may then have to be sought in other ways, for example by registering a trade mark or a design.

For code, the picture is mixed. The parts written or substantially reworked by the developers are protected in the usual way. Code kept secret may also be protected as a trade secret, regardless of copyright. Read more in the article on trade secrets and AI tools.

What AI creates alone, nobody owns. What humans create with AI, you can own, but only if the contracts and documentation are in place.

The opposite risk is that the AI output resembles someone else’s protected work. Under Copyright Act § 81, an infringer may have to pay remuneration and damages, and even a party who acted in good faith may have to pay reasonable remuneration for the use.

What should the business do?

  1. Decide what needs to be protected. Use people for what is meant to be a lasting asset, such as the logo, core code and key texts, and use AI as a tool along the way.
  2. Document the human contribution. Keep drafts, versions and decisions that show who made the creative choices.
  3. Update contracts with consultants and agencies. Require disclosure of AI use, an express assignment of rights, a right to make alterations and a warranty against infringing the rights of others.
  4. Address copyright in employment contracts. Do not rely solely on non-statutory rules for works other than computer programs.
  5. Check output before commercial use. Images, logos and code in particular should be checked for similarity with others’ works and for licence terms.
  6. Adopt internal rules. Specify which tools may be used for what, in line with the business’s AI policy for employees.

The law is still developing, and both court decisions and legislative changes may alter the picture. More articles on the subject are available on the topic page on artificial intelligence.

Questions and answers

Can a competitor copy text or images we have created with AI?

If the content was made by AI without substantial human creative input, it probably has no copyright, and you cannot then stop copying under the Copyright Act. Other rules may still give protection, such as trade mark law, the Marketing Control Act's rules on good business practice, or contracts.

Is writing good prompts enough to obtain copyright?

Probably not. The US Copyright Office has concluded that prompts alone do not give sufficient control over the expression. The question has not been decided under Norwegian law, but the same result is likely under the requirement of original and individual creative effort.

Can we be liable if the AI output resembles someone else's work?

Yes. If you use output that reproduces protected parts of someone else's work, it may be an infringement. Under Copyright Act § 81, even a party acting in good faith may have to pay reasonable remuneration. Check output that is to be used commercially, particularly images, logos and code.

Next legal review: 10 November 2026