Who actually owns the creation? The legal dilemma between your prompt and artificial intelligence

The European regulation on artificial intelligence has more than 100 articles and 13 annexes. None of them says who the author is of a text, an image or a logo generated with an AI system.

The silence is not an oversight. It is a demarcation of competence. Regulation (EU) 2024/1689 (the AI Act), as amended by Regulation (EU) 2026/1744, is a product-safety and compliance act. It sets obligations for providers and users of AI systems, classifies risks and imposes transparency requirements. It is not an intellectual property instrument and does not set out to say to whom the generated output belongs.

For companies, agencies and creators, this is precisely the decisive question. A logo, a campaign or a text produced with AI has economic value only if it can be defended against copying and transferred by contract. The answer is not found in the AI Act, but in national copyright law, interpreted in the light of the case law of the Court of Justice of the European Union.

1. What the AI Act regulates in the field of copyright

The AI Act touches on copyright at two points, and neither concerns ownership of the output.

a) When models are trained. Under Article 53(1)(c), providers of general-purpose AI models must put in place a policy to comply with Union law on copyright. In particular, they must identify and comply with the reservation of rights expressed by rightholders under Article 4(3) of Directive (EU) 2019/790. Under point (d), the same providers must publish a sufficiently detailed summary of the content used for training.

b) At the level of the output. Under Article 50(2), providers of systems that generate synthetic content (audio, image, video or text) must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Separately, Article 50(4) requires users of systems that generate deepfakes to disclose that the content is artificial.

These are transparency obligations, not obligations that allocate rights. The fact that content is marked as artificially generated says nothing about who owns it or whether it is protected.

2. Ownership: the answer in national law (Law no. 8/1996)

The question “who is the author” remains within the competence of the national legislator, within the limits of the harmonisation achieved by the European directives and the case law of the CJEU.

Under Romanian law, Article 3(1) of Law no. 8/1996 on copyright and related rights defines the author as the natural person or persons who created the work. The text is categorical. An AI system is not a natural person and therefore cannot be an author. Nor does the legal person that developed or operates the system acquire the status of author merely because the system functions. At most, it may hold derived economic rights, in the cases expressly provided for by law.

Article 7 of the same law protects original works of intellectual creation in the literary, artistic or scientific field, whatever the method of creation, the manner or the form of expression. The wording “whatever the method of creation” shows that the use of a technical tool, including an AI system, does not in itself exclude protection. The condition, however, remains the same: the work must be original, and the originality must come from a human being.

The interim conclusion is simple. An output generated with AI can be protected only if, and only to the extent that, it reflects the person’s own human contribution, expressed in free and creative choices. AI can be a tool of creation, just like a camera or a graphic-editing program. A generic prompt or the mere acceptance of the output delivered by the model may, however, be insufficient.

3. The originality test and the place of the prompt

The notion of originality has been harmonised at Union level by the CJEU. Starting with the Infopaq judgment (C-5/08) and continuing with Painer (C-145/10), Football Dataco (C-604/10) and Cofemel (C-683/17), the Court has held that a subject matter is original where it is the author’s own intellectual creation and reflects the author’s personality, through free and creative choices.

The Brompton Bicycle judgment (C-833/18) is particularly relevant for AI. The Court held that where the shape of an object is dictated solely by technical constraints that leave no room for creative freedom, the originality requirement is not met.

Added to this is Article 9(a) of Law no. 8/1996, which excludes ideas, theories, concepts, discoveries and inventions from protection. Copyright protects the form of expression, not the idea.

A prompt is, as a rule, the expression of an idea: “a minimalist logo for a café, in shades of green”. Between that instruction and the final image, the model itself takes hundreds of decisions on composition, colour, proportion and style. The user does not control those decisions and often cannot even anticipate them. That is why a simple prompt does not make the person who wrote it the author of the output.

Nor does the effort invested change the conclusion. In Football Dataco, the CJEU made clear that significant labour and skill do not, in themselves, justify protection if they do not express originality through free and creative choices. A prompt refined over dozens of iterations does not, merely through the effort put in, become the basis for authorship of the output.

4. The gradation of human contribution

The analysis is not binary. In practice, three situations emerge, with different legal consequences.

Situation I: a simple prompt, output used as is. The user formulates an instruction and takes the output without significant intervention. There are no free and creative choices of the user’s own regarding the form. As a rule, there is no protected work, and the output lies, in practice, outside the scope of copyright.

Situation II: iterative prompts, selection and combination of variants. The user refines the instructions over dozens of iterations, chooses between variants and combines elements from several outputs. This is a grey area. Selection and arrangement may, under certain conditions, constitute an original contribution, by analogy with derivative works and collections protected under Article 8 of Law no. 8/1996. The difficulty is one of proof: it must be shown that the choices were creative and not merely preferences between options offered by the model.

Situation III: AI used as a tool, with substantial human processing. The generated output is only working material. The user transforms it substantially through drawing, composition, rewriting or integration into a larger work. Protection may exist, at least for the identifiable human contribution. Elements that are generated and taken over unchanged, however, remain unprotected.

Two clarifications are needed. The prompt itself may be protected as a literary text, if it is sufficiently elaborate and original, but its protection does not extend to the generated output. Second, in litigation, the burden of proof lies with the party claiming authorship, under the general rule in Article 249 of the Code of Civil Procedure. The presumption of authorship in Article 4 of Law no. 8/1996, which operates in favour of the person under whose name the work was made known to the public, can be rebutted by proof that the work was generated by an AI system without creative human contribution.

An example. An agency delivers to a client a logo generated from a single prompt, with minimal retouching, and assigns to it by contract “all copyright”. A few months later, a competitor uses an almost identical graphic. The client wants to sue for copyright infringement, but must prove the existence of an original work and a creative human contribution. If it cannot do so, the assignment had no subject matter: the agency had nothing to transfer. The client is left without copyright protection, and the next question is whether the agency is contractually liable under the warranty it gave.

5. The reverse risk: infringing the rights of others

The fact that the user is not the author of the output does not shield them from liability. If the output reproduces, in whole or in part, protected works of third parties (a character, a photograph, a fragment of text, a recognisable composition), its commercial use may infringe the economic rights of the rightholders. Liability does not depend on knowing that the model “memorised” the work in question.

6. The contractual dimension

a) Platform terms. Many AI providers state that the user “owns” the outputs or that they assign their rights in them. Such clauses are of limited value. A contract can transfer only rights that exist. It cannot create a copyright where the law does not recognise one, in line with the principle nemo plus iuris ad alium transferre potest quam ipse habet. The clause has effect between the parties, for example through the provider’s waiver of any claim, but it is not enforceable against third parties as an intellectual property title.

b) Contracts with agencies and freelancers. These contracts should expressly regulate whether and to what extent AI may be used, the obligation to disclose generated elements, warranties of originality and of non-infringement of third-party rights, and liability under the warranty against eviction. A general assignment of “all rights” is no longer sufficient where part of the deliverable may not be protected at all.

c) Works created by employees. Under Article 45 of Law no. 8/1996, unless the individual employment contract provides otherwise, the economic rights in works created in the performance of the employee’s duties belong to the employee-author. The employer may nevertheless use the work within the scope of its business. For the part generated exclusively by AI, however, there may be no work at all and, therefore, nothing to transfer. Intellectual property clauses in employment contracts should be reviewed from this perspective.

7. Practical recommendations

• Document the human contribution. Keep the intermediate versions, the prompts, the rejected variants and the changes made by hand. In litigation, this is the proof of authorship.
• Review the terms of the tool you use. Check what rights the provider reserves, what warranties it gives and whether it offers indemnities against third-party claims.
• Update your contracts. Include clauses on the use of AI in contracts with agencies, freelancers and employees, with warranties and liability proportionate to the risk.
• Check the output before commercial use. Especially for visual identity elements, campaigns and content published at scale.
• Consider other forms of protection as well. A logo can be protected as a trademark, independently of copyright, and know-how can be protected as a trade secret.
• Do not assume you hold an exclusive right in content generated entirely by AI. Without a creative human contribution, that content may be used freely by anyone.

Conclusion

The AI Act and Law no. 8/1996 operate at different but complementary levels. The first regulates systems, risks and transparency. The second determines whether the output is a work and who its author is. Between them remains a space that every company must manage through documentation, contracts and verification.

The prompt belongs to you. The output, not necessarily.