The law has advanced a relatively stable response regarding the authorship of results generated through artificial intelligence systems: without sufficient human creative contribution, there is no work protected as a creation of its user. That response is narrow because copyright law needs to identify human creation. It does not explain what happens with the economic and cultural value that an output produces even if no one can claim it as their own work.
An unprotected image can replace a commissioned illustration. A synthetic track can displace a musical contract. A text without a legal author can capture visits, advertising or subscriptions. The lack of protection does not imply a lack of efficacy. The temptation is to call that value surplus. Within the framework of structural surplus, it is not.
What the law allows to be affirmed
Article 5.1 of the consolidated text of the Spanish Intellectual Property Law considers the natural person who creates a work to be the author. The starting point is not the ownership of an account or the property of a tool, but human creation.
The European Union has not approved a specific copyright for results generated by AI. The analysis by the European Parliamentary Research Service describes a comparative trend: content with minimal or no human creativity remains outside of protection; assisted works may protect the elements that express sufficient human creative decisions. The evaluation is case-by-case and is not resolved by counting prompts.
The European Parliament Resolution of 10 March 2026 on copyright and generative artificial intelligence is not a new law. It is a political position that calls for the effective application of the existing framework, transparency regarding materials used, licensing mechanisms, respect for rights reservations, and remuneration for rights holders. It also requests a study into the uncertainties regarding outputs and potential reproductions.
In the United States, the Copyright Office maintains a similar distinction, published in January 2025 in the second part of its report on artificial intelligence and previously applied to specific cases. In February 2023, it limited the registration of the comic Zarya of the Dawn, by Kristina Kashtanova, to the text and the selection and arrangement of elements, while excluding images produced with Midjourney. The general rule is as follows: human expression perceptible in a work that includes generated material may be protected, such as selection, coordination, arrangement, or creative modifications; the material produced automatically by itself may not. The US Supreme Court declined to review Thaler v. Perlmutter on 2 March 2026, and therefore did not alter the requirement for human authorship upheld in that litigation.
Convergence is not universal. In November 2023, the Beijing Internet Court recognised copyright over an image generated with Stable Diffusion because it considered that the user's choices in prompts and parameters expressed sufficient intellectual contribution. A similar result may receive opposing responses depending on the jurisdiction, and those distributing works across multiple markets cannot consider the matter resolved by the response of a single one.
These criteria do not establish that an assisted work lacks protection. They require separating contributions: what a person created and what the system produced. An extensive prompt may document intention and process, but it does not by itself guarantee expressive control of the result.
For a specific legal development, consult Artificial intelligence and copyright in Europe: what is decided in 2026 and Authorship and works made with AI in Spanish law.
Value without protection
The substitution of commissions for unprotected outputs is developed in What competes without being a work. Here another question matters: how labour is captured and rent is generated even when the result does not meet the requirements to be a protected work.
The market does not need every result to be a protected work to assign it a function. It can purchase access to a service, commission production volume, pay for speed or use an output within a larger product.
Substitution takes several forms:
- A company generates an image instead of commissioning an illustration.
- A studio reduces sketching or previewing phases.
- A platform offers functional music at scale.
- A media outlet automates low-differentiation texts.
- A user produces variants whose utility depends on their immediate availability.
There is a relevant Spanish case. According to the 2024 Tebeosfera Report, the University of Salamanca sponsored an exhibition in September 2024 which resulted in a comic of approximately 140 pages produced with artificial intelligence, without commissioning a professional in the sector. Author collectives protested, and the university maintained to the press that the work was the result of prior work by researchers and lecturers. The episode illustrates substitution better than a projection: a commission that could have existed did not occur, and the subsequent debate revolved around authorship, not displaced labour. I am aware of this through the report's summary, not through the university's documentation.
Not every substitution eliminates a complete job. It can reduce rates, hours, initial commissions or market entry capacity. It can also create integration, editing and supervision work. The issue is not resolved with an isolated net figure: it matters who captures the savings and who assumes the loss of bargaining power.
The study by CISAC and PMP Strategy published in 2024 projected for 2028 risks equivalent to 24 % of the income of musical creators and 21 % of audiovisual creators, with 22 billion euros accumulated between 2024 and 2028. These are estimates, not measurements of lost commissions. They must be read as scenarios constructed with hypotheses of adoption, substitution and impact, not as verified losses.
The absence of complete measurements does not eliminate the problem. It compels us not to present forecasts as facts and to build better data on rates, volume of commissions, employment, licences and income distribution.
The other end of the chain
Litigation is concentrated on the materials used to train and operate systems. European and American controversies discuss mining exceptions, rights reservations, copies, lawful access, memorisation, reproduction in results and provider liability.
The cases do not yet form a uniform doctrine. Like Company v Google, C-250/25, remains as a preliminary ruling before the Court of Justice of the European Union. The litigation of GEMA against OpenAI produced a first-instance decision in Munich in November 2025. Getty Images v Stability AI concluded in the British High Court with the rejection of the copyright claims that reached judgment and a limited trademark infringement. The Higher Regional Court of Hamburg dismissed Robert Kneschke's appeal against LAION in December 2025; the litigation remained pending before the German Supreme Court at the time of writing. In the United States, Bartz v Anthropic concluded through a settlement approved in 2026.
They should not be combined as if they answered a single question. Training, obtaining copies, scientific or commercial purpose, jurisdiction, production of results and trademarks are distinct planes. The only certain conclusion is that the legal cost of the system does not disappear because a specific output lacks an author.
Transparency is not distribution
From 2 August 2026, the obligations of Article 50 of the European AI Act apply, with a four-month transitional period for the marking of generative systems already marketed to certain providers and deployers, including measures relating to machine-readable marking of synthetic content. Its scope and exceptions depend on the type of content and use.
One year earlier, from 2 August 2025, Article 53 of the same regulation obliges providers of general-purpose models to implement a copyright compliance policy that respects the reservations expressed in accordance with Directive (EU) 2019/790, and to publish a sufficiently detailed summary of the content used for training, according to a template that the Commission published in July 2025.
Marking an output assists in identifying its technical provenance. It does not, in itself, inform about the works utilised in training, the workers who classified data, the energy consumed, the displaced fees, or the economic distribution. Surface transparency and infrastructure transparency are distinct issues.
The training summary goes a step further: it states the type of sources a model was nourished by. It does not distribute anything, either. Knowing that a corpus included works from a sector does not inform how much value was extracted from them, nor to whom a portion corresponds. Transparency is a condition for negotiation, not its outcome.
In Spain, the Ministry of Culture promoted a royal decree at the end of 2024 regarding the use of works by artificial intelligence systems, which, according to information published at the time, regulated extended collective licensing with the support of management entities. In January 2025, it withdrew its processing due to a lack of consensus and opened a dialogue with the sector. I record this based on the summary of the 2024 Tebeosfera Report and the press it cites; I have not consulted the text of the draft.
The same applies to attribution. Naming the user, the provider, or the model may be useful for information purposes. It does not remunerate the individuals whose works or labour were involved, nor does it determine who is liable for damages.
Why it is not surplus
What follows is a thesis, not a legal description.
I term structural surplus a relational difference that reorganises the field of a system's possible configurations without being absorbed by its primary function. A commercial generative system is organised to produce usable outputs in response to an instruction. When an image effectively replaces a commission, it fulfils that function. Its value may be considerable, but it does not exceed the system: it confirms its performance.
Value without an author is performance without an authorial subject, not an absence of labour. It condenses:
- Labour incorporated into prior works and data.
- Technical development and maintenance labour.
- Labelling, moderation, and evaluation.
- Computing and distribution infrastructure.
- Labour of the user who selects, edits, or integrates.
- Labour displaced from the commission that ceases to be produced.
Copyright observes creation and ownership. The market records costs, sales, and substitutions. Neither requires that set to be recognised as a work for it to produce value.
An economic category describes this value more accurately: rent. Whoever controls access to an infrastructure and an accumulated corpus can charge for its use without having produced everything that use contains. This is my own analogy and it has limits: classical rent refers to a scarce resource, whereas generative models produce abundance. Yet it captures something the word surplus conceals: value does not spring from the encounter; it is charged at the point of access.
Calling it surplus would ennoble the platform's efficiency by presenting it as an opening that transcends its purpose. An artistic event may exist within a practice involving generative systems, but it does not automatically derive from the fact that the result lacks an author or protection.
Distributed authorship does not resolve everything
Distributing authorship among programmers, datasets, the platform, and the user appears to acknowledge the complexity. However, it conflates contributions of a different nature. Whoever develops an infrastructure, creates a work included in the data, adjusts a model, or selects a result does not necessarily perform the same creative act with respect to the final output.
Contribution frameworks are useful for distributing credit, remuneration, and liability. They should not be used to manufacture a generic collective authorship that erases differences. At times there will be an assisted human work; at others, generated material without protection; at others, a composite work incorporating elements of a different status.
Documentation assists in separating these planes. Archiving and conservation of a work created with artificial intelligence proposes registering materials, tools, versions, decisions, and dependencies. That record does not guarantee rights, but it prevents the entire process from being reduced to the result.
Two questions that should not be confused
The debate contains at least two questions:
- Who is liable and who is paid. This is a legal, contractual, labour, and political matter. It includes licences, remuneration, liability, working conditions, competition, and distribution.
- Whether something occurred in the encounter that the function does not reabsorb. This is an aesthetic and ontological question that may be formulated before a painting, a generative work, or an interaction between machines.
Mixing them produces two errors. If the capture of labour is termed surplus, the platform's performance acquires an artistic depth it has not demonstrated. If only that which produces surplus is protected, the captured labour remains without a response because it does not meet an ontological requirement that labour or property law does not need.
Rights should not depend on a practice achieving artistic value. And artistic value should not be deduced from a system producing profitability without an author.
On the open conversation
This text distinguishes a provisional legal description from a thesis on structural surplus. If you work with data on contracting, fees, or effective substitution that would allow for a better measurement of the impact of generative systems on creative labour, you may write to me.
Sources
- Consolidated Text of the Intellectual Property Law.
- European Parliament, Resolution of 10 March 2026.
- EPRS, Copyright and AI-generated content, 2026.
- U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2.
- Supreme Court of the United States, Thaler v. Perlmutter.
- CISAC and PMP Strategy, economic study on generative AI.
- CJEU, Like Company v Google, C-250/25.
- Munich Regional Court I, GEMA v OpenAI.
- High Court, Getty Images v Stability AI.
- Regulation (EU) 2024/1689, AI Act.
- Directive (EU) 2019/790, on copyright in the Digital Single Market.
- U.S. Copyright Office, resolution on Zarya of the Dawn, 21 February 2023; Beijing Internet Court, Li v. Liu, 27 November 2023. References without verified link.
- Asociación Cultural Tebeosfera, 2024 Tebeosfera Report. The Triumph of Diversity.