BEYOND CHAPTER 03 · Explore the changes

AI art: separate the legal questions

If an AI-assisted image can be sold, does that mean it has copyright protection and its creation and use infringe no one else’s rights?

In this chapter

Discussions of AI images often combine four questions: whether training material may be used, whether the result contains protectable authorship, whether publishing it infringes third-party rights, and what a service’s terms permit. A favourable answer to one question does not automatically resolve the other three.

For artists, museums and collectors, a useful starting point is to identify specific human contributions, preserve process evidence and establish which rules govern use and transactions. The law can address registration, copying, attribution and licensing without first deciding whether AI is “creative”. Protection may also cover only some elements of a work.

Identify the protected expression

Section 102 of the US Copyright Act starts with original works of authorship fixed in a tangible medium, excluding ideas, procedures and methods themselves. Originality is neither novelty nor a measure of hours worked. In Feist (1991), the facts and alphabetical arrangement at issue lacked the required creativity. Effort collecting information did not confer exclusive rights over facts; repeated generation likewise is not a sufficient legal argument by itself.

Mechanical assistance does not automatically exclude human authorship. The Oscar Wilde photograph in Burrow-Giles v. Sarony (1884) involved the photographer’s arrangements of pose, clothing, light and composition. The useful question is which expressive choices came from a person. Cameras and generative models offer different forms of control, so describing both as tools does not establish identical legal treatment for every resulting image.

Nature and the limits of nonhuman authorship

Kelley v. Chicago Park District (2011) concerned the living garden Wildflower Works. The US Seventh Circuit considered natural growth and change and found the garden lacked the authorship and fixation needed for copyright. That conclusion does not mean garden plans, photographs or every work using natural materials are unprotected. The object for which protection is claimed must first be identified.

A central issue in Naruto v. Slater (2018) was whether an animal could bring an infringement claim under US copyright law. The Ninth Circuit found no statutory standing for animals; this did not establish that the photographer necessarily owned copyright in the monkey selfie. Gardens, animals and AI cases address different legal requirements, rather than supporting a rule that any nonhuman participation defeats protection.

After Thaler, examine the human contribution

Thaler sought US registration for A Recent Entrance to Paradise naming a system as sole author and describing autonomous creation. A district court upheld refusal in 2023; the DC Circuit affirmed on 18 March 2025 that a work must initially have a human author. The Supreme Court denied review on 2 March 2026, leaving the lower decision standing without issuing its own merits ruling.

That premise differs from a human making a mixed work with AI assistance. The Copyright Office’s 2025 copyrightability report says prompts alone generally are insufficient, while perceptible human expression, creative selection or arrangement, and modifications may qualify. Owning a system or paying for access does not establish authorship. Work-made-for-hire rules have statutory conditions and cannot turn a machine into an employee to bypass the human-authorship requirement.

Mixed works can receive partial protection

In its 2023 Zarya of the Dawn decision, the US Copyright Office recognised Kristina Kashtanova’s text and the selection, coordination and arrangement of text and images, while excluding the Midjourney images themselves. A narrower registration replaced the earlier one. The book’s arrangement and each image therefore require separate analysis: unprotected material does not erase human-authored text and arrangement.

That year, Théâtre D’opéra Spatial was refused as submitted because Jason M. Allen would not disclaim more-than-minimal AI material, rather than because every Photoshop edit lacked creativity. In SURYAST, Ankit Sahni’s photograph, style reference and strength setting were insufficient to establish authorship of the generated result; that did not extinguish possible rights in the original photograph. These decisions concern the claimed scope, rather than prompt counts or prizes.

China and the UK: different legal approaches

In Li v. Liu on 27 November 2023, the Beijing Internet Court recognised human authorship of the particular AI-assisted image based on prompts, parameter changes, selection and expression. Shenzhen’s 2019 Dreamwriter case also concerned team choices and arrangements behind automated financial reporting. Neither establishes copyright for AI itself or guarantees protection for other outputs; the reasoning connects the claim to human or organisational creative activity.

UK Copyright, Designs and Patents Act 1988 section 9(3) attributes authorship of specified computer-generated works to the person making the necessary arrangements; section 178 defines that category by the absence of a human author. It differs from ordinary AI-assisted creation. The government’s March 2026 report proposed monitoring this protection and removing it absent evidence of value, but a proposal is not repeal. Necessary arrangements and other protection requirements still need analysis.

Training permissions differ from output copyrightability

Whether an output has copyright does not establish permission to copy training materials. US fair use depends on the particular use and statutory factors, rather than a blanket rule that all training is lawful or infringing. The Copyright Office’s 2025 training report is not a judgment and remained labelled a pre-publication version when checked. Acquisition, training use, outputs and market effects require distinct evidence.

EU Digital Single Market Copyright Directive Articles 3 and 4 provide distinct text-and-data-mining exceptions involving lawful access, research purposes, eligible institutions and rights reservations. UK section 29A instead addresses qualifying non-commercial research, not an automatic permission for general commercial training. A museum possessing a high-resolution image may not control every right in the artwork, photograph and associated personal data needed for a vendor’s proposed use.

Check third parties and contracts when using outputs

Even an output without its own copyright may implicate others’ works, trademarks, likenesses, voices, privacy or misleading representations; applicable rights vary by place. Conversely, exposure to an artist’s work during training does not make every output infringing. Examine specific protected expression, people and proposed uses rather than deciding solely from stylistic resemblance or the involvement of AI.

A platform’s commercial-use permission is usually a contract between user and provider; it cannot create copyright where law provides none or license third-party rights. Commission agreements should address permitted tools, input clearance, transfers of human-authored material, exclusivity and complaints. Confidentiality, reuse for training, indemnity scope and service termination also matter. A promise that “you own the output” cannot settle all these issues.

EU AI rules address transparency, not copyright grants

The EU AI Act requires covered general-purpose AI model providers to maintain a copyright-compliance policy and publish a training-content summary, subject to the applicable model and timetable. These obligations support questions about training practices; they do not certify permission for every input or grant copyright to all AI images. A cultural institution deploying another provider’s system must distinguish its role from the model provider’s.

Article 50 transparency obligations apply from 2 August 2026, with transitional arrangements for machine-readable marking by some existing systems. Direct interaction, synthetic-content marking and deepfake disclosure have different conditions. Artistic, satirical and fictional works receive specific treatment, not a blanket exemption. For an exhibition, identify providers, deployers, content type and presentation before deciding what disclosure is required.

Make the process explainable through evidence

A practical process file can retain original photographs or drawings, input sources and permissions, model versions, significant generation stages, human editing layers, arrangement decisions and approvals. Its purpose is to show who made which expression, not to exchange document volume for rights. Collaborative projects also need records of contributions, ownership arrangements and agreements about later modification.

US Copyright Office registration guidance calls for disclosure of human contributions and exclusion of more-than-minimal AI-generated material; registration and litigation procedures elsewhere may differ. Exhibition labels, certificates and sale terms should describe production consistently, distinguishing artistic attribution, control of an account and full copyright ownership. Significant cross-border commissions or disputes require assessment against their actual facts and local rules.

Your turn to investigate

Which clue would you check first?

An artist seeks US registration for a book with self-written text, human arrangement and AI-generated images. Which approach best fits the issues?

Choose a clue to see what it can tell you.

READING COMPANION

A few useful terms

Human authorship: Protectable expression contributed by a person; US analysis neither excludes all tool-assisted work nor treats mere operation as sufficient.
Originality: In US copyright analysis, independent creation with a minimal degree of creativity, rather than novelty, expense or time spent.
Fixation: The US requirement that expression be embodied sufficiently stably in a medium to be perceived, reproduced or communicated for more than a transitory duration.
Scope of a claim: Identifying which text, images, modifications or arrangements are claimed and which materials are excluded.
Text and data mining: Automated analysis of text or data to obtain information such as patterns; exceptions depend on jurisdiction and conditions.
Fair use: A US copyright limitation assessed through factors including purpose, nature, amount and market effect, not a universal permission for training.
Computer-generated work: A specific UK category for work generated by a computer without a human author, distinct from all computer-assisted creation.
Licence: Permission to use rights under specified conditions; a licensor can grant only rights it is entitled to license.
Deepfake: AI-generated or manipulated image, audio or video that may appear authentic; legal definitions and obligations vary.

Explore primary and secondary markets

YOUR TURN

If platform terms say that the output belongs to you, what would you still ask the commissioned artist?

Consider another way to think about it

Ask about human creation and editing, inputs and permissions, the version of tool terms, third parties and whether the commission needs display, adaptation, exclusivity or sublicensing. Clear requirements reveal which rights the artist and platform can actually provide.

TAKE THIS THOUGHT WITH YOU

Assess authorship, training legality, output use and contractual rights separately, reading every conclusion in its jurisdiction and date.

Reading notes & references

Legal sources checked on 2 October 2026. Cases are identified by jurisdiction, procedure and decision date; administrative registration decisions, judgments and policy proposals serve different roles.

Connect the ideas

Continue into people, works and research.