Creative AI examples range from a Go move to an auction portrait to an award-winning image, and in every documented case the system supplied surprising material while people set the goal, chose among outputs, and took responsibility for the result. In the United States, the Copyright Office's report of January 29, 2025 concludes that purely AI-generated material is not copyrightable, while human selection, arrangement, and modification can be.
Key takeaways
- AlphaGo's Move 37, played in March 2016, is the clearest case of a system finding a strong idea that experts did not expect, but it happened inside a game with a fixed scoring rule.
- A portrait made with a generative adversarial network sold for USD 432,500 at Christie's on October 25, 2018, after people chose the data, the model, and the final print.
- The U.S. Copyright Office states that, with current technology, prompts alone do not give enough control to count as authorship.
- Canada has no settled answer yet: a federal consultation found that most participants wanted human authorship to stay central to copyright.
A move that experts did not expect
In March 2016, DeepMind's AlphaGo beat Lee Sedol four games to one in Seoul. DeepMind's own account singles out Move 37 of game two, which it describes as a move with a one in 10,000 chance of being used by a human player. The move helped AlphaGo win the game. In game four, Lee Sedol answered with Move 78, which DeepMind says had similarly long odds and helped him win that game.
Go gives a clear test for creativity of this narrow kind. A move is either part of a winning line or it is not, so novelty can be judged against a result. Most creative work has no such referee, which is why the later examples below are harder to settle.
A portrait sold at auction
"Edmond de Belamy" was made by Obvious, a Paris collective of three people (Hugo Caselles-Dupré, Pierre Fautrel, and Gauthier Vernier), using a generative adversarial network, or GAN. A GAN pairs two programs. A generator produces images, and a discriminator tries to tell them apart from real ones. The Obvious model learned from a dataset of 15,000 portraits painted between the 14th and 20th centuries.
Christie's sold the print on October 25, 2018 for USD 432,500, roughly 45 times the pre-sale estimate, and promoted the sale as the first auction to feature work created by artificial intelligence. Reports differ on the exact estimate, so only the ratio is repeated here. The sale shows how much of the work lies outside the algorithm: selecting training images, tuning the model, choosing one output out of many, and deciding to print and sell it.
Who contributed what
The table separates what each system produced from what people decided. It is a simplified reading of the sources cited below, not a legal opinion.
| Example | What the system produced | What people decided |
|---|---|---|
| AlphaGo, Move 37 (2016) | A move judged highly unlikely for a human | The game, the training setup, and the scoring rule |
| Edmond de Belamy (2018) | An image generated from learned portrait patterns | The dataset, the model, the selected output, and the sale |
| Théâtre D'opéra Spatial (2022) | An initial image from Midjourney | At least 624 prompts and revisions, then edits in Photoshop and an upscaling tool |
| Zarya of the Dawn (2023) | The comic's images, generated with Midjourney | The text, plus the selection, coordination, and arrangement of text and images |
What the U.S. Copyright Office decided
The last two rows of the table were tested by the U.S. Copyright Office. For Zarya of the Dawn, a decision of February 21, 2023 covered the human-written text and the human arrangement of the work, but not the Midjourney images. For Théâtre D'opéra Spatial, which had won a prize at the 2022 Colorado State Fair, the Office's Review Board affirmed a refusal on September 5, 2023, because the applicant, Jason Allen, would not disclaim the AI-generated content.
The Office's Part 2 report on copyrightability, published on January 29, 2025, generalizes these decisions. It concludes that AI used as an assistive tool does not affect protection, that purely AI-generated material is not protected, that prompts alone do not currently provide sufficient control, and that human authors keep rights in perceptible human expression, creative selection and arrangement, and creative modifications. Each case is judged individually.
A related lawsuit asked a different question: whether an AI system can be listed as the sole author. The Supreme Court declined to hear that case, Thaler v. Perlmutter, on March 2, 2026, leaving the lower-court rulings against AI authorship in place.
The Canadian position
Innovation, Science and Economic Development Canada ran a consultation on copyright and generative AI from October 12, 2023 to January 15, 2024. Its published summary of the consultation says about 1,000 people answered the online questionnaire. Most participants supported keeping human authorship central to copyright, and many held that only AI-generated content with enough human contribution should be protected. Opinion was split on whether current law is flexible enough or needs clarification. The page does not announce an amendment, and anyone relying on a creative AI work in Canada should check the current Copyright Act and take advice from a lawyer.
Questions to ask about any creative AI claim
- Is there a measurable goal, as in Go, or is quality a matter of taste?
- Who chose the training data, and did the creators have the right to use it?
- How many outputs were discarded before one was shown?
- Which parts were edited, arranged, or written by a person?
- Is the claim about art, about law, or about marketing?
This article offers general information and is not legal advice. Registration and licensing questions depend on the facts of a specific work.
FAQ
Did AI really create the Edmond de Belamy portrait?
A trained GAN generated the image, but the collective selected the data, ran the model, picked an output, and prepared it for sale. Authorship is therefore shared between people and software.
Can AI-generated art be copyrighted in the United States?
Material generated entirely by AI cannot. A work that combines AI output with human expression can be protected for the human contribution, which the Copyright Office assesses case by case.
Was AlphaGo's Move 37 creative?
It was novel and it was effective, which meets a practical test. Whether that equals human creativity is a philosophical question that the sources here do not settle.
Does Canadian law treat AI works differently?
The consultation record shows broad support for human authorship, but the report reviewed here does not announce new rules. The existing Copyright Act and any later amendment control.
Related reading
- Can AI be truly creative?
- Ethical limits of AI-generated art
- AI tools for making and refining art
- Humans versus AI on divergent thinking
More in AI Basics.
Last verified: October 2026
Sources
- AlphaGo, Google DeepMind, 2016 and later page updates
- Christie's Is First to Sell Art Made by Artificial Intelligence, But What Does That Mean?, Smithsonian Magazine, 2018
- Copyright and Artificial Intelligence, Part 2: Copyrightability, U.S. Copyright Office, January 2025
- Second Request for Reconsideration for Refusal to Register Théâtre D'opéra Spatial, U.S. Copyright Office Review Board, September 5, 2023
- Artificial Intelligence and Copyright (Zarya of the Dawn and registration decisions), U.S. Copyright Office, 2023 to 2025
- Docket 25-449, Thaler v. Perlmutter, Supreme Court of the United States, March 2, 2026
- Consultation on Copyright in the Age of Generative Artificial Intelligence, summary report, Innovation, Science and Economic Development Canada, 2025