The Ethics of AI Art Cannot Be Measured in Millions of Images
Refik Anadol wants the art world to accept a reassuring distinction: other AI companies may exploit artists, but his studio begins with permission. In a 60 Minutes interview, he says he has used ethically sourced datasets since 2020, researches where information comes from and makes permission the first step. Yet the programme never establishes what those permissions cover, who granted them, or whether the underlying material can legally be used to train commercial AI systems.
That is not a minor omission. Anadol's reputation, his multimillion-dollar art market and his proposed AI museum all benefit from the public accepting that his approach is ethically different. An assurance is not an audit, and a carefully curated dataset is not automatically a licensed one.
The 153 million-image question
The strongest test case is not an abstract debate about whether AI can make art. It is Artificial Realities: California Landscapes, a 2024 project built from 153 million publicly available images of California's national parks.
That figure is confirmed by both Refik Anadol Studio and the official LUMINEX exhibition press release. But neither public project description identifies the individual image sources, photographers, copyright holders, licence conditions or permissions covering commercial AI training.
Refik Anadol+1
There is a critical distinction here: publicly available does not mean copyright-free, and it certainly does not automatically mean authorised for commercial AI training.
The U.S. Copyright Office makes this distinction explicitly in its 2025 report on generative AI. It warns that material described as publicly available may simply mean that it can be found on the internet. The report discusses how internet scraping can incorporate copyrighted works without the knowledge or permission of their owners.
U.S. Copyright Office
Anadol may have obtained permission for all 153 million images. I have found no evidence establishing otherwise. But the public documentation does not allow an independent reader to verify that claim.
That is precisely the question 60 Minutes should have asked.
NASA does not automatically clear 200 million images
The other enormous dataset deserves the same scrutiny.
For his Earth imagery project, Anadol used approximately 200 million landscape images. Google Arts & Culture describes the collection as drawing on satellite data and found open-source imagery. The 60 Minutes report also identifies NASA as an important source of information.
Google Arts & Culture+1
There is a legitimate route to using much of NASA's scientific material. NASA's Science Data Portal states that data from NASA-led missions generally carry a Creative Commons Zero licence unless a restrictive notice applies.
But NASA also makes an important qualification: other data may have different rights, NASA may not be their original source, and users should validate the original source and associated usage rights.
Science Data Portal
That distinction matters when an artwork combines satellite observations with hundreds of millions of images gathered from other sources.
A NASA connection is not a blanket licence for every photograph in a mixed dataset.
I would want to see the original dataset manifest, source repositories, applicable licences and any permission agreements. Without those records, neither the accusation of theft nor the assurance of ethical sourcing has been established.
The MoMA project actually provides evidence in Anadol's favour
There is an important qualification that the copyright debate often misses.
Anadol's 2022 installation Unsupervised at the Museum of Modern Art used publicly available information about the museum's collection. Former director Glenn Lowry described the use of metadata to reconstruct relationships across more than 200 years of art.
MOMA Press+1
MoMA's own API terms clarify something particularly important: its collection metadata is released under a CC0 public-domain dedication, and images are not included in that dataset.
MoMA API
That makes Unsupervised a very different case from an image generator trained on unauthorised copies of copyrighted paintings.
The distinction matters. We should not accuse Anadol of stealing MoMA's artworks when the documented training material was publicly released metadata.
It also demonstrates that responsible AI art can be built using appropriately released institutional data.
Dataland has provided more evidence since the interview
The 60 Minutes programme presented Anadol's 20,000-square-foot AI museum, Dataland, as a project under construction. That is now outdated. Dataland opened in downtown Los Angeles on June 19, 2026.
Refik Anadol
More importantly, on October 7, 2026, Dataland published a substantial account of how its Large Nature Model collects and processes ecological information.
The studio describes a dataset containing more than 1.2 billion ecological data points, environmental recordings, scientific observations and field collections. It identifies:
- Field expeditions beginning in 2023 across Australia, Indonesia and California.
- LiDAR scanning and photogrammetry to capture forests in three dimensions.
- Environmental recordings of birds, rainfall, insects and streams.
- Sensor measurements covering rainfall, soil moisture, atmospheric conditions and plant-water relationships.
- Named research contributors, including Professor Kim Calders of Ghent University, TERN Australia and researchers at the University of Chile.
Dataland
This is meaningful evidence of actual research and first-hand data collection. It would be dishonest to dismiss the entire operation as indiscriminate internet scraping.
But there is another distinction.
Documenting where a dataset came from is not the same as documenting the rights to use every item in it.
Anadol has also identified partnerships with the Smithsonian Institution, National Geographic and London's Natural History Museum.
upov+1
Those institutions have different rules for different materials. The Smithsonian says much of its collection data is available under CC0, but individual images can carry restrictions. Its Department of Botany, for example, prohibits commercial use of images that are not specifically marked CC0 without written permission. National Geographic's Special Collections similarly restricts downloading and makes access dependent on individual usage rights.
Smithsonian National Museum of Natural History+2
This does not establish that Anadol used restricted material. It establishes why institutional partnerships alone cannot prove that every image in a training dataset is cleared.
The October Dataland publication is a welcome improvement in transparency. It is also a publication by the organisation whose practices are being examined, not an independent rights audit.
If Dataland wants to establish a global standard for ethical AI art, publishing a dataset-level rights register would be a much stronger demonstration than publishing another enormous number.
Molly Crabapple's demonstration is not proof of copyright infringement
The programme also presented an illustration by Molly Crabapple depicting Aleppo, Syria, and compared it with an AI-generated image requested in her artistic style.
The output was visually similar. It was a compelling demonstration of how AI can imitate an artist's visual language.
But it did not establish that the particular Aleppo illustration had been included in the model's training data. Nor did it establish that the generated image reproduced protected elements of her original work.
This distinction matters legally.
The U.S. Copyright Office explains that artistic style, by itself, is not protected by copyright. However, stylistic imitation can still damage an artist's market, and an AI output that reproduces substantially similar protected expression may constitute infringement.
U.S. Copyright Office+1
The demonstration therefore supports Crabapple's concern about competition and imitation. It does not prove her specific allegation that her illustration was stolen.
Her broader statement that no artist has ever received compensation also deserved qualification. The Copyright Office documented existing licensing arrangements involving image libraries, music, voice recordings and news publishers, including Shutterstock and Getty-related partnerships.
Those agreements do not establish that Crabapple was compensated or that every artist whose work entered an AI dataset was treated fairly. But they make the categorical claim that no compensation has ever occurred too sweeping.
U.S. Copyright Office
The programme allowed a striking visual comparison to carry more evidential weight than it could support.
Fair use is not a universal permission slip
AI companies have argued that training on copyrighted material can constitute fair use. Artists have argued that commercial training without permission deprives them of control and damages their markets.
Neither position can be settled by a slogan.
The U.S. Copyright Office's May 2025 report on generative AI training recognises the potential public benefits of AI while warning that unlicensed training can cause substantial market harm. It also identifies licensing as an existing and developing alternative.
Importantly, the report does not establish a universal rule that all AI training is either lawful or unlawful. It says the answer depends on the circumstances of each use.
U.S. Copyright Office+1
Two 2025 California district court decisions illustrate the distinction.
In Bartz v. Anthropic, Judge William Alsup treated training on lawfully purchased books as fair use but rejected the argument that downloading and retaining pirated books in a permanent general-purpose library was also fair use.
Justia Dockets & Filings
In Kadrey v. Meta, Judge Vince Chhabria granted Meta summary judgment on the claims of thirteen authors because their evidence did not establish the necessary market harm. He expressly warned that the decision did not establish that Meta's use of copyrighted material was universally lawful.
Justia Law
Neither case decided whether Anadol's image datasets were lawfully assembled. Both demonstrate why the provenance, purpose and actual use of individual datasets matter.
The 60 Minutes interview could have explained this developing legal landscape instead of leaving viewers with a simplified argument between fair use and theft.
Anadol understands the importance of copyright when he is the rights holder
There is one particularly revealing aspect of Anadol's own position.
In a December 2024 interview with the World Intellectual Property Organization, he was asked whether he objected to people copying his work without permission.
He answered yes.
He explained that unauthorized copying hurt because his studio spent years developing projects, and that he wanted to protect his work as its value increased among collectors. He also said that his studio identified its data sources at MoMA and obtained the necessary permissions for its projects.
upov+1
I have no objection to an artist protecting his intellectual property. He has every right to do so.
But the same principle must apply in both directions.
Anadol cannot reasonably expect other people to respect the rights behind his creations while asking the public to accept his own dataset permissions without meaningful verification.
That is not an allegation of hypocrisy. It is a demand for consistent standards.
The verdict: An ethical claim still needs evidence
The investigation does not establish that Refik Anadol stole artists' work. It also does not independently establish that every image in his enormous training datasets was properly authorised.
There is documented evidence of legitimate open-data use, institutional collaboration and increasingly sophisticated first-hand ecological research. There are also significant gaps in the public documentation of the 153-million-image California dataset and the 200-million-image Earth collection.
Dataland now has the opportunity to do something more useful than proclaim ethical intentions. It can publish the sources, ownership, licences, permissions, restrictions and compensation arrangements behind its datasets.
That would give artists something concrete to inspect, researchers something to verify and the public something more reliable than a promise.
The central failure of the 60 Minutes report was not that it celebrated AI art. Nor was it that it gave critics a platform.
It was that it treated an artist's assurance of ethical sourcing as an adequate answer to a question that demanded documentary evidence.
An AI artwork may contain billions of data points. Its ethical foundation still comes down to a much simpler question:
Who had the right to contribute the material?
References
1. AI art divide: Revolutionary or a gimmick? — CBS News, February 22, 2026. The original investigation, including Anadol's statements, Crabapple's allegations and the MoMA discussion.
2. Artificial Realities: California Landscapes — Refik Anadol Studio, 2024. Confirms the 153-million-image dataset and its stated source.
3. MRI of the Earth — Google Arts & Culture, April 2022. Describes the 200-million-image collection and its combination of satellite data and open-source imagery.
4. Science Data Licenses — NASA Science Data Portal. Establishes the distinction between CC0 NASA mission data and other material requiring source-level rights verification.
5. Terms of Use — Museum of Modern Art. Establishes that MoMA's public collection metadata is CC0 and that the dataset excludes artwork images.
6. Training AI to Understand Nature — Dataland, October 7, 2026. Provides the latest published account of the Large Nature Model's ecological datasets, fieldwork and research contributors.
7. Smithsonian Data Access and Image Use Policy — Smithsonian Institution. Demonstrate that collection assets have different licensing and commercial-use conditions.
8. Requesting Photos & Media — National Geographic Special Collections, updated October 29, 2026. Explains why access to archival material does not automatically establish commercial reuse rights.
9. Copyright and Artificial Intelligence, Part 3: Generative AI Training — U.S. Copyright Office, May 9, 2025. Provides the legal analysis of publicly available material, fair use, licensing and potential market harm; this remains a pre-publication report, not a judicial ruling.
10. Bartz v. Anthropic — U.S. District Court, Northern District of California, June 23, 2025. Distinguishes lawful book acquisition for training from the retention of pirated copies.
11. Kadrey v. Meta — U.S. District Court, Northern District of California, June 25, 2025. Explains why a particular fair-use judgment cannot be treated as blanket legal approval of AI training.
12. Painting with data: How media artist Refik Anadol creates art using generative AI — World Intellectual Property Organization, December 1, 2024. Records Anadol's own statements about dataset provenance, permission, institutional partnerships and protecting his intellectual property.