AI Artwork Is Not Proof of Theft

The claim: AI image equals stolen image

An image was made with a generative tool. Someone labels it “stolen art,” and apparently the investigation is complete. Are we serious? The tool used is one fact about the workflow. It does not identify a source image, show copied protected expression, establish how the model’s training material was acquired, or prove what the person publishing the image knew.

That does not make every AI image ethical or lawful. It means the accusation has to name the act. Training a model, generating an output, editing the output, publishing it, and marketing it are separate events. Collapsing them into one dramatic word may be efficient, but it is not evidence.

Reality check: copyright asks what was actually used

U.S. copyright law gives copyright owners specific exclusive rights, including reproduction, adaptation, distribution, and public display. Infringement occurs when someone violates one of those rights without authorization or another lawful basis. The analysis therefore needs a protected work, a relevant act, and evidence connecting the two.

An output that reproduces or is substantially similar to protected expression can create a genuine infringement question. An output that copies a protected character, composition, or other original expression does not receive immunity because a model produced it. But “a model produced it” does not itself show that the output copied any particular protected work. The label identifies a technology. It does not perform the comparison.

Training is a separate—and unresolved—question

Model training can involve copying works. Whether a particular use is licensed, covered by an exception such as fair use, or infringing depends on facts including the source material, purpose, model behavior, outputs, and market effect. The U.S. Copyright Office’s generative-AI training report rejects a one-rule-fits-all answer and discusses uses along a spectrum, from more transformative research or constrained systems to models designed to produce substantially similar expressive outputs.

The source of the training material also matters. A developer that knowingly builds a commercial library from pirated copies presents a different record from a system trained on licensed, openly licensed, public-domain, or creator-supplied material. Adobe, for example, states that its Firefly models are trained on licensed and public-domain content rather than Creative Cloud customers’ personal content. That is the company’s account of its dataset—not a universal certificate for every tool—but it is enough to disprove the slogan that all generative image systems necessarily use the same sources.

The user did not choose the developer’s dataset

The developer chooses or approves the training sources, builds the dataset, trains the model, and controls its safeguards. A person using the finished tool chooses prompts, reference inputs, edits, selection, publication, and marketing. Those responsibilities can interact, but they are not interchangeable.

If a user uploads an artist’s protected work without permission, asks for a close copy, publishes a reproducing output, falsely attributes it, or sells it in violation of a license, evaluate that conduct. If the complaint concerns unlicensed training, investigate the developer and the dataset. Assigning the developer’s acquisition act to every downstream user is the visual-art version of blaming the passenger for how the car company sourced its steel.

AI assistance can sit inside human creative work

The U.S. Copyright Office says AI used as an assistive tool does not bar copyright protection for human-authored expression. Human selection, arrangement, and sufficiently creative modifications may be protected even when a work also contains generated material; purely machine-generated material is not protected merely because someone supplied a prompt.

Copyrightability and infringement are different questions. “Can the user claim copyright in this contribution?” is not the same as “Did this output unlawfully copy somebody else’s work?” A work may contain unprotectable generated material without infringing anyone. A human-created work can infringe. Yes, we really do have to ask both questions.

What would count as actual receipts

• The allegedly copied source work and a side-by-side comparison of protected expression.
• Reliable records showing that a developer acquired or used particular material without authorization.
• The license, terms, or platform rule that governed the input, model, output, or publication.
• Prompt, reference-image, edit-history, or provenance records that connect a person to the relevant act.
• Evidence of false attribution, deceptive marketing, identity misuse, or another specifically identified harm.

An artist may reasonably object to unlicensed training, economic displacement, imitation, disclosure practices, or the treatment of creative labor. Those arguments deserve to be stated accurately. “I believe this system was built or used unethically” is an ethical claim. “This person stole this artwork” is a factual accusation that needs a work, an act, an actor, and evidence.

What creators are actually accountable for

A person publishing AI-assisted or generated artwork is responsible for the inputs they supply, the outputs they choose, the edits they make, the rights they claim, the disclosures a platform requires, and the way the work is marketed. Developers are responsible for their training sources, licenses, model design, safeguards, and representations. Platforms are responsible for their own rules and enforcement.

Criticize the conduct you can document. Demand better licensing, provenance, disclosure, or compensation where the evidence supports it. But “AI artwork exists” is not proof that its user stole a specific image any more than a paintbrush is proof that the painter forged a Monet. Name the work. Show the copying. Then we can discuss the receipts.

This explainer provides general information, not legal advice about any particular image, model, dataset, license, or dispute.