The phrase "AI art" gets applied to work made with a single text prompt and to paintings where an artist spent forty hours on canvas and used an AI tool only to test a composition idea beforehand. This is not a minor sloppiness. It affects how work gets credited, whether it qualifies for certain exhibitions or competitions, how platforms moderate it, and in some jurisdictions, whether it can be copyrighted at all. This article defines the three terms that matter most, states where the boundaries between them actually fall, and walks through how to apply the definitions to a real piece of work.
None of these definitions are universally standardized yet. Different platforms, competitions, and legal bodies draw the lines slightly differently. Where that is true, this article says so explicitly rather than presenting one convention as settled fact.
Why Terminology Became a Problem
Before 2022, the working vocabulary for digital art was simple: digital painting, 3D render, photo manipulation, vector illustration. Each term described a method with a traceable chain of human decisions. The arrival of diffusion-based image generators (Midjourney, Stable Diffusion, DALL-E) broke that chain in a way existing vocabulary could not describe, because a single text prompt could now produce a finished image with no intermediate human-authored steps.
Competitions and stock agencies reacted first, and inconsistently. The Colorado State Fair's 2022 digital arts category accepted an AI-generated piece, "Théâtre D'opéra Spatial," which won first place and triggered public debate specifically because the category had no AI-specific rules at the time. Getty Images banned AI-generated submissions outright in September 2022, citing unresolved copyright liability. Adobe Stock, by contrast, began accepting AI-generated content in December 2022 but required contributors to label it as such.
This patchwork of policy responses is the direct result of undefined terms. A rule that says "no AI art" is unenforceable without a definition of what counts, because "I used AI to remove a background" and "I typed a prompt and generated the final image" are categorically different acts that the same three words currently describe.
AI-Generated: Definition and Boundaries
AI-generated refers to a work where the final visual output, or substantially all of it, was produced by a generative model from a prompt or seed input, with no manual drawing, painting, or pixel-level editing by a human after generation. The human's contribution is confined to the prompt text, parameter settings (aspect ratio, seed number, guidance scale), and the selection of which generated output to keep.
This category includes work from text-to-image models like Midjourney, Stable Diffusion, and DALL-E 3, as well as image-to-image generation where a rough sketch is used only as a structural guide and the model fills in all final detail, color, and rendering. It also includes AI upscaling or AI inpainting when that process generates new content rather than simply enlarging existing pixels.
The boundary gets tested in a few recurring situations:
- Running the same prompt twenty times and picking the best result is still AI-generated; curation is not authorship of the pixels.
- Adjusting the prompt iteratively across many generations (prompt engineering) is still AI-generated, even if it took three hours and forty attempts, because the model still authored every pixel in the final file.
- Applying a basic color-correction filter to a generated image after export is a gray zone that most current platform policies still classify as AI-generated, since the correction did not touch the model's rendering decisions.
The U.S. Copyright Office has taken a formal position relevant here: in its March 2023 policy statement and in the 2023 Zarya of the Dawn decision, it stated that purely AI-generated images, where prompts are the only human input, do not qualify for copyright protection because copyright requires human authorship of the expressive content. This is a legal determination specific to U.S. law, not a global standard, and artists working across jurisdictions should consult a qualified intellectual property attorney rather than rely on general articles for their specific case.
AI-Assisted: Definition and Boundaries
AI-assisted describes a work where a human performs substantive creative labor, drawing, painting, compositing, or sculpting, and uses an AI tool for a bounded, specific function within that process rather than for generating the final image wholesale. The defining test is whether removing the AI tool would still leave a recognizable, nearly-complete work behind.
Common examples include:
- Using an AI denoiser or upscaler on a digitally painted piece that the artist rendered by hand.
- Generating AI reference images for lighting or pose study, then painting the final piece traditionally or digitally without tracing the AI output.
- Using AI-assisted selection tools (such as Photoshop's Generative Fill used only for object removal, not content addition) within an otherwise manually composited photograph.
- Using an AI color-palette generator to pick a scheme, then hand-painting the entire piece using that palette.
The line between AI-assisted and AI-generated gets genuinely contested when generative fill or inpainting is used to add new content to part of an otherwise hand-made image. If an artist hand-paints 90 percent of a canvas and uses Generative Fill to create a background element covering the remaining 10 percent, platforms differ on classification. Adobe's own contributor guidelines as of 2024 require disclosure of any generative AI content regardless of the percentage involved, while some art competitions set a threshold (commonly discussed informally as "any AI-generated pixels in the final visible image disqualifies AI-assisted status") without publishing an exact percentage rule.
Human-Made With AI Reference: A Third Category
This category covers work where AI tools were consulted during the process, generated images, mood boards, color studies, pose references, but no AI output appears in or directly informs the pixel content of the final piece. The AI tool functioned the same way a photo reference, a Pinterest board, or a photograph of a model would function: as research material, not as production material.
An artist who generates fifteen AI images of a dragon to study wing anatomy, then closes those images and paints an original dragon from imagination and traditional anatomy references, falls into this category. The resulting work is functionally no different from an artist who studied photographs of real bats and lizards for the same purpose. The AI tool here is a research aid, comparable in function to a library.
Some competitions and galleries do not require disclosure for this category at all, treating it as equivalent to any other reference use. Others, concerned about edge cases where "reference" bleeds into "near-copy," request disclosure regardless. Artists submitting to a specific venue should check that venue's written policy rather than assume this category is automatically exempt, because policies vary and some are stricter than the general convention described here.
Quick Comparison
| Category | Who renders final pixels | Typical disclosure requirement |
|---|---|---|
| AI-generated | The model | Almost always required |
| AI-assisted | Human, with AI tool for a bounded sub-task | Often required, policy-dependent |
| Human-made with AI reference | Human, entirely | Rarely required, but check venue rules |
Why the Label Does Not Equal a Value Judgment
A classification of "AI-generated" states a fact about process. It does not state that the work is low-effort, low-skill, or low-value, and it does not state that an AI-assisted or human-made piece is automatically superior. Prompt engineering for a complex, multi-element generative composition can take hours of iterative refinement and genuine aesthetic judgment; conversely, a human-made painting can be derivative or poorly executed. Process category and quality are independent variables.
Where the label does carry practical weight is in contexts with explicit rules tied to it: copyright eligibility in the U.S. as discussed above, competition entry categories, stock agency acceptance criteria, and client contracts that specify "original human-authored artwork" as a deliverable. In those specific contexts, misclassifying a work is not a stylistic error but a factual misrepresentation that can lead to disqualification, contract breach, or in professional settings, loss of client trust.
Buyers and collectors who want a value judgment separate from the label should ask about technique, time investment, and the artist's documented process rather than inferring quality from the category name alone.
How to Apply These Definitions to a Specific Work
Use this sequence when you need to classify a specific piece:
- Identify whether a generative model produced any part of the final visible pixels. If yes, move to step 2. If no, the work is either human-made or human-made with AI reference, depending on step 4.
- Estimate what proportion of the final image's pixels came directly from model output versus hand-executed work. If the generative portion is the entire image or the dominant visible element, classify as AI-generated.
- If the generative portion is a minor, bounded element within a majority hand-made image (a generated sky behind a hand-painted figure, for instance), classify as AI-assisted, and disclose the specific tool and function used.
- If no generative output appears in the final file at all, but AI tools were used earlier for reference, mood boards, or planning, classify as human-made with AI reference and check whether your specific platform or venue requires disclosure of this use.
- Document your process notes (prompts used, tools used, stages involved) as you work, not after the fact. Retroactive reconstruction of a multi-hour process is unreliable and can be contested if the classification is ever challenged.
Exceptions and Disputed Cases
Several situations resist clean classification under any current convention. AI-powered "smart" selection and masking tools built into mainstream software (Photoshop's Select Subject, for example) are generally not treated as AI-assisted in the sense discussed here, because they automate a mechanical task rather than generating new creative content; most platforms exempt these tools from disclosure requirements, though this exemption is a convention, not a universal rule.
Style transfer, where an AI model repaints an existing human-made image in the visual style of another artist or movement, is disputed because the underlying composition is human-authored but the final rendering is model-generated. Most competition rules as of 2024 classify heavy style-transfer output as AI-generated, since the final pixels were produced by the model, but this is actively debated and not settled across all venues.
Collaborative pipelines where one artist generates AI content and a second artist substantially hand-paints over or extensively reworks it (a practice sometimes called "AI-to-hand" or "paintover") occupy a genuine gray zone. The honest approach is to disclose both stages explicitly ("AI-generated base, hand-painted finish, approximately 70 percent of final surface repainted") rather than force the work into a single label that obscures the actual process.
Common Mistakes
Treating "I used AI somewhere in my process" as equivalent to "this is AI art" collapses three distinct categories into one and misinforms viewers, buyers, and competition judges about what they are actually looking at.
Assuming copyright protection exists for a purely AI-generated image without checking current law in the relevant jurisdiction; in the U.S. this assumption is currently false for prompt-only generation.
Failing to document process at the time of creation, then attempting to classify or defend the work's category months later from memory, which is unreliable if the classification is ever disputed by a platform, client, or competition.
Next Steps
If you are an artist, keep a running process log for each piece: tools used, prompts if any, approximate proportion of hand-executed versus generated content, and the sequence of steps. If you are submitting to a specific competition, platform, or client, read that entity's written AI policy directly rather than relying on general industry convention, since definitions and disclosure thresholds vary. If copyright status matters to your specific situation, especially for commercial use, consult an intellectual property attorney familiar with current case law in your jurisdiction rather than treating any article, including this one, as legal advice.
