Evaluating an AI-generated artwork requires a different set of questions than judging a traditional painting or photograph, but it is not a separate discipline invented from scratch. Many of the same criteria apply: composition, color handling, conceptual clarity. What changes is the need to account for process, since the tools used to produce the image can obscure or inflate the amount of human decision-making involved. A checklist helps because it forces the same questions to be asked every time, regardless of how polished the final image looks.

This article lays out a structured way to look at AI-assisted or AI-generated work, whether you are a curator selecting pieces for a show, a collector considering a purchase, or a juror reviewing competition entries. The goal is not to settle the debate over whether AI art "counts" as art. The goal is to give you a repeatable method for separating surface polish from substantive creative work, and for spotting the difference between a project built through iteration and one generated in a single prompt.

Separating Technical Quality From Artistic Intent

Technical quality and artistic intent are two different axes, and conflating them is the most common error in early evaluations of AI-generated work. Technical quality covers things a diffusion model or rendering engine can produce reliably on its own: coherent anatomy, consistent lighting, clean edges, resolution suitable for print or screen. Artistic intent covers the choices that shape why the image looks the way it does: subject selection, framing, symbolic content, the relationship between parts of the composition.

A useful test is to ask whether the technical polish is serving a specific idea or simply demonstrating that the tool works. An image of a hyperrealistic eye with visible iris texture and correct catchlights is technically accomplished, but if it exists only to show off the model's capacity for detail, it says little about artistic intent. Compare that to a piece where the same technical skill is used to direct attention toward a single asymmetry in an otherwise symmetrical face, supporting a stated theme about imperfection.

When scoring technical quality, check for these specific markers: consistent perspective across all elements in the frame, correct number of fingers and joints in figures, coherent shadow direction relative to a single light source, and absence of repeated texture patterns that reveal the model's default tiling behavior. None of these markers tell you anything about the strength of the idea behind the work, which is why they should be scored separately from conceptual criteria.

Checking Provenance and Process Documentation

Provenance documentation for AI-generated art typically includes the model or tool used, the date of generation, the prompt or prompt sequence, and any post-processing steps such as inpainting, upscaling, or manual retouching. The absence of this documentation is not automatically disqualifying, but it does limit how much you can verify about the claims an artist makes regarding their process.

Ask for the following when it is available:

• The base model or tool name and version number, since different versions of the same model produce noticeably different output quality and style.
• A record of the prompt or prompts used, including negative prompts if the tool supports them.
• Seed values or generation parameters, which allow a third party to reproduce a similar result and assess how much variation the artist introduced across attempts.
• A list of manual edits made outside the generation tool, such as compositing in Photoshop or color grading in a separate application.

If an artist cannot produce any of this, treat the work as undocumented rather than fraudulent. Some artists lose track of early prompt logs, especially from informal experimentation sessions. The correct response is to note the gap in your evaluation record and weigh it against other evidence of process, such as intermediate files or described revision steps, rather than to assume bad faith.

Assessing Iteration History and Revision Depth

Revision depth refers to how many distinct generation or editing passes a piece went through before reaching its final form. A single prompt that returns a usable image on the first attempt represents minimal revision depth. A piece built from 40 or more generation attempts, several rounds of inpainting on specific regions, and manual color correction represents substantial revision depth.

Request a visual record of the iteration history if the artist can provide one. This might take the form of a contact sheet showing early, middle, and final versions, or a short video capturing the editing session. Look for evidence of directed change: did later versions address a specific weakness in earlier ones, such as a distorted hand or a flat background, or do they appear to be random regenerations with no clear through-line?

None of these patterns is inherently superior for every context. A competition that explicitly welcomes raw single-prompt generations should not penalize an entrant for staying within that category. The point of checking revision depth is to understand where a piece falls, not to assume more revisions always means more merit.

Looking for Conceptual Coherence

Conceptual coherence means that the visual elements of the piece work together to support a single idea or a clearly related set of ideas, rather than reading as a collection of visually striking but disconnected details. This is where AI-generated work often falls short, because generation tools are skilled at producing individually compelling elements, such as dramatic lighting or intricate texture, without any underlying logic connecting them.

To test for coherence, try to state the central idea of the piece in one sentence based only on what you see, without reading the artist's statement. Then check whether each major compositional element supports that sentence. If the central figure, the background, the color palette, and the framing all point toward the same idea, coherence is likely present. If removing or replacing any single element would not change your reading of the piece, that element may be decorative rather than integral.

A practical example: a piece depicting a decaying library with floating books and a central figure reading by candlelight suggests a coherent theme around memory or the persistence of knowledge. If the same piece also includes an unrelated dragon in the background purely because the prompt included it as a stylistic flourish, that addition weakens coherence even though it may be technically well rendered.

Comparing Against the Artist's Stated Intent

Every evaluation should include a comparison between the artist's stated intent, usually found in a title, caption, or artist statement, and what the image actually delivers. This comparison is not about punishing ambition. It is about checking whether the claimed concept is legible in the work itself or whether the statement is doing work the image cannot do on its own.

Read the statement first, then look at the piece for no more than 30 seconds and write down three things you notice. Compare your three observations against the claims in the statement. If there is little overlap, the piece may rely too heavily on text to communicate its meaning, which is a legitimate critique regardless of whether the work was made with AI tools or traditional media.

Pay attention to statements that describe process rather than content, such as "this piece explores the tension between control and randomness in generative systems." Statements like this should be checked against the actual revision history. If the artist describes a deliberate exploration of randomness but the piece came from a single unedited generation, the stated intent and the demonstrated process do not match, and that gap belongs in your evaluation notes.

Red Flags That Indicate Minimal Human Input

Certain patterns reliably indicate that a piece involved little beyond writing a prompt and selecting an output. None of these patterns proves the work has no value, but they should lower your confidence in claims of extensive artistic labor.

• Inability to describe why specific compositional choices were made, beyond "it looked good."
• No variation across a claimed series, where multiple pieces share nearly identical composition and only differ in subject skin, suggesting a single prompt template run repeatedly.
• Presence of common model artifacts left uncorrected, such as text that resembles writing but is not legible, or hands with an incorrect number of digits in a final published piece.
• Artist statements that use generic language applicable to almost any image, such as references to "duality" or "the human condition" without any specific visual evidence pointing to those themes.
• Refusal or inability to provide any process documentation when asked directly, combined with claims of extensive manual work.

If you encounter two or more of these red flags in a single piece, document them specifically rather than issuing a general judgment of low quality. Specific notes, such as "six images in the series share identical pose and lighting, suggesting a repeated template," are more useful for later reference than a vague note like "feels generic."

Common mistakes

Evaluators new to AI-generated work tend to make a small number of recurring errors. The first is treating technical flawlessness as proof of artistic merit, when in fact current generation tools can produce flawless technical output with zero conceptual input. The second is the opposite error: dismissing a piece entirely because it used AI tools, without applying the same evaluation criteria used for traditional media.

A third common mistake is accepting an artist statement as a substitute for visual evidence, rather than checking the statement against what the image itself communicates. A fourth is failing to distinguish between an artist who used AI as one tool among several in a longer process, and an artist who generated a single image and stopped. These represent very different levels of involvement and should be scored differently.

Putting the Checklist to Use

When you apply this checklist in practice, work through it in a fixed order rather than jumping to the section that interests you most. Starting with technical quality and provenance before moving to conceptual coherence prevents a strong concept from masking weak execution, and prevents strong execution from masking a weak or absent concept.

Use the following one-page version for quick reference during a review session:

1. Score technical quality separately from conceptual content, using specific markers like consistent perspective and correct anatomy.
2. Request the model name, version, prompt record, and list of manual edits. Note any gaps rather than assuming either full transparency or deception.
3. Identify the revision pattern: single generation, selected generation, generation plus targeted editing, or generation used as raw material.
4. State the central idea of the piece in one sentence based on visual evidence alone, then check whether each major element supports that sentence.
5. Compare the artist's stated intent against your own independent observations, noting any mismatch.
6. Check for the five red flag patterns listed above and document any that apply with specific examples, not general impressions.
7. Record your findings in a format you can refer back to later, since memory of individual pieces fades quickly when reviewing many works in sequence.

If you are evaluating work for a formal context such as a juried exhibition, a licensing agreement, or a sale involving significant sums, consult with a colleague experienced in digital provenance or, where contractual or copyright questions arise, a qualified legal professional. This checklist supports consistent critical judgment, but it does not resolve questions of ownership, licensing rights, or authenticity claims that carry legal weight.