When a critic sits down to write about a piece of AI-generated art, there is a temptation to spend the entire review explaining how the image came to exist. The model name, the prompt structure, the number of diffusion steps: all of this can crowd out the single question that criticism exists to answer, which is whether the work succeeds on its own terms. A reader who wanted a technical manual would have sought one out. A reader of criticism wants to know whether the piece rewards attention, and why.

This article sets out a working method for writing about AI art that treats the tool as relevant context rather than the entire subject. The aim is not to ignore the mechanics of generation, since they often matter, but to keep them in proportion to the formal, historical and emotional questions that criticism has always asked of pictures, objects and performances. What follows draws on established critical practice and applies it, with some adjustments, to a medium that is still finding its vocabulary.

The Risk of Letting the Tool Overshadow the Work

Early writing about photography suffered from a similar imbalance. Nineteenth-century commentators spent long paragraphs marveling at the chemistry of the daguerreotype before they learned to discuss composition, light and subject with the same seriousness they applied to painting. AI art criticism risks repeating that pattern. An essay that opens with three paragraphs on transformer architecture and closes with a single sentence about color and framing has not reviewed the artwork. It has reviewed the software.

The tool-centered approach also flattens differences between artists who use the same underlying model. Two people working with the same diffusion system can produce work that differs as much as two photographers working with the same camera body. If a critic's vocabulary stops at "this was made with a text-to-image model," the review cannot distinguish a careful, iterative practice from a single lucky prompt. That distinction matters to readers who want to know which artists are doing serious work.

A useful test is to ask whether a paragraph about the technology would need to change if the artist switched tools next year. If the answer is no, the paragraph is probably doing descriptive housekeeping rather than critical work, and it can often be shortened. Housekeeping has its place near the start of a piece, but it should not crowd out analysis.

Balancing Technical Explanation With Formal Analysis

Readers do benefit from some technical grounding, particularly when a specific method shapes what appears on screen. Knowing that an artist used inpainting to rebuild a damaged region of a scanned photograph, for instance, helps explain why certain textures repeat in a way that a camera lens would never produce. That single fact belongs in the review because it affects how the viewer should read the surface.

The discipline is to introduce technical detail only when it answers a question the viewer would otherwise ask. A critic might write: "The recurring lattice pattern in the background is not decorative; it is a known artifact of the upscaling process the artist used, and here it has been left visible rather than smoothed away." That sentence does two things at once. It explains a mechanism and it makes a claim about artistic intention, which is the proper subject of criticism.

A practical way to hold the balance is to allocate space deliberately. In a piece of roughly 1,200 words, two or three short passages, each no longer than three sentences, are usually enough to cover the technical ground a general reader needs. The remaining space should go to composition, color, scale, sequencing, and the relationship between the work and other art, whether generated or not.

• Explain a method only when it changes how the viewer should interpret a visible feature.
• Name the tool once, early, and avoid repeating the brand name as a substitute for description.
• Translate technical terms into visual consequences: not "latent space interpolation" alone, but what that interpolation produces on the surface of the image.
• Reserve longer technical digressions for sidebars or footnotes rather than the main critical argument.

Asking the Same Questions You Would Ask of Any Artwork

Criticism has a long-standing set of questions that apply regardless of medium: what is the subject, how is the composition organized, what does the color or tonal range contribute, how does scale affect the viewer's bodily relationship to the piece, and what does the work ask of its audience. These questions work for a Flemish still life, a video installation, and a generated image, because they address the experience of looking rather than the method of making.

Consider a series of AI-generated portraits that recombine historical costume with contemporary facial features. A critic can ask the same questions asked of any portrait series: does the sequencing build a narrative, does the lighting treat each subject with consistency or variation, and does the work comment on the history of portraiture or merely borrow its surface. None of those questions require mentioning the generation process at all, and the review is stronger for addressing them directly.

It helps to draft a short list of questions before writing and to answer each one from direct observation of the work, not from the artist's statement. An artist's statement can be quoted, but it should not substitute for the critic's own account of what is visible. If a statement claims the piece "interrogates memory" and the image itself shows no formal device that supports that claim, the critic is obligated to say so.

Comparing the generated work to a non-generated precedent is often the fastest way to anchor a review in shared critical ground. A piece that reworks family photographs through a generative model might be usefully set beside Gerhard Richter's blurred photo-paintings, not because the techniques are equivalent, but because both bodies of work raise questions about the reliability of the photographic image.

Useful Vocabulary That Avoids Jargon

Writers covering this field often reach for terms that are precise within a technical paper but opaque to a general reader. "Hallucination," "latent space," and "checkpoint" each have a specific meaning to practitioners, and each can be replaced or glossed with language that serves the critical argument rather than the technical one.

Beyond individual terms, the larger habit worth building is writing about outcomes rather than mechanisms whenever an outcome-based sentence will do. Instead of "the artist used a low guidance scale," a critic might write "the forms remain loose and under-defined, closer to a sketch than a finished rendering." The second sentence tells the reader what to look for, which is the proper job of criticism, and it remains accurate even for a reader who has never adjusted a guidance parameter.

Examples of Criticism That Gets the Balance Right

Some published criticism has managed this balance well, and it is worth studying the method rather than only the conclusions. A review of Refik Anadol's data-driven installations typically spends a sentence or two on the data source, such as museum archives or weather records, and then moves quickly to the experience of scale, motion and color inside the installation space. The technical fact grounds the viewer's understanding of where the imagery originates, but the bulk of the writing addresses what it feels like to stand in the room.

Writing on Sougwen Chung's collaborative drawing performances offers another model. Critics have described the robotic arm's technical capacity in a single paragraph, then devoted the remainder of the piece to the choreography between the artist's hand and the machine's mark-making, treating the collaboration as a question of gesture and rhythm rather than engineering. The technical note exists to make the gesture legible, not to replace discussion of it.

A shorter example: a review of an AI-assisted print series might open with one sentence noting the artist trained a model on a personal archive of family snapshots, then spend the remaining paragraphs on how the resulting images handle facial ambiguity, repetition across the series, and the uneasy resemblance between generated faces and remembered relatives. The technical origin explains why ambiguity appears; the critical work is describing what that ambiguity does to the viewer.

Common mistakes

The most frequent error is leading with the tool and never returning to the work, so that the review reads as a product description. A second common mistake is quoting technical terms without translation, which excludes readers who have not studied machine learning and adds nothing for readers who have. A third is accepting an artist's statement about intention as equivalent to an observed formal achievement, rather than checking the claim against the image itself. A fourth, less obvious mistake is over-correcting into silence: refusing to mention the generative process at all, even when a single sentence of explanation would resolve a visible oddity the reader is certain to notice.

Practical next steps

A critic preparing to write about a piece of AI art might begin by viewing the work twice without reading any artist statement, noting formal observations on paper both times. Only after that should the method of generation be researched, and only the facts that explain a visible feature should make it into the draft. Drafting a short list of the same questions asked of any artwork, covering subject, composition, color, scale and sequencing, before writing a word of technical explanation keeps the priorities straight on the page.

Editors reviewing such pieces can apply a simple check: count the sentences devoted to mechanism against the sentences devoted to observation and interpretation. A ratio weighted heavily toward mechanism suggests the draft needs revision. Writers who want to build facility with this balance might practice on a single still image, writing one short paragraph that would satisfy a reader who knows nothing about the generative process and a second that would satisfy a reader who wants the technical grounding, then deciding how much of the second paragraph the final piece truly needs.