We spend most of our days with our noses close to paper, canvas, and screens, so we notice things before we can name them. A smudge that doesn't match the pressure of the stroke around it. A highlight that sits a little too perfectly in the center of a curve. Over the past year, as AI-assisted and AI-generated images have moved into galleries, print shops, and client portfolios, we started keeping notes on what actually gives these pieces away, and what doesn't. This checklist is the result of that note-taking, built from side-by-side comparisons of hand-finished work and machine-assisted work under raking light and at full zoom.

We want to be upfront about what this is and isn't. It is a set of observations we trust enough to use in our own studio reviews. It is not a forensic tool, and it will not give you certainty. Treat it the way you'd treat a loupe: useful for looking closer, not for passing judgment on its own.

Why we built this checklist, and its limits

We started this project after a print buyer asked us, point blank, whether a batch of botanical illustrations had been touched by AI tools somewhere in the process. We didn't have a good answer, because we didn't have a shared vocabulary for what to look for. So we pulled twenty pieces, ten we knew were hand-drawn and ten we knew involved AI generation or heavy AI retouching, and we compared them under the same lighting and magnification.

What we found is that no single cue is reliable on its own. A repeated brushstroke pattern can come from a human working quickly on a deadline, just as easily as from a diffusion model's texture habits. The checklist works best as a cumulative read: three or four small signals pointing the same direction matter more than one dramatic one.

We also want to name the limit plainly: current AI tools improve every few months, and any visual tell we list here has a shelf life. We revisit this list every quarter. If you're reading this more than a year after publication, assume some items are outdated and look for updated guidance before relying on them.

Checking edges and transitions

Edges are where a lot of the tells live, because edges are where a material has to make a decision, wet meets dry, pigment meets paper tooth, one color layer stops and another starts. Hand work tends to show the evidence of that decision: a slight bleed, a scratchy boundary where a brush ran low on paint, a visible correction.

AI-generated or AI-assisted edges often resolve too smoothly across regions that would normally have different physical behavior. We look closely at:

• Where hair or fur meets skin: real rendering usually shows individual strokes of varying length; AI output sometimes blurs this into a soft gradient with no individual marks.
• Where fabric folds: shadows that should deepen with each fold sometimes flatten or repeat at regular intervals instead of responding to the specific drape.
• Where a hand meets an object: fingers wrapping around a cup or brush handle are a classic stress point; we count fingers, check that each one has a consistent light source, and look for the small overlap errors AI models still produce here.

One studio trick: zoom to 300 to 400 percent on a transition point and ask whether you can tell which material was laid down first. If the layering order is ambiguous or physically impossible, that's worth flagging.

Looking for repetition and pattern drift

Humans get tired, and that fatigue actually helps authenticate work: a hand-inked crosshatch pattern will drift in spacing and angle across a large area because the wrist and the eye aren't machines. AI-generated texture, by contrast, can repeat a motif with a regularity that feels off once you notice it, even when the overall image looks varied at a glance.

We look at background foliage, brick patterns, woven textures, and crowd scenes first, because these are the areas most likely to contain tiling artifacts. Cup the image in your hands, or crop it on screen, and compare two sections that should be independently drawn, two trees, two sections of a stone wall. If you find near-identical leaf clusters rotated and rescaled, or bricks that repeat in an exact rhythm, that's a pattern-drift flag.

Text inside an image, a street sign, a book spine, a logo on a shirt, remains one of the more stubborn tells. We still find garbled or invented lettering in AI output often enough that we check it early, though we expect this gap to close over time.

Reading surface and material cues

This is where our own trade knowledge matters most, because we handle the physical outputs, not just the screen image. If a piece claims to be a traditional watercolor, we check whether the paper shows the right kind of cockling, the slight buckle that happens when water hits cotton or cellulose fiber. A flat sheet with a painted-on look of texture, printed from a file that started as AI output, usually feels different under raised light, almost too even.

We also check pigment behavior against the stated medium. Real gouache sits opaque and slightly chalky; real ink pools at stroke ends where the nib or brush released extra fluid. When a "gouache" piece shows a photographic softness with no pooling, no chalkiness, and no visible brush loading, we start asking questions, not accusing anyone, just asking.

Scanning resolution matters here too. We examine files at their native resolution before any upscaling, because upscaling tools, AI or otherwise, introduce their own smoothing that can mimic or mask the cues above. If a file has clearly been run through an upscaler, we note that separately and try to find an earlier, lower-resolution version to compare.

Comparing stated process with visible evidence

Every piece that comes through our review has some stated process, a caption, an artist statement, a materials list. We treat that statement as a hypothesis to test against the object, not as a fact to confirm.

If an artist says a piece took eleven hours of layered oil glazing, we expect to see the evidence of layering: slight ridges where glaze pooled, variation in gloss across the surface, maybe a visible underpainting peeking through at an edge. A flawless, uniform surface claimed to be eleven hours of glazing is not proof of AI involvement by itself, but it's a mismatch worth logging.

We keep a simple two-column note for this comparison: what the artist says happened, and what the surface actually shows. When those columns line up, we move on quickly. When they don't, we slow down and look again before forming any opinion, and we try to talk to the artist directly rather than guessing.

When to simply ask the artist

We've learned that the fastest, most reliable step in this whole checklist is often the one we're tempted to skip: asking. A direct, respectful question, "can you walk me through how you built the texture in this corner," gets better information than an hour of pixel-peeping, because the artist can point to process photos, timelapse footage, or leftover materials that settle the question outright.

We ask this question when two or more checklist signals point the same direction, not on a single flagged detail. We also ask it the same way regardless of what we suspect, because a neutral, curious tone gets a more useful answer than a loaded one.

If an artist can't or won't describe their process in any specific terms, that's information too, though it's not proof of AI use. Some people genuinely don't track their own process closely. We weigh a vague answer alongside the visual evidence rather than treating it as a verdict.

Using the checklist without turning it into an accusation

A checklist like this can slide into a witch hunt if we're not careful, and we've seen that happen in online communities where a single repeated leaf pattern gets treated as a confession. We try to hold two things at once: genuine curiosity about process, and genuine respect for the fact that we might be wrong.

In practice, that means we keep our notes private or shared only with the artist first, not posted publicly as an accusation. It means we use language like "I noticed" and "can you tell me about" instead of "this is obviously." And it means we accept an answer we can't independently verify, because we are not running a forensic lab, we are running a studio and a publication, and our job is to look closely and ask good questions, not to issue rulings.

Common mistakes

• Treating one cue, like a repeated texture, as definitive proof rather than one data point among several.
• Comparing a heavily compressed web image against a high-resolution original and drawing conclusions from compression artifacts instead of the actual image.
• Skipping the conversation with the artist because the visual evidence feels strong enough on its own.
• Applying the checklist only to work you already suspect, which biases what you notice and ignores the same flaws elsewhere.
• Assuming last year's visual tells still apply; AI image tools change quickly, and a checklist item that worked six months ago may no longer hold.

Where to go from here

If you're reviewing a piece and something feels off, start small: pick one section, zoom in, and check it against two or three items from this list before forming any impression of the whole piece. Keep a written note of what you saw rather than relying on memory, because specific detail holds up better than a general feeling when you eventually ask the artist about it.

If the stakes are high, a sale, a commission dispute, a competition entry, we'd treat this checklist as a starting point only and bring in someone with specific expertise in image forensics or, where contracts and money are involved, a professional familiar with intellectual property and disclosure requirements in your region. Our checklist helps us look more carefully. It isn't a substitute for that kind of specialized judgment when the outcome actually matters.