How to Tell If Art Is AI: A Craft-Based Method

Artists are now asked to judge images constantly: as jurors, commissioners, moderators, or as the friend someone sends a screenshot to asking "is this image AI generated?". Knowing how to tell if art is AI relies on the eye this library trains for drawing, because generators fail where a person makes deliberate decisions. This tutorial starts with provenance, then gives a table of ai art tells grouped by craft area with a reliability rating for each, and a table of the mistakes behind false accusations. It belongs to the process and provenance pillar.

Check provenance before you squint

The fastest honest answer comes from the file. Before you look for signs of ai art, run the original image through the Content Credentials verifier. A signed manifest tells you which app or camera produced the file, what edits were recorded, and whether a generative source was declared. A declared generative source is a clear answer; a painting-app record listing only crops and exports is strong support for hand-made work. Reading these records is covered in Content Credentials and C2PA for artists.

Most images in the wild carry no manifest, because social platforms strip it. Even then, do not squint yet. Follow how to check where an image came from: reverse search for an earlier or larger copy, look at the account's history for consistent style and posted process, and check for the same image at higher resolution somewhere older. Ten minutes of provenance work answers more cases than an hour of pixel hunting, and never produces a false accusation.

How to tell if art is AI: the tells by craft area

When provenance gives nothing, you look. Group the tells by the craft skill they test, because each group has a different reliability: construction tells are strongest, since a generator has no internal model of a skull or a hand, and surface tells are weakest, since surface is what generators do best.

Generation tells grouped by craft area, what a hand-made piece does in the same spot, and how much weight each tell carries in 2026.

Craft area	Tell	Hand-made piece instead	Reliability
Construction	Eyes on different wrapping lines; ears at different heights; jaw not joined to the skull	Features sit on one construction cross, even in a loose drawing	High
Construction	Fused or extra fingers, joints bending the wrong way, thumb on the wrong side	Clumsy hands still keep five fingers and a plausible palm	High, but generators fix hands first
Construction	Paired objects that differ: one earring, mismatched buttons, a strap ending in nothing	Both sides drawn from one idea, so they match	High
Edges	Uniform softness everywhere, or hard edges where nothing needs focus	Hardest edge at the focal point, softer with distance, some edges lost on purpose	Medium
Light	Eye highlights on different sides; cast shadows disagreeing with highlights	One key light; the highlight sits on the same side in both eyes	Medium to high
Detail	Background logic drifts: a fence becomes a railing, chairs merge into a table	Background objects stay what they are across the frame	Medium, decays fastest
Text and symbols	Lettering that almost reads; logos close to real but wrong; impossible keyboards or sheet music	Text is legible or clearly abstracted, and consistent	High today, shrinking
Symmetry	Faces mirrored with no natural drift; patterns that lose their rule halfway	Slight asymmetry in faces; patterns keep their rule or break it deliberately	Low to medium

Reading construction

Start with the head. Drop an imaginary cross through the face, a centre line and an eye line wrapping around the ball of the skull. Both eyes should sit on that line, ears between eye line and nose line. In a generated portrait the eyes often sit at different heights or tilt differently while the surface stays flawless; a hand-made drawing can be off by a lot and still be consistent, because both eyes came from one decision. Then hands, then paired objects: count fingers, check joint direction and thumb side, compare left and right earrings, buckles and laces. This is the most reliable family of tells and the first to improve with each new model, so re-check yearly.

Reading edges, light and detail

A painter spends edges on purpose, hard at the focal point and soft or lost elsewhere, as set out in hard and soft edges. Look for a picture where every edge has the same firmness, or where a background element is as crisp as the face. Then find the key light from the strongest cast shadow and check that every highlight agrees with it, especially the two eye highlights. Finally scan the parts of the frame nobody is meant to look at, read any text as an alphabet you know, and count repeats. Symmetry is the weakest group, since many illustrators mirror on purpose.

Why tells decay as generators improve

Every tell in the table is a failure of a specific model family at a specific time. Hands were the famous tell in 2022 and are largely fixed now; text is following. Background logic and paired-object consistency improve more slowly, because they need the model to hold one idea across the whole frame. Ha et al., in the 2024 University of Chicago study Organic or Diffused, tested images from five generators across seven art styles and found that the strongest automated detector and expert artists both did well yet failed differently, and expected those weaknesses to persist as models change rather than disappear.

So the surface tells you learned last year are the ones a current generator has most likely fixed, while construction and provenance keep their value. A judgement resting on one tell rests on a moving target; one resting on provenance, process and a cluster of construction tells holds. This is why NIST AI 100-4 treats detection as one method alongside provenance and watermarking rather than as a standalone answer.

Ask for process, politely

When provenance and tells leave you uncertain, ask. A working artist can usually produce roughs, a layered file or a timelapse within minutes, the record described in how to document your art process. Ask in private, name what you have already checked, and say what would settle it. If the context is a contest or a commission, put process evidence in the rules in advance, so nobody is singled out.

What detection tools can and cannot tell you

Automated detectors are a first pass, never a verdict. They are classifiers trained on labelled image sets, so they report resemblance to training data, and stylized, smooth or heavily compressed hand-made work resembles generated output more than a photo does. In the Chicago study the leading detector was weaker against perturbed images while expert artists produced more false positives. Treat a score as a reason to look closer.

Mistakes that lead to false accusations

The errors behind most false accusations of AI art, why each misleads, and what to do instead.

Mistake	Why it misleads	Do this instead
Judging a compressed repost or screenshot	Recompression softens edges and texture, two of your tell groups	Find the original upload or ask for the full-resolution file
Treating smooth rendering as a tell	Airbrush and soft-blend styles predate generators by decades	Check construction and paired objects; ignore polish
Treating one anatomy error as proof	Artists draw bad hands constantly, and generators now draw good ones	Look for a cluster of construction errors under a flawless surface
Trusting a detector score	Scores measure resemblance to a training set; false positives on stylized art are common	Use the score only as a prompt to check provenance and ask for process
Accusing in public first	The artist cannot undo the accusation, even after showing process	Ask privately, state what would settle it, and allow a real answer

A fair review checklist

1. Check the file for a Content Credentials record: who signed it, what it lists.
2. Reverse search for an earlier, larger copy; read the account's history.
3. Check construction: eye line, hands, every paired object.
4. Check edges and light: one focal point, one key light, matching eye highlights.
5. Scan backgrounds, repeats, text and pattern rules.
6. Ask for process before any conclusion.
7. If still unsure, say so. Uncertain is an honest answer.

How to tell if art is AI, in short

Provenance first, then process, then tells. Read the file and the history before you read the pixels, group the visual tells by craft area so you know which carry weight, and ask for process before you conclude. Used this way, knowing how to tell if art is AI protects hand-made work from wrongful accusation as much as it catches generated images. If you are on the other side of the question, the playbook is in what to do when someone says your art is AI.

Common questions

How can I check if an image is AI-generated?

Check the file before the pixels. Run the original through a Content Credentials verifier, reverse search for an earlier or larger copy, and look at the account's posted process. Then check construction (eye line, hands, paired objects), edges, light, backgrounds and text. If still unsure, ask the artist for a layered file or timelapse.

Can you tell if art is AI-generated just by looking?

Sometimes, and less often every year. Construction errors under a flawless surface, mismatched paired objects and near-legible text are still strong clues, but leading generators fix the famous tells within months. A trained eye should reach a suspicion, never a verdict. Provenance records and process evidence settle the question.

What are the most reliable AI art tells?

Construction tells: eyes on different wrapping lines, hands with impossible joints, and paired objects such as earrings or buckles that do not match. They come from the generator having no internal model of the form. Surface tells such as smoothness, glow or soft edges are the least reliable, because many hand-made styles share them.

Why do AI detectors flag hand-made art?

Detectors are classifiers that measure resemblance to a training set, so smooth gradients, flat cel shading and heavy compression push real paintings toward the generated side. A 2024 study found even expert human judges produced false positives on real art. What to do about a flag is in why AI image detectors flag real art.

Sources and further reading

Primary references for the facts and definitions in this piece, all from independent publishers:

• Ha et al., Organic or Diffused (2024): how automated detectors and artists compare at telling human art from generated images.
• NIST AI 100-4, Reducing Risks Posed by Synthetic Content (2024): detection, provenance and watermarking as complementary approaches.
• Content Credentials: Verify: reading the signed provenance record on a file before judging the pixels.