Why AI Image Detectors Flag Real Art

Plenty of artists have now had the same unpleasant experience: a painting made stroke by stroke is run through an online ai image detector, and the tool announces that it is probably generated. Sometimes a stranger did the checking; sometimes a contest, a shop or a client. This guide explains what the detector is measuring, why stylized hand-made work is the hardest case, what the best available research found, and how an artist, a judge and a moderator should each treat a score. It belongs to the process and provenance pillar.

What an AI image detector is actually measuring

Most automated detectors are classifiers. They are trained on large collections of images labelled real or generated, and they learn statistical regularities that separate the two groups: tiny patterns in noise, texture, colour transitions and edges that people cannot see. When you upload a picture, the detector compares its pixel statistics to those learned patterns and returns a probability. It never watched you draw. It has no access to your sketches, your layers or your timelapse. It only knows how closely the finished pixels resemble the two piles it was trained on.

That training set is the whole story. If the real pile was mostly photographs and the generated pile mostly smooth renderings, then anything smooth and polished looks generated to the model, whatever its origin. Every false positive on hand-made art comes from that overlap.

Why stylized hand-made work gets flagged

The features that make digital illustration look finished are the same features detectors associate with generation. That is an uncomfortable overlap for anyone who paints in a clean style, and it explains why the false positives cluster in particular kinds of work.

Features of hand-made art that trigger false positives, why a classifier reads each one as generated, and what an artist can do about it.

Feature	Why a detector reads it as generated	What to do
Smooth gradients and soft blending	Airbrushed skin and glowing light, built with the stack in layer blend modes for shading, produce even transitions with none of the sensor noise photographs carry	Keep the layered master; the layer panel shows the blend was built, not sampled
Clean cel shading and flat fills	Large areas of perfectly uniform colour leave none of the irregular texture a classifier learned from cameras and physical media	Export a work-in-progress crop with visible construction lines under the flats
Heavy compression from social platforms	Resizing and recompression wipe the fine detail many detectors rely on and push real and generated images toward the same blurred statistics	Never argue from a repost; supply the original export at full resolution
Small crops and screenshots	A thumbnail or a screenshot of a screenshot gives the model almost nothing; uncertain inputs produce confident-looking, unreliable outputs	Ask which file was tested and offer the original
Popular anime and character styles	These styles are heavily represented in generator training data, so their look is close to what the classifier learned as generated	Show a consistent account history and posted process in the same style
Generated texture used honestly	A generated element carries generated statistics into the file	Disclose it in the description before anyone asks

What the 2024 study of eight detectors found

The most useful study for artists so far comes from researchers at the University of Chicago, published in 2024 as Organic or Diffused: Can We Distinguish Human Art from AI-generated Images? by Ha et al. The team collected real human art across seven styles, generated matching images with five generative models, and tested eight kinds of detector: five automated tools and three groups of people, including 180 crowdworkers, more than 4,000 professional artists, and 13 expert artists experienced at spotting generated work.

The best commercial detector and the expert artists both performed very well, but they failed in different ways. The automated tool was weaker when images had been deliberately perturbed, while the experts produced more false positives, calling real art generated. The authors concluded that a combined team of human and automated review gave the best mix of accuracy and robustness, and they expected those weaknesses to remain as models continue to evolve. Neither a tool alone nor an expert alone should be the last word, and the human experts were the ones more likely to wrong a real artist.

Standards bodies take a similar view. The US National Institute of Standards and Technology, in NIST AI 100-4, its 2024 overview of technical approaches to synthetic content, treats detection as one tool among several, alongside provenance records and watermarking, rather than a standalone answer. A signed provenance record on the file, checked with the Content Credentials verifier, says something a classifier cannot: which app produced the file and what was done to it.

How a score should be treated by artists, judges and moderators

The same number means different things depending on who holds it. Each role needs a rule for what the score justifies and what it does not.

What an ai image detector score justifies for each role, and what it cannot justify on its own.

Role	What a score justifies	What it does not justify
Artist	Checking that your original export, not a compressed repost, was tested; gathering your process record; a short factual reply	Re-exporting or editing the piece to change the number; arguing about the tool instead of showing process
Contest judge	A closer look by a trained eye; a private request for process evidence under rules published in advance	Disqualification, public naming, or any decision before the entrant has answered
Moderator	Queuing the post for human review; asking the poster for a layered file or timelapse	Removal or a label based on the score alone, especially on a screenshot or a thumbnail
Client or shop	A polite question about how the piece was made, and a look at the artist's history	Cancelling a commission or withholding payment without process evidence being requested and refused
Curious viewer	Nothing beyond your own curiosity	A public accusation, which the artist cannot undo even after proving the work

If a detector flags your work

Build the record before you need it

Every step above is easier if the evidence already exists. The habit is small: timelapse on by default, layers named as you go, an export at the lines, flats and render stages, and the layered master kept after delivery. That is the whole of how to document your art process, and it costs a few minutes per piece. An artist with this record answers a flag in ten minutes. An artist without it is left arguing about statistics.

The record also changes the tone of the exchange. When someone says your art is AI, a calm reply with three dated crops ends most conversations before they become public. The full playbook, including what to share and what to keep private, is in what to do when someone says your art is AI. If you run a contest or a queue: publish process evidence rules in advance, have a trained eye review flagged work, give the artist a real chance to respond before any public decision, and record how each case was decided.

The honest limits of any AI image detector

No detector and no human judge is perfect; real paintings get flagged and edited generations slip through. That is a strong reason to change the question. Instead of asking what an ai image detector says about a picture, ask what evidence exists of the process behind it. A working artist almost always has that evidence. A generated image almost never does. Built into your habits, that difference protects you far better than any score, and it is the same evidence that a fair judge, following how to tell if art is AI, will ask for first.

Common questions

Are AI image detectors accurate?

Some are good on the kinds of images they were trained on, and all of them make mistakes. The 2024 Chicago study found the best commercial detector performed well but was weaker against deliberately altered images, while expert human judges produced more false positives on real art. Treat any score as a prompt to check provenance and ask for process.

Why did an AI detector flag my art?

Because the finished pixels resemble the detector's generated training pile. Smooth gradients, flat cel shading, a hyper-polished render, heavy platform compression or a small screenshot all push a hand-made image toward the generated side. The detector never saw your layers or your timelapse, so the score says nothing about how the piece was made.

How do I prove my art is not AI?

With process, never with a better score. Keep the original export untouched, then share a timelapse clip, dated work-in-progress crops or a screenshot of the layer panel, and run the file through a Content Credentials verifier if your app attached one. Ask which tool and which file were tested, and note the exchange in case it happens again.

Should I edit my art so it passes a detector?

No. Adding noise or texture to change a number teaches you nothing and makes you look like you have something to hide. Keep the original file exactly as exported and keep the process record instead. A layer panel and three dated stages answer every fair reviewer, and no score is needed.

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): the study of eight detectors, including professional and expert artists.
• NIST AI 100-4, Reducing Risks Posed by Synthetic Content (2024): why detection is one approach among several.
• Content Credentials: Verify: checking the provenance record on your own file.