How to Test if an Image Is AI-Generated: Practical Workflow (2026)

An AI image detector can flag likely AI-generated visuals in seconds. That is useful for social posts, user-submitted photos, product images and viral news content. But one tool alone cannot deliver a final verdict on authenticity — and knowing which part of the answer is evidence and which part is a guess is what makes the difference between a check and a verdict.

This guide covers how to test if an image is AI: what the free check actually looks at, how to read the result, which cases produce false positives, and a repeatable workflow for when the answer matters.

Start with the strongest signal: the file itself

Before any model looks at the picture, read what the file says about itself. This is the cheapest check, it takes seconds, and when a marker is present it is conclusive rather than probable.

• Provenance markers — C2PA / Content Credentials manifests, IPTC/XMP AI markers. A signed generative marker is the one kind of AI evidence that is not a guess.
• Generation parameters — prompt, seed, sampler, model, sometimes a whole node graph, written into the file by the tool that produced it.
• Generator software tags — a Software field still naming the generator after several rounds of editing.
• Contradictions — a camera make and model that postdate the capture date, an editing history that does not match the claimed story, or a capture device that cannot produce the file's dimensions.

Read them with the EXIF Viewer. The absence of a marker proves nothing — plenty of real photos have stripped metadata, and plenty of AI images have had theirs removed.

Then run the pixel-level check

When the file carries no metadata, the remaining evidence is statistical. Detectors look at the places where camera and model pipelines differ:

• Noise floor and PRNU — every sensor leaves a pixel-level fingerprint; generated images are often unnaturally clean.
• Local texture variance — real scenes vary in fine detail; generated surfaces repeat.
• Colour entropy and channel correlation — synthetic images tend to couple their channels more tightly than a camera pipeline does.
• High-frequency energy — a Laplacian-style measure of whether fine detail resembles a photograph.
• Generator-typical dimensions — an image whose edges match a well-known model default is worth a closer look, on its own it means little.

How to read the result

A raw number is not a verdict. The band it falls into is, because that is what a reader can act on:

• Below 30 — Likely Real Photo. Nothing suspicious was found. This is an absence of evidence, not proof of authenticity.
• 30 to 65 — Mixed Signals. Some cues fired, some did not. Look at the file metadata before concluding anything.
• 65 and above — Likely AI. Enough cues fired to treat the file as synthetic unless the source explains why it looks that way.
• Verified provenance — a signed marker was found. This overrides the score entirely, because signed evidence outranks statistical inference.
• Inconclusive — the image was too small or too uniform to measure. A thumbnail or a solid-colour frame reports no score rather than a low one, so an unmeasurable image never looks like a passing one.

The free workflow

1. Read the metadata. Provenance markers, generation parameters, generator tags, internal contradictions.
2. Run the detector on the original file. PhotoRadar's local analysis runs in your browser, so the image is not uploaded. Add the optional Deep Scan only when the softer cues are not decisive.
3. Note the band, not just the number. A move from 28 to 29 changes nothing; a move from 62 to 66 changes the conclusion.
4. Cross-check the source. Reverse image search for the earliest publication, and check whether the account, the timeline and the claimed event hold up. Context resolves what pixels cannot.
5. Record what you did. Score, tool, timestamp, and the exact file you tested. A conclusion you cannot reproduce is not one you can defend.

When AI image detectors get it wrong

False positives are normal, and they cluster in predictable places. If your image falls into one of these categories, treat a high score as a prompt to look closer, not as a conclusion.

• Portraits with beauty filters: smoothed skin and contour edits resemble generated texture.
• Night photos with heavy denoising: flattened texture reads as a synthetic surface.
• Long repost chains: every re-save adds compression artefacts that distort the statistics in ways no clean photo has.
• Screenshots and scans: moiré, sharpening halos and resampling noise all shift the measurements.
• Heavily edited real photos: filters can manufacture the very artefacts a detector looks for.
• Hybrid content: a real photo with an AI-retouched region is genuinely hard, and is the most common reason a confident score is wrong in both directions.

What one tool cannot answer

Even a perfect detector answers a narrow question: does this file carry the signatures of generation? It does not tell you whether a scene happened, whether a person in it exists, or whether an image has been manipulated in ways that leave no generation trace at all. For manipulated-but-not-generated images — splices, removed objects, altered dates — the tools that matter are different ones, such as a tamper heatmap and manual forensics review.

For newsroom and OSINT workflows, the longer checklists in Social media image verification and Image forensics for journalists cover the rest of the chain.

Who should test, and when

• Creators and marketers — check your own output before publishing, so you know what a third-party detector will see.
• Journalists and OSINT investigators — treat the result as one input and document it as such.
• Anyone buying or listing online — a "likely AI" result on a product photo is a reason to ask for a real one, not a reason to accuse anyone.