How to Tell If an Image Is AI Generated

Start by classifying what kind of file you have, because the method depends on it. An untouched camera original, a social media download, a screenshot from a chat, and a heavily edited image all require different levels of caution and will produce different score ranges from the same underlying content. Establishing this first prevents you from over-reading a weak result.

Then establish intent before you invest effort. Why are you checking this file? A curiosity question needs one pass. A decision involving money, publication, employment, or safety needs the full three layers plus a documented trail. Most wasted checking effort comes from applying a publication-grade investigation to a casual question, or a casual glance to a consequential one.

Work the layers in ascending cost. Provenance requires only asking questions and running a reverse search. Visual inspection requires zooming. The scored check requires an upload. Doing them in that order means the free, most-decisive layer always runs first — and when it produces a clear answer, you may not need the rest.

Finally, write down the decision and the evidence if it affects anyone else. Not for the file's sake, but because a recorded basis is what distinguishes a defensible decision from an assertion, and it is the only thing that survives a challenge three months later.

High-yield tells: hands and fingers, lettering and text within the frame, jewellery clasps and watch bands, partially occluded limbs, and complex repeating patterns. These fail most often because they require the model to solve geometry it has only weakly learned, and they are also the regions that receive the least attention during generation.

Medium-yield tells: shadows pointing in inconsistent directions, reflections showing a scene that does not match the visible one, objects floating without contact or intersecting without occlusion, lighting that seems to come from two sources at once, and implausible water behaviour. These are worth a look but individually explainable by bad photography or editing.

Low-yield tells: overall cleanliness, colour smoothness, and the general absence of sensor noise. These are the most commonly cited and the least reliable — an aggressive denoising filter or a modern phone's computational pipeline produces all three on a genuine capture. Treat them as a reason to look closer, never as evidence.

One more worth knowing: repeated textures. The same brick, leaf, window, or pattern stamped across a frame with identical shape is a strong signal, because real scenes vary and generated ones repeat. It is common in generated backgrounds and it survives compression better than most fine artifacts.

Generators were trained to produce images that look good, not images that obey physics. They render plausible texture and plausible composition, then fail on the details a camera cannot avoid: the exact grain structure of a specific sensor, the way a lens falls off toward the corners, micro-contrast at an edge where two surfaces meet, and the small optical imperfections that every real lens has.

The gap widens when images move through the wild. Every platform recompresses on upload, every messaging app strips metadata and downsizes, every screenshot crops and re-encodes. Each round flattens the high-frequency texture regularities that detection depends on, which is why an image shared across three platforms scores closer to the middle than the same image uploaded directly — regardless of which model made it.

Hybrid pipelines widen it further. A common workflow starts from a real photograph, applies generative fill to remove or add an element, upscales, and finishes with a filter. Only part of the file is synthetic, so scores land in the ambiguous middle — correctly. Binary rules break against this reality, which is why the middle band should trigger more evidence gathering rather than a stronger opinion.

The practical consequence: treat any single answer as provisional. Provenance plus inspection plus a scored check gives a defensible conclusion. A glance no longer does, and neither does a score on a twice-compressed thumbnail.

Match the threshold to the cost of the two error types, because they are rarely symmetric. Publishing a fabricated image as fact and rejecting a genuine one are not equally bad, and a policy that treats them equally will be wrong constantly. For most editorial use, a false positive costs a follow-up email; a false negative costs credibility.

A workable default: low score plus low stakes proceeds; the middle band always queues for human review; a high score holds the action until provenance is confirmed. Publish the policy internally so reviewers apply it identically — inconsistency between reviewers, not the threshold itself, is what makes these systems feel unfair.

Never automate punishment. Scores drift with lighting, camera quality, filters, and skin tone, so an automatic ban or rejection on a threshold alone produces systematically unfair outcomes. The score should route attention; a human delivers any consequence.

Record consequential decisions: the score, the evidence gathered, the reviewer, and the outcome. That record protects people wrongly flagged, lets you tune thresholds against real outcomes instead of intuition, and gives you a defensible position if the decision is later challenged.

No detector can read intent. A generated image may be honest concept art, an obvious joke, or a deliberate fraud, and the pixels do not distinguish between them. Intent lives in the caption, the account, and the surrounding claims — which is why provenance carries more evidential weight than any score.

No detector certifies authenticity. A low score means no strong generative cues were found in this file. It does not mean the story is true, that nobody manipulated the image outside a generative pipeline, or that the person depicted consented to appear. Honest photography and ordinary editing both produce camera-like results.

No detector identifies the creator or the tool. Scores describe statistical similarity to generative output, not authorship or model provenance. For genuine attribution you need platform metadata, C2PA Content Credentials records, or publication history — and for questions of culpability, formal forensic examination with documented chains of custody.

Keeping these three boundaries clear is what stops a useful signal from becoming a dangerous shortcut. The tool answers one question well. The decision belongs to you.