Practical guide

How to tell if an image is AI generated

No single trait proves an image was made by AI. But generated images break in a consistent order: they convince at a distance, then fall apart in the detail. Work from coarse to fine, and use a detector score as a pointer for where to zoom — never as a final verdict.

1. Start with the story, not the pixels

Ask: does this image fit whoever posted it? A private seller with a flawless studio backdrop, a dating profile with only model shots, a news photo with no traceable source. Context is the cheapest check there is.

2. Zoom all the way into text

Letters are the weakest point of nearly every generator: care labels, street signs, book spines, logos, displays. Text that looks like language but forms no words at full zoom is a strong signal.

3. Count and trace physical details

Fingers, teeth, earrings left versus right, zips that lead nowhere, a necklace that disappears under a collar and returns displaced. One anomaly can be chance; a pattern of anomalies is not.

4. Check light and reflection

Shadows falling in two directions, a mirror or window with no photographer in it, eyes with mismatched catchlights. Generators paint light that looks good, not light that comes from somewhere.

5. Find the image elsewhere

Drop it into Google Images or Bing. If it appears under another name, seller or date, the current presentation is wrong regardless of how it was made.

6. Ask for the unrepeatable

One extra photo with a detail you choose, a short live video, a note with today's date. Whoever owns the subject or was there delivers that within a minute; a generated image cannot be reproduced on request.

7. Use the detector as a triage, not a judge

The image detector flags patterns common in synthetic imagery. A high score says: zoom here. A low score says only: this detector saw no strong known pattern. Neither is proof of origin, authorship or intent.

What the measurement says

On the measured set, the current lane drew a strong warning on 78.1% of 653 AI images and a false warning on 4.0% of 795 real ones, with the newest generators as the weakest spot. Full figures, samples and intervals are on the accuracy page.