How to tell if an image is AI-generated

Ten checks you can run on any photo in a couple of minutes, ordered by how quickly they usually pay off — followed by the cases where your eyes are not enough and a forensic scan is worth running.

The 10 checks

1. 1

   Zoom to 100% before judging anything

   Almost every tell lives in fine detail, and thumbnails hide all of it. View the largest version you can find and inspect regions one at a time rather than reacting to the whole image.

2. 2

   Read any text in the frame

   Signage, labels, keyboards, book spines and license plates are still the fastest giveaway. Generated text tends to be almost-letters: plausible shapes that spell nothing, inconsistent letter spacing, or a font that changes mid-word.

3. 3

   Count and follow the hands

   Check finger count, joint direction, and whether fingers actually wrap the object they hold. Also check where limbs disappear behind bodies and reappear on the wrong side.

4. 4

   Test the lighting story

   Pick the brightest highlight, decide where the light must be, then verify every shadow agrees. Missing contact shadows under feet and objects, and two subjects lit from opposite sides, are strong signals.

5. 5

   Check reflections and transparency

   Mirrors, windows, sunglasses, water and glossy floors are hard to synthesize consistently. A reflection that omits the subject, or duplicates it at the wrong angle, is close to conclusive.

6. 6

   Look at texture and skin

   Generative output often lands on plastic, evenly lit skin with no pores, hair that merges into strands with no individual ends, and fabric weave that repeats like wallpaper.

7. 7

   Hunt for a local edit seam

   If only part of the photo is fake — an added helmet, watch, person or logo — sharpness, grain and colour temperature change at that object's boundary while the rest of the frame stays consistent. This is the most common real-world case and the easiest one to miss.

8. 8

   Judge the background separately

   Cover the subject and study the background alone. Melted architecture, repeating crowd faces, doors to nowhere and warped straight lines show up clearly once the subject is not drawing your eye.

9. 9

   Chase the provenance

   Reverse-image search the file, look for an earlier or higher-resolution version, and check whether the account posting it has history. Inspect metadata and any C2PA content credential — present provenance is meaningful, though absent metadata proves nothing since platforms strip it.

10. 10

    Run an automated forensic check

    Human eyes miss sensor-level evidence: noise that is too uniform for a real camera, resampled compression structure, and detail that survives where a lens would lose it. DetectFake analyses these and returns a verdict, a confidence score and every signal it found so you can weigh them against your own inspection.

Tells that are no longer reliable

Advice written for early image models has aged badly. Six-fingered hands, garbled eyes and watermark corners are all largely fixed in current generators, and "it looks too perfect" flags plenty of genuine studio photography. Treat those as weak hints at most, and put your weight on lighting consistency, reflections, local edit seams and provenance.

FAQs