Generated images have improved fast enough that most "how to spot AI" checklists are already outdated. Some visual tells remain, but the reliable methods are contextual.
Visual Tells That Still Sometimes Work
- Text and symbols. Signage, book spines, watches, licence plates and logos are still frequently garbled or subtly wrong, even in otherwise convincing images.
- Repetition in crowds and patterns. Duplicate faces in a crowd, bricks that don't line up, fence posts that merge, jewellery that changes form along its length.
- Physical inconsistency. Shadows falling in different directions, reflections that don't match what's in front of the mirror, light sources that can't exist.
- Anatomy at the joints. Not fingers specifically any more — modern models handle hands well — but where limbs meet bodies, where glasses meet ears, where hair meets shoulders.
- Impossible continuity. A strap that disappears behind a shoulder and re-emerges in the wrong place; a chair leg that doesn't reach the floor.
- Uncanny perfection. Skin with no pores, teeth all identical, symmetrical faces, a "photograph" with no sensor noise, no chromatic aberration and no lens character.
Note the direction of travel: every one of these is being fixed, so treat their *absence* as meaningless. A photo with perfect hands is not evidence of anything.
Metadata and Provenance
More reliable than looking:
- Check EXIF. A real camera photo carries camera make, model, lens, exposure settings and a timestamp. A generated image usually has none of that, or has software fields naming a generation tool. Caveat: metadata is trivially stripped or faked, and social platforms remove it, so absence proves nothing.
- Look for Content Credentials (C2PA). A growing number of cameras, editing tools and generators attach a signed provenance record describing how an image was made and edited. Where present, this is the strongest available signal.
- Reverse image search. Find the earliest appearance and the original context. A "news photo" that exists only on one account, with no wire-service or publisher copy, deserves suspicion.
- Check the source, not the image. Who published it? Do they have a record? Does any independent outlet have the same scene from a different angle? For events, multiple independent images are the strongest evidence.
Why Detector Tools Are Not Enough
Automated AI-detection services report a probability, not a verdict. They produce both false positives (flagging real photos, especially heavily edited or upscaled ones) and false negatives (missing generated images that have been re-compressed or screenshotted). Screenshots and social-media recompression strip exactly the statistical traces they rely on.
Use them as one input among several. Never use one as sole grounds for accusing a person of fabrication.
A Practical Verification Order
- What does it claim to be? The stakes decide the effort. A stock-looking landscape needs no investigation; a photo purporting to document a news event does.
- Where did it first appear? Reverse image search for earlier copies and original context.
- Does the provenance data exist? Check Content Credentials and EXIF.
- Does it corroborate? Are there independent images or reports of the same event?
- Only then, look at the pixels for the tells above.
The honest summary: for a well-made generated image with no provenance data, there may be no reliable way to tell from the file alone. That's precisely why provenance standards exist, and why the durable skill is verifying sources rather than inspecting pixels.