The advice everyone repeats is to count the fingers. That worked in 2023. Current image models get hands right most of the time, and the tells have moved somewhere less convenient.
Nothing below is a guaranteed detector. Treat these as reasons to check further, not verdicts.
1. Read the text in the image
Generated text is still the most reliable giveaway. Signboards, number plates, book spines, shop names, jersey numbers, logos in the background.
Look for letters that are shaped almost right but not quite, words that dissolve into scribble at the edges, or two signs in the same photo using slightly different fonts for the same brand. Real photographs have consistent, readable, correctly spelled text, even when it is blurry.
2. Follow the straight lines
Trace a railing, a roofline, a tiled floor, a road marking. In a real photo these follow consistent perspective and stay continuous.
Generated images tend to lose track. A railing changes thickness partway along, floor tiles shift alignment behind a person, a wall edge does not quite meet the ceiling. Zoom in and follow the line from one end to the other.
3. Check where things meet
Errors cluster at boundaries. Where hair meets a shoulder. Where fingers wrap around a glass. Where a strap crosses a shirt. Where teeth meet lips.
Look for a strap that vanishes behind an arm and reappears in the wrong place, or hair that blends into the background instead of ending in strands.
4. Look at reflections and shadows
Do the shadows point the same way? Is there a light source that explains them? A mirror, a window, sunglasses, a wet road, a glass of water.
Models are notoriously loose with reflections. The reflection is often plausible-looking but shows something slightly different from the scene, or the shadow direction contradicts the visible light.
5. Watch for the too-clean look
Generated images have a characteristic finish: even lighting, no sensor noise, everything in sharp focus, skin with no pores, backgrounds that are attractive but vague.
Real photos have flaws. Dust, motion blur, an ugly plug socket, a chipped tile, someone’s arm cut off at the edge of the frame. If a photo feels like a stock image of something that should have been a snapshot, be suspicious.
6. Count and inspect the background people
Crowd scenes are where generators cut corners. Faces at the back merge, two people share a leg, a person is standing at an impossible angle, someone’s arm ends without a hand.
Repeated patterns are another sign. The same face appearing twice, or three people wearing identical clothes in a way real crowds rarely do.
7. For video, watch the edges and the loops
Play it at quarter speed if you can.
Look at where the subject’s outline meets the background: generated video often shows a faint shimmer or warping there. Watch for objects that change size slightly between frames, jewellery that disappears and returns, and backgrounds that morph when the camera moves.
Also check the length. A lot of AI video is a few seconds long and cuts before anything complex happens.
8. For talking-head video, check the mouth and neck
Lip-sync deepfakes usually replace only the mouth region. The tells are a mouth that is slightly too sharp or too soft compared to the rest of the face, teeth that do not change shape as the jaw moves, and a chin or neck that stays oddly still while the mouth works.
Sound matters too. Synthetic voices tend to have flat breathing, even pacing, and no room acoustics.
9. Reverse image search, always
This catches more fakes than every visual tell combined, and it is the step people skip.
Upload the image to Google Images, TinEye, or Yandex. You are looking for the earliest version of it. If a photo being shared as breaking news turns up on a stock site from four years ago, or on a page in another language from a different event, you have your answer.
Cropped and altered images often still match well enough to find the original.
Check the source before you check the pixels
Most viral fakes fail on context, not on technical detail.
Who posted it first? Is the account new, or does it have a history? Is any established outlet carrying the same image? Does the weather, the season, or the clothing match where the event supposedly happened? Do the details in the caption survive a plain search?
An image with no traceable origin is worth doubting even if it looks perfect.
What about detector tools?
AI detection tools exist and they are unreliable in both directions. They flag real photos as fake, especially ones that have been compressed, filtered, or screenshotted, and they miss recent generations.
Use them as one weak signal among several. Never as proof.
Some images now carry C2PA content credentials, a form of embedded provenance data. When present it is useful. It is easily stripped by screenshotting, so its absence tells you nothing.
Frequently asked questions
Can AI images be detected with certainty? No. Detection is probabilistic and the generators improve faster than the detectors. Verification of source is more reliable than analysis of the image.
Do these checks work on AI video too? The text, lighting, and boundary checks apply to both. Video adds temporal tells: flicker, warping at edges, and objects that change between frames.
What should I do if I think something is fake? Do not share it, including to debunk it. Reverse search it, check whether any reliable outlet has reported it, and report it on the platform if it is being presented as real.
Are watermarks on AI images reliable? Visible watermarks are trivially cropped out. Invisible watermarking is being adopted by some providers but is not universal and does not survive every edit.
