Start with the least damaged file
If you have a choice, do not upload a screenshot of a screenshot. Use the original image file, a direct download, or the largest version you can get. Every repost can resize the image, remove metadata, and add compression marks that make the result less clean.
For example, a product photo saved from a marketplace thumbnail is weaker evidence than the same image downloaded from the seller's full-size gallery. A profile picture cropped from a chat app is weaker than the original uploaded portrait.
Use the detector score as a triage signal
A high AI signal is a reason to slow down. It is not a reason to make an accusation. A low score means the detector did not find strong generation signals in that file. It does not prove the image is real, recent, unedited, or honestly captioned.
The useful question is: "What should I check next?" A high score points you toward source history, captions, and alternate versions. A low score still leaves ordinary verification work.
Look for visual clues, but do not overread them
Unnatural hands, unreadable signs, repeated textures, strange reflections, melted jewelry, mismatched earrings, overly smooth skin, and impossible shadows can all be warning signs. They are stronger when several appear together.
One odd detail is not enough. Motion blur, camera lenses, beauty filters, HDR processing, bad lighting, and social media compression can make real photos look strange. The best visual review compares multiple areas of the image instead of focusing on one artifact.
Check the source trail
Search for the image title, caption, visible text, account name, and any older copies. If the image is attached to a serious claim, source trail matters more than the detector number. A real image from 2021 can be used falsely in 2026. A generated image can be harmless if it is clearly labeled as concept art.
Use careful wording: "This image needs more review" is usually more accurate than "this image is fake."
A practical order of checks
- Save the best available file and note where it came from.
- Run the AI image detector and record whether the result is high, low, or uncertain.
- Look for metadata: dimensions, software, camera clues, and missing fields.
- Search for earlier copies or different captions.
- Decide whether the image supports the claim being made.
FAQ
Can I tell with 100% certainty?
No. AI image detection is probabilistic, and both generated images and edited real photos can produce confusing signals.
What if there is no metadata?
That is common. Social platforms, messaging apps, screenshots, and export tools often remove metadata. Missing metadata is a clue, not proof.
What is the strongest warning sign?
The strongest warning is not one weird finger or one missing field. It is several independent clues pointing in the same direction.