Borderline cases: the images nobody can decide

Why borderline cases exist

The question “real or AI?” assumes two clearly separated groups. In practice the origin is often mixed: a real photo with a generatively filled background, a phone snapshot computed from several exposures and then denoised, an AI image that was compressed, screenshotted and re-uploaded.

Every one of those steps removes evidence and adds artefacts. After three rounds of compression and re-upload, the traces of the original origin are largely gone. That is not a gap in your ability to see — it is a gap in the material.

Case 1: a real photo that looks generated

Modern phone cameras compute. They combine several exposures, brighten shadows, sharpen aggressively and smooth noise. The result can look exactly like the polished output of a generator: even light, no grain, edges that appear too clean.

What helps here is not a closer look but a different question: where does the picture come from, who published it first, and is there a second photo of the same scene taken by someone else?

Case 2: an AI image that copies a camera

Some generated images are deliberately made to look like imperfect snapshots: slight underexposure, film grain, a tilted horizon, motion blur in the corners. Everything people learned to rely on as a clue is reproduced on purpose.

Against such images, stylistic arguments fail. What remains is content logic: does the scene exist, is the lettering readable, do the shadows follow one light source, does the geometry of the room hold together?

Case 3: lettering that proves nothing on its own

Unreadable text is one of the most reliable hints — and at the same time a classic misjudgement. Real photos of signs, shop windows and menus are frequently blurred, cropped, photographed at an angle or written in a script you cannot read.

The fair test is not “can I read it?” but “does this look like writing that failed, or like writing that is simply out of focus and foreign?” The distinction is fine, and that is exactly why it stays a hint.

Case 4: hands, and why they became less useful

Hands were the classic giveaway, because fingers follow strict anatomical rules. Current models handle them well, and at the same time real hands get blurred by movement, hidden behind objects or covered by gloves.

So hands are still a good place to look, but a failed look at hands is no evidence in either direction.

Case 5: the screenshot of a screenshot

Most people do not see images in their original form but as a screenshot of a social post, forwarded through a messenger, saved and shared again. Screenshots strip metadata, content credentials and often resolution.

That means the most common form of circulating image is precisely the one that carries no verifiable origin — which is why provenance questions have to be asked of the posting account, not of the file.

What to do instead of guessing

Name your confidence in levels rather than forcing a verdict: certainly AI, probably AI, unsure, probably real. Then ask for the original, check who published the image first, look for independent reports, and treat the picture as a claim until the claim is corroborated.

• Ask for the original file instead of judging the copy
• Check the publishing account: does it show a history, or was it created last week?
• Look for the same scene from another source — a second photographer is strong evidence for a real scene
• Say “unsure” out loud when you are unsure, and be careful with public accusations

How experts actually decide

In practice, professionals rarely decide by staring at details. They work with corroboration: several independent sources, the earliest publication date, a known photographer, a venue that confirms the event. Inspection supplies hints; corroboration supplies answers.

This is also the honest takeaway of this whole learning area: no single feature — not text, not hands, not grain, not metadata — proves that an image was generated. Several independent observations together, plus verifiable origin, move you from suspicion to a defensible judgement.