← All notes Guides

How to tell if a photo is AI-generated

Four ways to check whether a picture came out of a camera or a model, what each one proves, and the order to try them in when it actually matters.

· 6 min read · Original or AI

Look at the picture, check where it came from, read the file, then score it. Only the last two survive a screenshot, and only the score works on a picture you cannot trace.

There are four ways to answer this question and they are not interchangeable. Each one fails in a different situation, and the situation you are in decides which to reach for. Most advice online covers the first and stops, which is why it works on the pictures in the article and not on the picture you actually have.

Here they are in the order that costs you the least effort first.

1. Look at it properly

Your eyes are free and they catch a fair share. Generators still struggle with anything that has to obey a rule: the number of fingers on a hand, the way a reflection has to contain what stands in front of the mirror, whether the shadows in a scene agree about where the sun is.

  • Hands, teeth and ears, which are expensive to get right and easy to get subtly wrong
  • Text on signs, packaging or screens dissolving into shapes that resemble letters
  • Jewellery and glasses whose frames change thickness or fail to meet behind the head
  • Repeating patterns such as brick, tiling or fabric that drift out of alignment
  • Skin lit evenly across the whole face, without the falloff a real lens produces
  • Hair that merges into the background rather than ending in separate strands

2. Find out where it came from

Provenance beats analysis whenever it is available. A reverse image search that finds the same picture on a stock library three years ago settles the question faster than any model can, and a picture whose earliest appearance is an anonymous account posted an hour ago tells you something too.

Ask the person who sent it for the original file rather than the copy you were forwarded. Someone who took a photograph can produce the version straight off their phone, with its full resolution intact. Someone who downloaded it cannot.

This is the strongest method and the least available one. It needs the picture to exist somewhere else, or a person willing to answer.

3. Read the file itself

Two kinds of evidence can be embedded in a picture. EXIF metadata records the camera, the lens and the settings. Content Credentials, the C2PA standard, add a signed record of what made the file and what was done to it since.

When a valid Content Credential is present it is genuinely strong evidence, because it is cryptographic rather than statistical. The problem is how rarely it survives.

Off the sensor Exported Posted online Screenshotted EXIF camera data Survives Survives Lost Lost Content Credentials Survives Survives Lost Lost Invisible watermark Survives Survives Survives Lost Pixel signal Survives Survives Survives Survives

Most platforms strip metadata on upload. A screenshot creates a new file carrying none of it.

What is still attached to a picture at each stop on its way to you. Only the bottom row is present by the time most pictures need checking.

A text editor can also write "Canon EOS R5" into a generated picture in about four seconds. Treat metadata as a bonus when it survives, never as a foundation.

4. Score the pixels

This is the only method that still works on a picture that has been posted, re-saved, screenshotted and forwarded, which describes almost every picture anybody needs to check. A model trained on both populations reads the statistical residue that generation leaves behind and returns a likelihood.

  1. Get the best copy you can

    Compression removes the fine texture the analysis depends on. The original file scores more reliably than a screenshot of it, often by several bands.

  2. Read the number and the band together

    A score means nothing without the scale it sits on. Bands are published and fixed, so 72 today means what 72 meant last month.

  3. Check the region map before you act

    An even reading across the frame and one hot region among eight quiet ones can produce a similar headline number and mean completely different things.

Which method survives which situation
SituationEyesProvenanceFilePixels
Screenshot from a chat appPartlySometimesNoYes
Picture posted to social mediaPartlySometimesNoYes
Original file from a phonePartlyYesYesYes
Heavily compressed forwardPartlySometimesNoWeaker
Anonymous account, no historyPartlyNoNoYes

The pattern in that table is worth stating plainly. The methods that prove the most are the ones that vanish first. By the time a picture is contentious enough that somebody wants it checked, it has usually been through the exact processing that removes the strongest evidence, and the pixel score is what is left.

That is not an argument for trusting the score more. It is an argument for reading it as one input rather than as an answer, and for chasing the original file whenever chasing it is possible.

The order in this article is deliberate. Your eyes cost nothing and rule out the easy cases. Provenance costs a search and settles the question outright when it works. Reading the file costs seconds and occasionally hands you a signed record. Scoring the pixels always returns something, which is why it comes last rather than first: it is the fallback that works when the better methods have run out.

Putting them together

No single method is sufficient, and the good news is that they fail in different places. Your eyes work on an obvious fake and fail on a careful one. Provenance is decisive when it exists and useless when it does not. The file evidence is proof when present and absent most of the time. The pixel score always returns something, and what it returns is a likelihood rather than a verdict.

Where a decision affects a real person, use all four and stop before treating any of them as proof. A statistical model with a published error rate is not a forensic examination, and the gap between those two things matters most in exactly the cases where the stakes are highest.

Questions people ask

Can I tell just by looking?
Sometimes, and less often each year. The visible tells are the parts generators are actively improving, so a picture that passes your eye today would have failed it two years ago. Use your eyes as a filter for what to check properly.
Does checking the metadata work?
Rarely, because it is usually gone. Nearly every platform strips metadata on upload, a screenshot removes it entirely, and what remains can be edited by anyone with a text editor. A valid Content Credential is strong evidence; its absence proves nothing.
What if the score comes back in the middle?
Then the honest answer is that the picture cannot be called either way. Screenshots, heavy compression, upscaling and strong filters all push genuine photographs into that band. Find a better copy of the same picture before concluding anything.