← All guides Platforms

How to check if a Facebook photo is AI

Facebook recompresses everything it touches, which changes what a detector sees. How to get a usable copy of the image, and how to read the result once you have it.

· 9 min read · Best AI Image Detector

Open the photo at full size and save that file rather than screenshotting the post. Facebook strips metadata and recompresses on upload, so a screenshot of a screenshot loses most of what a check reads and pushes the score upward on genuine photographs.

What Facebook does to a picture

An image posted to Facebook is not the image the poster selected. It is resized to fit the platform's own dimensions, re-encoded with its own compression settings, and stripped of the metadata the original file carried.

All three of those matter here. Resizing throws away pixel detail, re-encoding replaces the original compression fingerprint with Facebook's, and stripping metadata removes any Content Credential that might have settled the question in one step.

The practical effect is that every image on the platform has been pushed slightly toward the middle of the scale. Genuine photographs score higher than they would as originals, and generated images score lower, which compresses the gap between them.

None of this is Facebook behaving badly. Recompression is what makes a site with billions of images load quickly, and metadata stripping is a privacy protection that removes location data most people would not want published. It is simply a cost that lands on verification.

Getting the best available copy

  1. Open the photo itself, not the post

    Click the image so it fills the viewer. The version in a feed is a smaller preview than the one the viewer loads.

  2. Save the image file

    Right click and save, or use the download option where the post allows it. This gets you Facebook's stored copy rather than a re-render of your own screen.

  3. Do not screenshot

    A screenshot re-encodes the picture a second time using your own device's settings, on top of what the platform already did.

  4. Look for the original elsewhere

    If the same picture appears on a personal site, a news page or a photographer's portfolio, that copy has usually been through less processing.

  5. Then run the check

    Drop the saved file in and read both the score and the region map.

What each hop costs you
Camera original
14
After upload
29
Saved from the viewer
34
Screenshot of the post
56
Screenshot, re-shared
68

One real photograph throughout. Every rise is processing, not manipulation.

The same genuine photograph, at each stage of its journey through a social platform.

The bottom two rows are the ones that generate false alarms. A picture that has been screenshotted and re-shared twice can cross the decision line while being an entirely ordinary photograph, and the score is measuring the journey rather than the origin.

Reading the result on a platform image

Adjust your expectations rather than your conclusions. On a file that has been through a social platform, treat the middle of the scale as genuinely uninformative rather than as a weak signal in either direction.

The two ends still carry meaning. A score below twenty on a recompressed file is strong evidence of a photograph, because compression pushes upward and it did not. A score above eighty-five has survived a process that works against it, which makes it more meaningful, not less.

The region map holds up better than the headline number here, because a localised replacement stays localised through recompression even as the absolute values drift. One tile standing well clear of its neighbours is still a finding on a degraded file.

How to read a score on a file saved from Facebook
ScoreOn an originalOn a platform copy
Under 20PhotographPhotograph, more confidently
20 to 45PhotographProbably a photograph
45 to 65UncertainUninformative, ask for the original
65 to 85Probably generatedUncertain, get a better copy
Over 85GeneratedGenerated, more confidently

The checks Facebook itself gives you

The platform carries context a detector cannot. When the post was made, whether the account has a history, whether the same image appears on other pages, and whether the picture is labelled by the platform as AI-generated.

That labelling is worth understanding. Where a generator embeds a marker that the platform reads, a label may appear automatically, and it is a strong positive signal when present. Its absence proves nothing, because most images arrive with no marker at all.

Post history is the check people skip and should not. A photograph posted three years ago, with comments from people who were there, is corroborated in a way no pixel analysis can match. A picture posted an hour ago by an account created last week is not.

When the picture is of somebody you know

The most common reason people check a Facebook photo is that it shows a person they recognise doing something that seems out of character. That case needs the region map rather than the score, because the frame is mostly a real photograph.

It also needs restraint. A high score on a picture of a named individual is not something to post in the comments. It is a reason to ask the person, or to report the post, and to keep the conclusion to yourself until somebody with more information has looked.

Groups, marketplace and shared posts

The context around a Facebook image changes how much the check is worth. A photograph in a community group, posted by a member with a history, sits in a very different evidential position from one in a marketplace listing by an account with no reviews.

Marketplace is where a check earns the most. Listing photographs are frequently taken from elsewhere, and a generated product image is now cheap enough that it appears in listings for goods that do not exist. Run the check, then run a reverse image search, because the stolen-photo case is more common than the generated one.

Shared posts are the weakest case of all. Each share is another opportunity for the image to be screenshotted, cropped or re-uploaded, and by the third hop you are analysing a copy that bears little pixel-level resemblance to what was originally posted.

Where a post has travelled, the useful move is to find the earliest version rather than to check the one in front of you. A reverse image search usually surfaces it, and the original will give a far more reliable reading.

Questions people ask

Why does everything from Facebook score higher?
Because the platform re-encodes every image on upload, and re-encoding is one of the things the model reads as a sign that software touched the pixels. It affects genuine photographs and generated ones alike, which narrows the gap between them and makes middling scores less useful.
Can I check a photo without downloading it?
Yes, paste the page URL and the tool will pull the images out of it. That usually retrieves the same file the viewer loads, which is better than a screenshot. It will not work on posts that are private or behind a login, because there is nothing publicly fetchable.
Does Facebook remove Content Credentials?
Metadata is stripped on upload as a matter of course, which removes credentials along with camera settings and location. Some platforms preserve provenance data selectively, and coverage changes over time. In practice, assume any credential is gone once an image has been posted.
The photo has an AI label. Is that reliable?
A label is a strong positive signal, because it means a marker survived inside the file or the uploader declared it. It is not reliable in the other direction: most generated images carry no marker at all, so the great majority of them are never labelled.
What about profile pictures specifically?
Profile photos get more processing than ordinary posts, because they are cropped, resized and often filtered before upload. Expect them to score higher, and check the account history alongside the picture rather than reading the number alone.
Someone shared a screenshot of a post. Can I still check it?
You can, and the answer will be weak. A screenshot of a post is at least two re-encodings away from the original, which is enough to move a genuine photograph well up the scale. Find the original post and save the image from there instead.