How Facebook and Instagram Know Your Image Is AI (The Full Detection Chain)

There is a specific kind of frustration in posting a photograph you took yourself, with a camera, of a real place — and watching the platform quietly tag it as AI.

And a mirror-image frustration: carefully cleaning up a generated image, posting it, and getting labelled anyway.

Both happen for the same reason. Meta does not run one AI test. It runs a sequence, and the first thing that fires wins. Once you can see the sequence, both outcomes stop being mysterious.

This is a spoke of our AI image watermarks pillar guide.

The cascade, in order

Steps 1 to 3 are the container. Steps 4 and 5 are the picture. Step 6 is you. That division explains essentially every confusing labelling outcome people report.

Steps 1-3: what your file says about itself

This is where most labels originate, because it is the cheapest possible check — no model inference, no ambiguity, just read a field.

• A C2PA manifest is a cryptographically signed record of how the file was produced, including the generating tool and sometimes the edit history.
• The IPTC digital source type field is a single declaration: this content is synthetic, or composite, or captured.
• Generator tags in EXIF and XMP name the software and often the model version, written automatically by generation and editing tools.

None of these render in the picture. All of them are read on upload in milliseconds.

And all of them vanish when an image is decoded to raw pixels and re-encoded. Not scrubbed field by field — the entire container is simply absent from the new file.

For the full anatomy of these fields, see our AI image metadata guide and C2PA Content Credentials explained.

Step 4: invisible watermark detection

Some generated imagery carries a statistical watermark distributed through the pixel values. Platforms participating in provenance initiatives can test for the schemes they have access to.

This step does not care about your metadata and is not affected by removing it. It is also the step nobody can honestly promise to defeat — robust watermarks are built and tested specifically against compression, resizing, cropping and screenshots. Our SynthID guide covers why.

Step 5: the classifier

Independent of every signal above, a model looks at the pixels and estimates whether they look generated. Common tells include over-smooth skin and texture, physically inconsistent lighting, frequency-domain regularity that cameras do not produce, and the structural oddities that generation still leaves behind.

This is probabilistic. It misses real AI images and it fires on real photographs — which is the mechanism behind almost every "why is my genuine photo labelled?" complaint, particularly for heavily retouched or upscaled work.

Our pillar guide on AI image detection covers how these classifiers work in depth.

Step 6: your own disclosure

The simplest input of all, and the one with the fastest-growing legal weight. In several jurisdictions, disclosing synthetic content is moving from courtesy to obligation — see AI image labelling laws and agreements.

Why partial cleanup fails

Two very common mistakes, both now explainable in one line:

• Cropped the badge, kept the metadata. The badge was never the trigger. Step 1 fired.
• Stripped the metadata, still labelled. Steps 1-3 were cleared; step 4 or 5 fired.

What to do before you upload

1. 1Inspect. See every provenance signal your file carries — most people are surprised by the list.
2. 2Strip the container. Re-encode from raw pixels. Steps 1-3 cleared, with certainty.
3. 3Handle any visible mark. Crop or patch it.
4. 4Read the estimated detection likelihood. That is your honest read on steps 4-5.
5. 5Decide on disclosure knowingly rather than by accident.

Read next: the laws and agreements forcing platforms to label at all →