How to check whether an image is AI-generated (and why it's hard)
Published 10 October 2026
There's no reliable test that tells you whether any image was made with AI. There are clues, though, and the most useful ones are in the file's metadata, when they survive.
Clues in the metadata
Many AI image tools label what they make. Our inspector looks for these labels and lists them under "AI labels found":
- Content Credentials. A signed record, added by some generators and editors, of how an image was made. It can say an image was "created with generative AI" or that parts of it were generated. What Content Credentials are.
- The IPTC digital source type. A standard field that says whether an image was captured by a camera, edited, or generated. "Trained algorithmic media" means AI-generated.
- Generation settings. Some tools, such as Stable Diffusion web interfaces and ComfyUI, save the prompt and settings inside PNG files.
- Tool names. The software fields can name the generator, such as Midjourney, DALL·E or Firefly.
When one of these is present, it's a strong hint. It's still metadata: anyone can add or edit it.
Why missing labels prove nothing
Most images you see online have no metadata at all. Social networks, messaging apps and many websites remove it when images are uploaded or compressed, and a screenshot never has any. An AI image that was shared a few times has usually lost its labels, and someone who wants to pass off an AI image as real can remove them in seconds.
So "no AI labels found" means only that: the file doesn't say. It doesn't mean the image is real.
Why tools that look at the pixels aren't reliable
Some services claim to detect AI images from the pixels alone. They make mistakes in both directions: real photos flagged as AI, AI images passed as real, especially after resizing or compression. Each new generator makes the old patterns less useful. A detector's percentage isn't evidence.
Invisible watermarks, such as Google's SynthID, are more reliable for images from the companies that add them, but they can only be checked with that company's own detector, and only for its own tools.
That's why our inspector doesn't analyse pixels: it reports what the file says, and says when the file says nothing.
What you can do
- Check the metadata, if you have the original file. Labels and Content Credentials are the clearest evidence when present.
- Find the source. A reverse image search can lead to the first place an image appeared, and who posted it.
- Look for context, not just artefacts. Does the image match other photos of the same event, from different people?
- Look closely, knowing that the classic tells (hands, text, reflections) are getting rarer.