The quick answer
Sometimes there is clear evidence, such as a disclosed editing record. Appearance alone or a detector score is not enough to prove how a photo was made.
Check the original source and any recorded history. If the evidence is missing, say that the editing method is unknown.
In this article
Keep these in mind
- Separate a visible mistake from proof that AI caused it.
- Content Credentials can document history, but do not prove the scene is true.
- Use several pieces of evidence before making an accusation.
Decide what you need to know
“Was AI used?” covers very different changes. A photo might have a removed dust spot, a generated sky or a completely invented person. Ask about the part that matters to the job.
For a property listing, the key issue may be whether the room and view are accurate. For a portrait, it may be whether the face was changed. Naming the concern makes the check more useful than asking for a single real-or-fake label.
Treat visual clues as questions
Look for unreadable labels, broken reflections, strange edges or repeated textures. Compare these areas with the original when available. A visible mismatch tells you to investigate that area.
It does not establish the editing method. Blur, heavy retouching and a poor composite can also look odd. In a photographer discussion, a client suspected AI because of polished headshots, while replies questioned that conclusion. Do not diagnose someone’s process from smooth skin alone.
Read more: User discussion: concerns about overly polished headshots
Read Content Credentials carefully
Content Credentials are records attached to a file. They can show how an image was captured and edited. Check who signed the record and which steps it describes. It may cover only part of the image’s history.
The Content Authenticity Initiative explains that credentials are not a truth label. A photo of an AI image can still be a real camera capture. Missing credentials also do not prove deception: the file may never have had them, or that history may not be available.
Read more: C2PA: Frequently asked questions · Content Authenticity Initiative: What Content Credentials show
Use detector results as limited evidence
An AI detector can wrongly flag a real image or miss a generated one. NIST’s report explains why results depend on the test and how the tool is used. False accusations can harm people.
Ask which pictures the detector was tested on. Which image types can it check? How often is it wrong? A percentage does not answer all those questions. Check the detector’s data rules before uploading private client files.
Follow a short verification process
Keep the image and the claim about it separate. Even an untouched photo can have the wrong caption or date.
- Find the earliest reliable source you can, rather than relying on a repost.
- Ask the creator for the original and a description of important edits, when appropriate.
- Inspect available Content Credentials and compare source files.
- Check visible facts against independent references for the place, product or event.
- State what the evidence supports and what remains unknown.
Make your own process easier to understand
Keep an original, a layered working file and a short change log. With permission, show a before-and-after crop that explains the edit. For example: “Photographed product; generated room; original label retained.” Use that wording only when it matches the work.
Preserve supported Content Credentials and check the exported file. Share only the source material your client allows. A clear account of your work is more useful than claiming that an image is impossible for a detector to question.
Sources & further reading
Product features can change. Check the linked documentation for the version and service you use.