Sometimes there is strong evidence that an image was generated or substantially edited by AI. In other cases, the available file and context are insufficient. No visual trick or generic detector score should be treated as a reliable answer for every image.
The practical goal is to establish what the evidence supports, not to force every uncertain picture into a binary label.
Define what you are trying to establish
An image can be wholly generated, partly generated or conventionally edited. Those are different questions from whether its caption is accurate. A camera photograph of a real event can be misrepresented as showing a different place or date.
Start with the claim attached to the picture. If the issue is an alleged event, confirming that some pixels were camera-captured does not verify the event description. Our explanation of generated images and edited photographs shows why mixed workflows complicate simple labels.
Use visual clues as leads
Conflicting geometry or incoherent details can justify closer inspection. They do not independently identify the software used. Low resolution, compression and ordinary editing can also make genuine photographs look unusual.
Conversely, a synthetic picture can look coherent. A checklist of yesterday's common errors is not a guarantee against newer outputs. Describe a suspicious detail as a clue and avoid presenting it as a finding about the whole file.
Understand what a detector result means
Automated detectors can make false-positive and false-negative errors. Their usefulness depends on what they were evaluated against and how closely the image being checked resembles that evaluation material. NIST's report on synthetic-content transparency discusses detection alongside other approaches and emphasises that no single technique provides a comprehensive solution.
A displayed percentage is not automatically a calibrated probability that your specific image is synthetic. Read the tool's explanation of its score and limitations. If that explanation is unavailable, report the result narrowly as the tool's output rather than a factual verdict.
Do not average scores from unrelated detectors and call the result a scientific consensus. Their scales and errors may differ, and they may depend on similar underlying signals.
Check the source and available history
Look for the earliest publication you can establish, accompanying descriptions and available original files. A reverse-image search can reveal earlier appearances or changed captions, but finding no match does not prove that an image is new or AI-generated.
Where Content Credentials are present, examine the record and its validation status. Such records concern provenance and integrity, not a guarantee that every real-world claim is true. The C2PA explainer makes that distinction explicit.
The absence of a credential is also inconclusive. Read watermarks and Content Credentials explained before interpreting either presence or absence as a verdict.
Protect private files while checking
Uploading an image to another website creates a separate privacy decision. Do not distribute sensitive or abusive material to multiple tools simply to compare scores. Where the issue involves targeting or harassment, preserve relevant links and seek an appropriate reporting route rather than increasing circulation.
For a low-stakes public image, a cautious description may be enough: “The source is unverified” or “The publisher identifies this as a generated illustration.” For consequential claims, seek stronger corroboration or qualified analysis. Uncertainty is an acceptable result when the evidence does not support more.
Related reading
AI-Generated Images vs Edited Photos: What Is the Difference?
