Start with the original source and its disclosure, inspect available Content Credentials, and use a relevant watermark verifier when you know which system may have created the image. Treat an inconclusive result as unknown. Visual appearance alone is not a reliable verdict.

Start with the question you need answered

“Was this AI-generated?” is often a proxy for something more concrete: is a product actually available, did an event happen, or is this person being misrepresented? Decide what evidence would answer that question. Even an ungenerated photograph can be old, staged, edited or accompanied by a false caption.

A practical checking sequence

  • Preserve the received file and note the original post or sender. Screenshots discard useful context.
  • Ask for the original capture or published source when it is available.
  • Inspect any Content Credentials, including the signer, edit history and validation warnings.
  • If the suspected generator uses an identifiable watermark, consult that provider’s official verification service.
  • Compare the claim with independent records before repeating it.

A generator label is a statement, not a verified signature

An ordinary software name, comment or XMP field can be written or changed. Finding a generator name is a lead to investigate, not automatic proof of who made the picture. A metadata reader displaying that text has not authenticated it.

For signed Content Credentials, use a validator and read its warnings, signer information and recorded actions. Even valid provenance does not prove that a caption is accurate or that the entire history is available. Our metadata tool does not perform this signature validation.

What a watermark check can establish

Google’s SynthID embeds signals into generated media and offers verification for supported content. Its scope is particular watermark systems, not every AI image ever made. A negative or uncertain result is not a certificate that an image came from a camera.

How to interpret a detector percentage

A classifier score is not a measurement of how much of the picture was generated. Read what the provider says the number means and which content it has evaluated. NIST identifies false positives and false negatives as important risks and highlights differences in subject matter, generators, compression and resizing.

Treat a result as one signal. If two detectors disagree, retain both results and investigate the source; averaging their percentages does not resolve the disagreement. Do not turn a single score into an accusation about a photographer.

Avoid a visual checklist verdict

Odd hands, text or reflections may justify closer inspection, but normal-looking details do not establish authenticity. Conversely, motion blur and unusual lighting in real photographs can look strange. Describe the observation you actually have rather than converting suspicion into certainty.

Illustrative example

A striking image of flooding appears without a location. Finding its original context and comparing local reporting may answer the practical question faster than scoring its pixels with a detector. If those checks remain unresolved, describe the location and origin as unverified.

Privacy before uploading

An external verifier may receive the image you submit. Check its terms and avoid uploading private photographs casually. This site’s metadata remover runs locally but is not an AI detector; a clean export provides no authenticity verdict.

Write down the result you actually established

  • Declared AI use: the file or publisher states that AI was used; say whether the statement was authenticated.
  • Watermark detected: name the verification system and preserve its result; do not infer unsupported details about the creator.
  • Classifier suggests AI: record the tool, date, file and score, together with its limits.
  • Origin unresolved: no reliable conclusion after the available checks. Missing metadata or a negative detector result does not change this to “camera original”.

Sources & further reading

Keep exploring