Read the exact output and its scope
An automated detector may report a confidence tier or a claim about a particular watermark system. That is different from proving every pixel is synthetic or identifying the event shown. Record the tool, date and exact result if using it in an investigation. A screenshot without those details can outlive a changed model or a revised explanation.
Understand the recent documented limitation
On 21 October 2026, Full Fact published an investigation in which five apparently genuine images were initially flagged by Google’s detector. After the organisation raised the cases, an update changed their public classification to Unsure. The account is a concrete example of why a detector output should not be treated as an infallible public accusation about a photographer.
Combine independent evidence
Look for the original publication, a credible explanation from the source and other photographs or footage of the event. A detection result can prompt questions, but it should not replace them. Likewise, no detected signal is not proof that an image is genuine: the content may come from another system, have been altered differently or lack an identifiable supported marker.
Use cautious conclusions without abandoning verification
If the available evidence remains ambiguous, state that uncertainty instead of selecting whichever label supports a preferred story. Do not upload private images to an unfamiliar service merely for a dramatic score. When publishing a correction, distinguish what the tool reported from what the surrounding evidence establishes. This protects both the reader and people whose authentic images may have been mislabelled.
Questions, answered
Can a detector result change after an update?
Full Fact documented that occurring in October 2026.
Does one detector label settle every authenticity question?
No. Read its scope and combine it with provenance evidence.