Begin with the original file and its context
An AI image detector estimates whether an image resembles examples associated with generated content. That estimate answers a narrower question than whether a photograph is truthful. A real photograph can depict a staged event, and an edited image can contain both camera pixels and generated material. Write down the question you are trying to answer before interpreting a score.
Keep the original file unchanged. Record where you obtained it and what the source claims. A screenshot, social platform download or recompressed copy may contain fewer useful clues than the camera file. Ask for the highest-quality source when possible, and compare the picture with independent records of the event.
Know which checks actually ran
Open Image Investigation, select the image, read the analysis acknowledgment and enter the requested context. After analysis, inspect the status of each check. The interface can combine local evidence with a connected detector when the website owner has configured that service. An unavailable engine has not examined your image and must not count as a reassuring result.
Local metadata, hashes and evidence maps can help organise a review. A file hash identifies a particular set of bytes; it does not certify that the depicted scene is authentic. Metadata can be removed or changed. An evidence map can highlight a pattern without establishing its cause. Keep these observations separate from an AI detector assessment.
Interpret confidence without turning it into a verdict
A detector score is not a measurement of how many pixels were edited. A confidence figure is also not automatically a calibrated probability that the image is fake. Its meaning depends on the engine and the assessment it implements. Read the result labels and limitations rather than assigning your own definition to the number.
Genuine photos can be flagged, and generated images can escape detection. Resizing, recompression, screenshots and unfamiliar generation methods can change the evidence available. When signals disagree, record the disagreement and investigate alternatives. Avoid presenting a single score as proof, especially when the conclusion could affect someone’s reputation.
Combine evidence and retain the limitations
Check whether the source account, publication date, scene details and independent photographs support the claimed context. Review metadata when it exists, while remembering that missing camera fields are common after sharing and do not prove AI generation. Compare available detector output with these observations.
Before downloading a report, read and accept the separate report acknowledgment. Retain the original alongside the report and record which checks were unavailable. A useful conclusion states what was observed, what remains uncertain and what further evidence would help. It should distinguish an indication of generation from a claim that the event did not occur.
Privacy and sensible next steps
Core image operations work in the browser, but an activated connected forensic check can send the selected image to an external service. Review the tool’s processing notice before submitting confidential or personal material. Website analytics is separate from the image-analysis workflow; do not put sensitive case details into page URLs.
Use the metadata viewer for a closer look at fields, and use Image Investigation for the broader assessment. If the available evidence is insufficient, an inconclusive result is appropriate. Preserve that uncertainty instead of treating the absence of a warning as an authenticity certificate.
Screenshots, partial edits and the wrong question
A screenshot can replace the original file structure, lose camera metadata and change the pixel pattern presented to a detector. Ask for the original when it exists, and document whether your assessment used a screenshot or a platform download. A clean result for that copy is not a certificate for the earlier file.
A high detector score does not tell you exactly what percentage of the image was AI edited. A small generated region and a fully generated scene can both influence a whole-image assessment. ImgProcess does not provide a validated edited-area percentage. Use completed findings to guide further review rather than inventing that measurement.
