When a business publishes an image or recording, it needs a record of where the file came from, what changed, who approved it, and which claims were checked. An “AI-made” label answers only a small part of that review.
Watermark detection can add useful information. Google describes its SynthID Detector as a way to check images, video, and audio for AI content from Google and participating partners. Its scope depends on the systems that apply supported watermarks.
A detected watermark can support a conclusion about how a file was made. An absent watermark leaves that question open. The result also says nothing about whether a product specification, caption, or spoken statement is accurate.
Content Credentials provide a related kind of provenance record. The C2PA standard can bind signed statements about a file’s origin, edits, tools, and use of AI to the asset. The standard is explicit about the limit: verified provenance does not tell you whether the content itself is true, accurate, or factual. Provenance can also be incomplete, and the metadata can be removed.
That distinction matters. A file can carry a valid record and still contain a bad claim. The person who created it may not have permission to use a logo, image, likeness, or voice. A real product photo may be paired with inaccurate specifications. A synthetic rendering may be completely appropriate if it is labeled correctly and checked against the actual product.
I would build the review around five records.
First, keep the original. Preserve the source file, the date, the tool, and the person or team responsible for it. If a prompt or reference asset materially shaped the result, keep that too.
Second, inspect the available provenance. Check Content Credentials and supported watermark systems. Record what the tool found, what it could not evaluate, and which ecosystem the result applies to.
Third, verify the content. Check names, dates, product features, performance claims, citations, trademarks, and any real person shown or imitated. This is where fact-checking, consent, licensing, and claims review happen.
Fourth, approve the destination. A visual that is appropriate in an internal concept deck may need different labeling or evidence before it becomes an advertisement, a sales document, a training asset, or public education.
Fifth, preserve the approved export. Keep the final file with its source and approval record so the next person does not have to reconstruct the decision from a compressed image downloaded from social media.
Dental technology makes the boundary easy to see. A synthetic product rendering can help someone understand the shape or workflow of a device, provided it matches the actual product and is identified as an illustration. Clinical evidence should use genuine images with their source records preserved. Generated illustrations should be clearly labeled and kept separate from patient findings. Diagnostic accuracy, regulatory status, and clinical performance each need their own supporting records.
Keep the detector result beside the approved file, then retain the separate checks for claims, permissions, and intended use. The next person reviewing the asset should be able to see what was verified and which questions the watermark could not answer.