Independent guidance · AID Editorial Team
Updated October 11, 2026
An AI image detector estimates whether an image shows signals associated with generation or manipulation. That is a different problem from detecting generated prose. Our published text tests cannot establish which image detector performs best.
Two kinds of evidence to keep separate
Detection signals
An image-analysis product returns its assessment of the submitted pixels or other available information. A low or high score remains a model output. Record the exact file, selected mode, date, and explanation so a later reviewer knows what was checked.
Provenance and editing history
Content Credentials use the C2PA specification to attach verifiable claims about an asset and its history. Such credentials help assess the recorded provenance; they do not certify that every depicted event is true. A missing credential does not itself prove AI generation.
Winston AI: a documented image workflow to explore
Winston’s image-analysis guide distinguishes Basic and Advanced scans. It lists Basic scanning on Essential, Advanced and Elite, with Advanced scanning limited to Advanced and Elite. Its trial includes up to five Basic scans and one Advanced scan. Terms were checked October 11, 2026 and should be confirmed in the account.
The vendor recommends an original, uncompressed file where possible. Its supported-image guidance explains coverage and limits. Treat these as vendor descriptions. We have not tested those modes, independently measured their accuracy, or assigned Winston an image score.
Read our Winston AI text review if you are considering the wider product, but keep its 4.6/5 text score separate from image performance. The same boundary applies to our Originality.ai review and our GPTZero review. Their inclusion here provides context for our existing coverage; it is not a claim that all three supply equivalent image tools.
A practical image-checking workflow
- Find the earliest available source. Record who published the image and what they claim it depicts.
- Preserve the original file. Keep a copy and note any resizing, cropping, compression, screenshots, or edits.
- Inspect available provenance. Check the signer and recorded changes where credentials exist; preserve the verification output.
- Choose a tool that supports the image type. Do not assume a model trained for photographs has been validated for diagrams, artwork, or screenshots.
- Run a documented check. Save the file, scan mode, explanation, and date. Do not repeatedly modify the image to chase a desired result.
- Investigate conflicting evidence. Look for corroborating sources and editing history. State uncertainty when the available evidence cannot settle the question.
What a useful image benchmark should contain
Before AID publishes a ranking, we need known-origin human photographs and generated images, documented generation and editing processes, and a consistent scoring protocol. The test should distinguish original files from cropped, compressed, and edited variants. Repeating the same photograph in several formats does not create several independent sources.
A useful report should identify false positives, missed generated images, and ambiguous cases. It should also state which content types and model families were tested. Text-detection credit units, outcomes, and confidence labels cannot fill those gaps.
Frequently asked questions
Can visible oddities prove an image is AI-generated?
They may guide an investigation, but an unusual detail alone is insufficient. Editing, compression, and ordinary photographic effects can also produce unexpected appearances.
Does a missing Content Credential prove anything?
No. Assess the available source and history rather than treating absence as a generation verdict.
Which tool has the highest AID image score?
None yet. Our image evaluation remains uncompleted. For finished text evidence, visit our text shortlist or our comparison.
Editorial basis: AID’s published initial tests and linked official guidance. Vendor feature descriptions are not independently tested findings.