1. Start with the original file
Original files can contain useful evidence: camera model, capture time, exposure values, software names, C2PA provenance, or generation workflow data. Screenshots and reposted images often strip that evidence away.
2. Check metadata and provenance
Camera metadata supports a real-photo interpretation, but it is not proof. Missing metadata is also not proof of AI generation. C2PA or JUMBF provenance records can help when they are present, especially if they declare editing or AI creation.
3. Look for generator traces
Some generated images include text fields such as prompt, negative prompt, seed, sampler, model, workflow, ComfyUI, Automatic1111, Stable Diffusion, Midjourney, or DALL-E. These are strong clues when they appear in the file.
4. Inspect visual details carefully
Hands, text signs, reflections, shadows, repeated textures, warped objects, and background people can reveal generation artifacts. But visual inspection can also be fooled by blur, compression, wide-angle distortion, heavy editing, and low resolution.
5. Compare with the source context
Save the source URL, author, publish date, and surrounding claim. Run reverse image search to find earlier copies. A detection result becomes much more useful when it is tied to where the image came from and what it is being used to prove.
6. Separate strong clues from weak clues
Strong file clues include a readable generation workflow, prompt, seed, model name, or a provenance claim that explicitly identifies AI creation. Supporting camera clues include a consistent group of camera model, lens, timestamp, exposure, ISO, and focal length fields. Camera fields support a capture history, but metadata can be copied or edited.
Weak clues include missing metadata, odd fingers, garbled signs, smooth skin, repeated textures, or inconsistent shadows. Each can occur in generated images, but also in screenshots, motion blur, aggressive denoising, panoramas, compositing, or low-quality compression. A responsible conclusion explains which category each clue belongs to.
7. Understand what a pixel model adds
When local file evidence is uncertain, ImgShield may send the current image through its same-origin backend to Sightengine's generative-image model. A pixel-model result is an additional signal, not a record of how the file was created. Models can be wrong on unfamiliar generators, edited images, illustrations, screenshots, and heavily compressed files.
On ImgShield, high and low model ranges can change the displayed direction while the middle range stays uncertain. Read the provider name and score together with metadata and source context. If the consequence is important, use a second independent method and a human reviewer.
Worked example: a reposted event photo
Suppose a social post claims to show a recent event. The downloaded image has no EXIF and a sign in the background looks distorted. Those are two weak clues, not proof. A better review saves the post URL and time, asks for the original, searches for earlier versions, and compares landmarks with other coverage. If an earlier high-resolution copy contains consistent camera data and appears before the disputed post, the context becomes more informative than the missing metadata in the repost.
If the original instead contains a ComfyUI workflow with a model and seed, that is a much stronger origin clue. The final note should say exactly that the file contains a generation workflow—not merely that a detector returned a percentage.
Worked example: an edited product image
A product image may contain Adobe software metadata but no camera fields. This proves only that Adobe software touched the file. Background removal, color correction, and resizing can all strip capture data. Check the seller's source files, look for consistent angles across the product set, and use reverse search. Classify the image as edited or uncertain unless independent evidence supports AI generation.
Document a conclusion another person can review
- Identify the exact file and whether it is an original, export, screenshot, or repost.
- List strong, supporting, weak, and contextual clues separately.
- Record missing evidence, conflicting evidence, and any third-party model result.
- Use calibrated wording: likely generated, likely captured/edited, or uncertain.
- State what would change the conclusion, such as obtaining the original or finding an earlier source.
Use the checker for the first pass
Open ImgShield to inspect one file, then read the methodology and image-processing disclosure. For a team queue, use the manual bulk review workflow.