AI Image Limits Checklist

For the method in What AI Image Tools Get Wrong, and Why Detectors Don’t Fix It. Use as a supporting signal only - never as a verdict.

Not for

Deciding with certainty whether a specific image is AI-generated. No checklist or detector can give you that. Use this to organize a weak signal, and prioritize finding the original source instead.

Step 1: Visual patterns worth a second look (weak signal only)

• Hands: extra/missing fingers, impossible poses
• Text in the image: garbled letters, inconsistent fonts
• Counting: wrong number of identical/repeated objects
• Consistency: same subject looks different across multiple images/frames
• Lighting/reflections: shadows or reflections that don’t physically match

None of these alone or together prove anything. Move to Step 2 regardless of what you find.

Step 2: Provenance check

• Does the file carry a Content Credential (C2PA)? If yes, what does it assert?
• Is there a platform AI-content label on the post?
• Does the metadata (if any) match the claimed date/location?

Step 3: Source check (the most reliable step)

• Reverse image search performed
• Original poster/account/outlet identified
• Original date and context compared to the claim
• Independent second source found (or noted as absent)

Step 4: Detector tools (if used at all)

• Detector name and score recorded: ______________________
• Explicitly NOT treated as the deciding factor
• Decision based on Steps 2-3, not this score

Final decision

☐ Likely genuine, source confirmed ☐ Likely synthetic/manipulated, source confirms or no credible original found ☐ Undetermined - documented as such, not resolved with a guess

Do not delay an urgent action (like reporting harmful content) while trying to resolve “undetermined.”