Separate observation from conclusion
State the specific evidence first, then the confidence level it supports. "The reflections do not match the light source and the fingers merge, so I lean toward AI-generated at moderate confidence" is reviewable: someone else can check the reflections and the fingers. "It looks fake" is not, because there is nothing to check. Keeping the two parts separate also stops you from upgrading a weak clue into a strong claim just because you already committed to an answer.
Matching words to evidence
Low confidence
- One weak clue, or several clues that could each have an innocent cause
- Example: a slightly odd hand and no metadata, in a heavily compressed screenshot
Moderate confidence
- Two or more independent clues pointing the same way, with no strong counter-evidence
- Example: mismatched reflections plus merged fingers plus text that dissolves into shapes
High confidence
- A strong positive signal, such as a verified credential naming a generative tool, or motion that breaks physics across many frames
- Example: a content credential whose signature verifies and names a generative model
When to stop
If the evidence is mixed or thin, say so and say what would change your mind — a higher-resolution original, the source file with metadata intact, or an independent recording of the same event. Committing to a confident answer on weak evidence is worse than an honest 'uncertain', because it is harder to correct later.