Do not rely on fingers, teeth or a single detector score. Start with the source and claim, check where the image appeared before, inspect available provenance or Content Credentials, and treat a missing signal as inconclusive rather than proof that the image is real.
Start with source and context
The question is not simply whether something looks AI-generated. Start with who published it, where it appeared first, what independent evidence supports it and whether provenance information survives from the original file.
Modern image generators have reduced many of the obvious visual artefacts that older spotting guides relied on. That is the anchor for the rest of the check: build confidence from source, context and corroboration rather than trying to force a verdict from one clue.
What matters in practice
Modern image generators have reduced many of the obvious visual artefacts that older spotting guides relied on.
Synthetic media quality keeps improving, while genuine media can look strange after compression, cropping or editing. That makes appearance alone a poor basis for a confident verdict.
Google can surface earlier uses and creation/editing information for some images through About this image.
Provenance can strengthen a conclusion when it is available, but missing provenance is not proof of fakery. Treat it as one piece of evidence alongside the original source and independent corroboration.
Content Credentials can provide cryptographically verifiable provenance when a participating creation or editing workflow preserves them.
The consequence of being wrong should set the verification threshold. A harmless meme can tolerate uncertainty; a claim that could move money, damage a reputation or trigger panic deserves a much stronger evidence trail.
Why visual clues are not enough
Verification rarely produces perfect certainty from one clue. Strong conclusions come from several independent pieces of evidence that point the same way, while missing or conflicting evidence should lower confidence.
- No provenance signal does not prove an image is genuine because many real images have no credentials and platforms may strip metadata.
- A valid provenance record describes a file history; it does not by itself prove that the depicted event or claim is true.
What could change the answer?
Confidence should move when a better source appears, provenance becomes available or genuinely independent reporting confirms or contradicts the claim. A good verification process leaves room for the answer to change.
What you can do now
Slow the claim down. Find the original source, look for context and corroboration, and increase the evidence threshold when being wrong would have serious consequences.
- Check the original account, page and surrounding claim before analysing pixels.
- Use reverse image/context search to look for earlier or contradictory uses.
- Treat detector percentages as supporting evidence, not a verdict.
A four-question reality check
When a claim sounds dramatic, work through these four questions before deciding how much weight to give it.
What can the system demonstrably do?
What tools, credentials and networks can it actually reach?
Can it keep operating after errors or intervention?
What real-world outcome follows if safeguards fail?
Bottom line
A good verification process is comfortable with “not enough evidence yet.” That is usually safer than forcing every suspicious image, video or claim into a confident real-or-fake label.
Do not rely on fingers, teeth or a single detector score. Start with the source and claim, check where the image appeared before, inspect available provenance or Content Credentials, and treat a missing signal as inconclusive rather than proof that the image is real.
Sources & further reading
These links are selected for the claims and methods discussed above. Primary or official sources are preferred where available.
Two useful questions
What is the biggest caveat?
No provenance signal does not prove an image is genuine because many real images have no credentials and platforms may strip metadata. A valid provenance record describes a file history; it does not by itself prove that the depicted event or claim is true.
What is the most useful next step?
Check the original account, page and surrounding claim before analysing pixels. Use reverse image/context search to look for earlier or contradictory uses.