AI Content Verification Checklist

Before you start

This checklist deliberately contains no list of visual artefacts. Fingers, teeth, garbled background text, and similar tells change with every model generation, and a practice built on them has to be rebuilt continuously.

It also does not ask you to run a detector. Detection tools produce both false accusations and false reassurance, and neither error is acceptable when the output is treated as authoritative.

The checklist

Provenance

• Can I reach an origin? A named publisher, an account with a history, a file with a source. If the trail ends at a forward or a screenshot, that is your finding.
• Does the file carry content credentials? Where present, signed provenance data records capture and subsequent edits, as defined by the C2PA specification [1]. Where absent, conclude nothing: credentials are missing from most content in circulation and are routinely stripped on upload.
• Does the account have a history consistent with the claim? An account created recently, posting exclusively about one topic, is a different proposition from an established one.

Independence

• Do the corroborating sources trace back to one origin? Count independent origins, not posts.
• Would this be reported elsewhere if true? For anything consequential, check whether the expected coverage exists.

The specific reference

• Does the cited study or document exist? Search the exact title. Fabricated citations are formatted correctly and name plausible journals.
• Does it say what is claimed? Open it. This step catches real citations used to support claims they do not make.
• Does its scope match the use? A study of one city being used to support a national claim is a scope failure, not a fabrication.

Expected evidence

• What else would exist if this were true? A public statement has a recording. A regulatory action has a filing. A weather event has satellite data. Generated content produces the artefact, not the surrounding record.
• Is there a version with more context? Longer recordings, wider crops, and original uploads frequently resolve the question outright.

Consequences

• What would acting on this cost if it is wrong? Raise your standard accordingly, and be willing to conclude that you do not know.

If you are dealing with a chatbot answer

Treat it as a lead. Ask for sources, open them, and check that they say what the summary reports. A fluent paragraph presents no visible selection process, so there is no ranked list to disagree with and nothing to inspect until you have the underlying sources in hand.

What not to do

Do not accuse a person of using AI on the basis of a detector score or a stylistic impression. False positives fall disproportionately on people whose writing is atypical for a detector’s training data, and the harm from a wrongful accusation is real and immediate.

Version and reuse

Version 1.2, reviewed August 2026. Reuse terms are described on the Terms of Use page.

References

Numbered citations in the text above correspond to the entries below.

1. Coalition for Content Provenance and Authenticity. (2025). C2PA technical specification, version 2.2. C2PA. https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html

Next step

Read the full article on what AI content changes

Topics covered: AI LiteracyInformation Manipulation