AI-Generated, Edited or Misleading? Four Different Questions

Question one: was the content generated?

This is the question most AI detectors try to answer. Was the image produced by a generative model? Does the voice contain synthetic speech? Does the text resemble machine-generated writing? It is a technical classification about the submitted material.

Even here, the edges are untidy. A person can generate a background, photograph a real product, combine them and retouch the result by hand. A writer can use AI for an outline and rewrite every line. The useful answer may describe which parts show synthetic signals rather than forcing the whole work into one box.

Question two: was it edited or manipulated?

Editing is older than generative AI. Cropping, color correction, noise removal and cutting pauses from an interview all change a file. So do beauty filters, object removal, speed changes and subtitles. Some edits are routine; others change the meaning.

A face swap is manipulation and often uses AI. A misleading crop may not use AI at all. When a detector reports little evidence of generation, it has not ruled out selective editing. Look for cuts, missing context and comparisons with an earlier version.

Question three: where did this copy come from?

Provenance is the history attached to a file: who or what captured it, which tools handled it and what actions were recorded. Content Credentials can carry this information in a tamper-evident form when compatible tools participate. A valid record can be far more direct than guessing from a suspicious shadow.

It still needs careful reading. C2PA explicitly does not make a value judgment that the recorded story is good, honest or important. It helps verify that provenance information is associated with the asset and has not been tampered with. Missing credentials are not proof of AI either; coverage is not universal and platforms can separate files from metadata.

Question four: is the claim true?

A genuine video of a flood in 2022 can be reposted as today's disaster in another country. An authentic photograph can be staged. A real voice note can come from a compromised account. None of those cases requires synthetic media, and a perfect AI detector would still miss the deception.

Write the claim as a sentence, then search its people, place and date. Find the earliest useful publication. Compare it with official records, local reporting or other independent evidence. The caption often deserves at least as much suspicion as the pixels.

A practical order that avoids wasted work

Begin with source and context because they can resolve the case quickly. Next check available provenance. Inspect the file for editing and mismatched elements. Then use specialist detection on the best-quality copy. The order is not sacred, but it prevents a dramatic score from taking control before you understand the claim.

• State exactly what the post claims happened.
• Trace the earliest attributable source and fuller version.
• Check Content Credentials or other available metadata.
• Separate visual, audio, text and face-manipulation findings.
• Finish with a calibrated conclusion and name what remains unknown.

Use four answers instead of one oversized verdict

A careful report might say: 'The visuals show no strong AI-generation signal; the audio is suspicious; no verifiable provenance was found; and the central claim remains unconfirmed.' That sounds less convenient than 'fake,' but it tells the reader what was actually checked.

This separation also protects honest creators. Ordinary editing is not automatically deception, and AI assistance is not automatically fraud. The problem is the evidence, claim and use - not the presence of a fashionable tool by itself.