Treat an alarming clip as a claim to investigate. Confirm its source, context and requested action through a separate channel.
A convincing video can make a claim feel settled before you have checked what it actually shows. An urgent voice message can do the same: recognition arrives first, verification later. AI-generated media makes that gap more consequential. Yet the answer is not to become an amateur image analyst, or to dismiss everything unfamiliar as fake. A more useful habit is to separate three questions: who supplied this material, what event does it claim to document, and what does someone want you to do after seeing it? Each question calls for different evidence.
Synthetic, authentic and true are different questions
A synthetic image is not automatically deceptive. It might illustrate a fictional story and be clearly labelled. A genuine photograph can mislead when paired with the wrong date or location. And an edited clip may preserve the speaker’s actual words while removing the context needed to understand them.
NIST’s November 2024 report treats provenance tracking, labelling and synthetic-content detection as different technical approaches. It explains that transparency can provide useful information about origins without guaranteeing trustworthiness. No individual technique supplies a comprehensive solution. The practical implication is to investigate the underlying claim even when the file itself looks legitimate.
A verification workflow for everyday readers
The following sequence is AI Affect’s suggested reading method. It is not a forensic certification process.
- Write down the claim. Replace “this looks shocking” with a sentence that can be checked: a named person allegedly said something, in a particular place, on a particular date.
- Find the earliest identifiable source. A screenshot of a repost is several steps away from an original recording. Look for the complete item and its publication context.
- Seek independent corroboration. Five accounts repeating one upload are one evidence chain. Look for a separate recording, an official transcript or reporting that explains how it verified the event.
- Check the action being requested. Sharing a clip, revealing account details and transferring money carry different consequences. Match the level of verification to the decision.
For voice impersonation, move to another channel
The US Federal Trade Commission’s guidance describes how cloned voices can support family-emergency scams. Its central advice is to verify through a number you already know, rather than trusting a familiar-sounding voice. That breaks dependence on the incoming call.
Consider an illustrative workplace request: a caller sounding like a manager asks for an urgent payment. The immediate question is whether the payment is authorised, not whether you can hear an audio artefact. Follow the organisation’s existing approval route and contact the manager through an established channel. Urgency is a reason to use the process, not bypass it.
What Content Credentials can add
C2PA Content Credentials can record signed information about a file’s origin and editing history. They are evidence about provenance, not an automatic judgment that the depicted claim is accurate. Missing credentials are also not proof of fabrication. Read the information they supply alongside other evidence.
The human cost is the extra work of verification
Our analysis is that synthetic media shifts part of the burden from recognising a familiar face to maintaining reliable verification habits. Families need known contact routes. Employers need clear approval procedures. Publishers need accessible source material and visible corrections. These habits help with ordinary misleading content as well as deepfakes.
You do not need to resolve every doubtful video before moving on. When evidence remains incomplete, “unverified” is a useful conclusion. It preserves the difference between not knowing and proving that something is false.