How to Spot AI-Generated Media Without Overtrusting Your Eyes
Verify the person through a known channel
The FBI recommends independently confirming identity rather than trusting a new number, voice message or video.
Treat artifacts as clues
Odd hands, shadows or mouth movement can raise suspicion but are not reliable enough to settle authenticity.
Check whether the account is expected
A real public figure appearing on an unknown or recently created account needs extra verification.
Search for the same statement elsewhere
Context changes what a headline means. Major public announcements usually leave multiple independent traces.
Inspect urgency and requested action
For day-to-day verification, the practical consequence is straightforward: Deepfake scams often push money, credentials or a move to a new messaging platform.
Escalate before acting
Use previously verified contact details when money, reputation or safety is involved.
Field evidence: the 2025 AI impersonation campaign
In 2025 the FBI described a real campaign in which malicious actors used text messages and AI-generated voice messages while impersonating senior U.S. officials. The Bureau’s practical advice did not depend on perfect deepfake detection: independently verify the sender through known contact details, inspect domains and account information, and avoid moving to attacker-controlled channels. This field example is useful because it shifts the decision from “does this voice sound fake?” to “can I confirm the identity through a separate trusted path?”
High-stakes synthetic-media claims deserve a higher threshold. For accusations, emergencies or market-moving material, wait for authenticated originals or multiple independent observations instead of sharing a clip because it merely looks real.
File metadata can reveal software, timestamps or editing history, but platforms often strip it. Content Credentials and other provenance systems can provide useful signed information when present; absence of those signals is not proof of manipulation. Check whether the account posting the media has a credible chain to the event and whether other observers captured consistent details.
Use metadata and context when available
For audio and video, compare the clip with longer recordings from the same event. Lip-sync problems, abrupt cadence and unusual background noise can matter, but transcoding and low-quality reposts complicate interpretation. Search for another angle or an official full recording before relying on a detector score.
Extra fingers, warped lettering and inconsistent reflections can still reveal synthetic images, but generation systems improve and ordinary editing can create similar artifacts. Treat visual oddities as prompts for provenance work rather than as a verdict. The stronger question is whether the file can be traced to a known photographer, camera sequence, original upload or independently documented event.
Do not make artifact hunting your only test
When a real person is depicted, compare speech with verified recordings from the same period and setting. Avoid biometric certainty from casual visual comparison; use it only to identify questions worth checking. Authentication should come from provenance, corroboration and source access.
A convincing synthetic clip can still fail a basic external check. Search for other photographs, local reporting, official schedules, venue footage or eyewitness accounts that should exist if the event happened as shown. Consistency across independent records is harder to fake than a single polished asset.