Replaces identity, keeps performance
Maps one person's face onto another person's real performance — the expressions and motion are the original actor's; only the identity changes.
Terminology
Deepfake vs. AI-generated image: not the same problem.
The two get used interchangeably in headlines, but they're different technically and different to verify. Knowing which one you're looking at changes what evidence is even available.
Definitions
What each term actually refers to.
| Deepfake | AI-generated image | |
|---|---|---|
| Starting point | A real photo or video of a real person | Text prompt or reference image, no required real starting point |
| Typical technique | Face-swap or face-reenactment models trained on the target's real footage | Diffusion or GAN models generating pixels from scratch |
| An "original" exists? | Usually yes — the underlying real video/photo the face was swapped onto | No — there's no real photo being altered |
| Best verification approach | Compare against known authentic footage of the same person; check for blending seams, blink rate, lighting mismatch at the face boundary | Metadata, generator signatures, reverse image search, artifact review across the whole frame |
| Legal/harm profile | Directly impersonates a specific real identifiable person | May depict no real person, or a fabricated likeness — different risk profile |
Why it matters for detection
You're checking a different thing in each case.
For a suspected deepfake, the useful comparison is the target person's other authentic footage — voice, mannerisms, known appearances — checked against the suspect clip for boundary artifacts around the face, inconsistent blinking, and audio-video sync issues. For a suspected fully-AI image, there's no "real" version to compare against, so the check shifts entirely to internal evidence: embedded metadata, frequency-domain artifacts, and physical-detail review (hands, text, reflections). Running the wrong checklist against the wrong category wastes the time you have to verify something that matters.
FAQ
Deepfake questions.
Is every deepfake a video?
No — face-swap techniques apply to still images too, though "deepfake" is used most often for video and audio impersonation of a specific real person.
Can this site's checker detect deepfakes?
The local checker is built for still-image generator metadata and compression analysis. Deepfake video detection needs frame-by-frame and temporal analysis this tool doesn't attempt.
Is an AI-generated image of a real person a deepfake?
Terminology varies, but most usage reserves "deepfake" for techniques that alter real source footage of that person, rather than a prompt-based generation of their likeness with no real source frame.
Worked example
What to look at, frame by frame, on a suspected face-swap.
A short clip shows a public figure saying something out of character. Working through it in order: check the face boundary — jawline and hairline edges where a swap meets the real footage often show a faint blur, color mismatch, or a seam that flickers between frames. Check blink rate and eye reflections — early face-swap models notoriously under-blink or show inconsistent catch-lights in the eyes, though this has improved in newer tools. Check audio-video sync at plosive sounds ("p", "b") — lip shapes for these sounds are distinctive and hard to fake precisely. Finally, check whether the claimed footage source (a specific broadcast, a specific event) actually exists — a fabricated setting is often easier to disprove than the video itself.
Technique names
Face-swap vs. face-reenactment: not the same technique.
Replaces performance, keeps identity
Drives a real target person's face with someone else's expressions and speech — the identity in frame is genuinely theirs, but the performance is puppeted.
Often paired with either
Increasingly combined with audio deepfakes trained on a target's real voice samples, which is why audio-video sync review matters as much as the visual check.