Deepfake Awareness
how to recognize deepfakes and AI face swaps: the visual tells, verification habits, and what to do when you suspect media is synthetic.
The short answer
Spot deepfakes by stable tells — edge artifacts around the face, teeth and ear inconsistencies, lighting mismatches, unnatural blinking — plus context checks (source credibility, cross-verification, provenance labels). But tells fade as models improve: the durable skill is verification habits, not pixel-hunting. When in doubt, assume nothing and check provenance.
The visual tells (while they last)
Classic generative tells: boundary artifacts — blurring, shimmer, or color mismatch at the face boundary, especially in profile or during motion; teeth and eyes — individual teeth blending together, inconsistent reflections in the eyes; lighting and shadow physics — face lighting that disagrees with scene lighting (the single hardest thing for compositors, the most reliable tell for stills); temporal tells in video — frame-to-frame identity drift, irregular blinking, mouth-to-audio sync errors. Face swaps specifically tend to fail on occlusions: hair, glasses, hands crossing the face.
The verification habits that do not expire
The tells above decay as models improve; process does not. 1) Source check: who posted it, and does the account have a track record? 2) Cross-check: is the event/image corroborated by independent outlets or the depicted person's own channels? 3) Reverse image search: many "new" fakes recycle older real photos as bases. 4) Provenance labels: C2PA-style content credentials and platform AI labels are becoming reliable signals — their presence is meaningful, and their absence on an extraordinary claim is a flag. 5) The extraordinariness test: the more out-of-character the content, the higher the verification bar.
What to do with a suspected fake
Do not amplify: no shares, no quote-posts, no engagement — attention is the payload of most deception. If it depicts you: document (screenshots with URLs and timestamps), report to the platform (synthetic-media and impersonation policies), and preserve evidence for takedown requests (the TAKE IT DOWN framework covers NCII; see the laws page). If it depicts someone else and seems harmful: report, and optionally notify the person. The ecosystem-level defense is collective restraint — the share is the vulnerability.
FAQ
How can I tell if a photo or video is a deepfake?
Visual tells — face-boundary artifacts, lighting mismatch, teeth/eye inconsistencies, occlusion failures — plus, more durably, process: source check, cross-verification, reverse image search, and provenance labels.
Are deepfake detection tells still reliable?
Less every year. Pixel-level tells decay as models improve; the durable skills are verification habits (provenance, corroboration, source credibility), not spotting artifacts.
What should I do if someone makes a fake of me?
Document everything (URLs, timestamps, screenshots), report to the platform under synthetic-media and impersonation policies, and preserve evidence for takedown — the 2025 federal TAKE IT DOWN Act covers non-consensual intimate imagery.