How to spot a deepfake: photos, video and voice

The honest answer: by eye, it is now very hard

Recent research is consistent:

• in an iProov study of 2,000 people in the UK and US (February 2025), only 0.1% correctly told apart all the real and fake content, and participants were 36% less likely to spot a fake video than a fake image, while over 60% felt confident in their skills;
• Ofcom's June 2024 polling found that fewer than 1 in 10 (9%) UK adults felt confident they could identify a deepfake;
• for audio, a University College London study (2023) found listeners caught synthetic speech only 73% of the time, and training only helped slightly.

The signs below are still useful, especially against sloppy fakes. But the absence of signs proves nothing: a flawless-looking clip can still be fake.

Spotting a fake photo

• Hands and fingers: wrong number of fingers, odd joints, hands melting into objects.
• Fine details: teeth that are too uniform, mismatched earrings, asymmetric glasses, hair blending into the background.
• Text: signs, labels or screens with warped or meaningless lettering.
• Light and shadows: shadows pointing different ways, reflections in the eyes that do not match.
• Skin: too smooth, waxy or plastic-looking, with no pores or blemishes.
• Background: bending lines, half-formed objects, distorted people behind the subject.
• Body joins (in fake intimate images): different skin tone or grain between face and body, neck or shoulders that do not line up.

Spotting a fake video

• Face edges: a blurry or "shimmering" border around the chin, ears or hairline, especially when the head turns.
• Blinking and gaze: too little or mechanical blinking, a slightly fixed stare.
• Mouth: lips slightly out of sync, a blurry inside of the mouth, teeth that change from frame to frame.
• Objects in front of the face: a hand, glass or microphone crossing the face often makes the fake glitch.
• Profile view: many fakes hold up face-on but distort side-on.
• On a live video call: ask the person to turn their head, wave a hand in front of their face or hold up a specific object. Newer tools can sometimes handle this, but it still trips up many live fakes.

Spotting a cloned voice

• Rhythm: speech that is too even, no hesitation, no breathing, flat intonation.
• Emotion: panic or urgency that sounds "performed".
• Sound: no background noise, or noise that is oddly uniform; small cuts between sentences.
• The script: urgency, a request for money or gift cards, a request for secrecy, a new phone number. This is the most reliable sign, far more than the audio itself.

What works: hang up and call the person back on the number you already have. Agree on a family code word to ask for if a call seems off. What to do if your voice was cloned.

Checks that actually work

1. Trace the source. Who posted it first? A new account with no history, or one that only posts shocking content, is a warning sign.
2. Run a reverse image search (Google Lens, Bing, TinEye) on a screenshot: you may find the original, unaltered photo or video.
3. See whether reliable outlets report it. A shocking statement by a public figure that nobody credible has picked up is suspicious.
4. Look for provenance data. Some tools attach "Content Credentials" (the C2PA standard) or an invisible watermark. Since November 2025, Google's Gemini app can tell you whether an image carries Google's SynthID watermark. Big caveat: these tools only see content from services that use them, so "nothing found" does not mean an image is real.
5. Be wary of online "deepfake detectors". They give a probability, not proof, and they get it wrong in both directions.
6. Contact the person through a separate channel before you act (send money, share, reply).

What if the deepfake is of you?

You do not need to prove technically that something is fake to get it taken down: if you never filmed that scene or said those words, lack of consent is enough to report it. Keep evidence (the page address and a screenshot), never pay a blackmailer, and report with the most precise reason. In the US, the TAKE IT DOWN Act requires platforms to remove non-consensual intimate images, including AI fakes, within 48 hours of a valid request. The first 7 steps.

If the person shown is under 18: use Take It Down (NCMEC) in the US or Report Remove (Childline and the IWF) in the UK. Our page for under-18s.