Spotting Deepfakes: How to Tell When a Photo, Voice or Video Is AI-Generated

Two years ago, deepfakes mostly looked plasticky and obviously fake. Today the better ones can fool a casual viewer. Voice clones can replicate a familiar voice with under a minute of source audio. AI-generated images of people are routinely indistinguishable from photos at thumbnail size.

We can't go back to a world where every video is real. But we can build sharper instincts.

What to look for in images

Hands and ears are still tricky for many AI image models. Count the fingers. Look at how the rings sit on them. Look at the shape of the ear, the symmetry, the small folds.

Background details: signs with text often have garbled letters. Repeating patterns sometimes glitch. Reflections in windows or mirrors don't always match what's in front of them.

Skin texture: AI-generated faces often have slightly too-smooth skin, particularly on cheeks and foreheads. Real skin has tiny imperfections.

What to look for in videos

Mouths and eyes are the hardest. Watch the lip-sync carefully — does it match the consonants exactly? Watch the eyes — do they blink at natural intervals, do they track movement smoothly, do they reflect light realistically?

Edges of the face: where the face meets the hair, where the chin meets the neck. Deepfakes often have a faint blur or a slight wobble at these boundaries, especially when the head turns sharply.

Audio quality vs video quality: if the video is high resolution but the voice sounds like a phone call, that's a hint that the audio was added separately.

What to listen for in voice clones

Voice clones often have unnatural pacing — words come slightly too fast or too slow, with breathing patterns that don't match natural speech. Background noise tends to be too clean or weirdly inconsistent.

If you receive a voice message from a family member asking for money urgently, never act on the audio alone. Call them back on their normal number. Voice cloning scams targeting parents and grandparents are a growing problem.

Context is the most reliable filter

Even when the content itself is convincing, context betrays it. Did this video appear from an unknown account with no history? Is the claim emotionally extreme — designed to make you angry, scared, or shocked enough to share before thinking? Is the source a screenshot of a screenshot of a screenshot, with no original linked?

Real news, real announcements, real official statements come from real sources you can verify. AI fakes thrive in the gap where people share before they verify.

What to do if you're not sure

Reverse-image search the photo. Search for the same claim in major news outlets. Wait an hour and see if anyone else has pointed out the manipulation. The internet is faster than ever at debunking fakes — sometimes within minutes.

If a video targets you personally — claiming to be a family member or boss — verify on a different channel. Call. Don't reply on the channel where the message came from.

Healthy skepticism, not paranoia. The goal is to slow down for half a second before forwarding or believing. That tiny pause is the most powerful defence we have against the new wave of synthetic content.