How to detect AI deepfakes
Learn what AI deepfakes are and how to recognise fake videos, photos and cloned voices — including the scams where someone copies a relative's or leader's voice to trick you.
A 'deepfake' is media — a video, photo or audio clip — created or altered by artificial intelligence to show someone saying or doing something they never did. The technology has become cheap and easy, so deepfakes now appear in political rumours, fake celebrity endorsements, and scams targeting ordinary families.
You may have seen videos where a famous Nepali figure appears to promote an investment scheme, or audio where a 'relative' calls asking for money. Many of these are AI fakes. As the tools improve, the old advice of 'just look closely' is no longer enough on its own.
This guide explains the common types of deepfakes, the visual and audio clues that can still give them away, and — most importantly — the habits that protect you even when a fake looks perfect.
The main types of deepfakes
Deepfakes come in a few forms, and knowing them helps you guess what you are looking at.
- Face-swap or talking-head videos: a real person's face is made to say new words, often for fake endorsements or political smears.
- AI-generated images: completely invented 'photos' of events, people or documents that never existed.
- Voice cloning: a few seconds of someone's real voice is enough to generate new sentences in that voice — used in 'family emergency' phone scams.
- Cheap fakes: not true AI, but real clips slowed down, sped up, mislabelled or edited to change their meaning. These are still very common and very effective.
Visual clues in fake videos and photos
AI is improving fast, so no single clue is proof. But when several of these appear together, be very suspicious.
- Hands and fingers: wrong number of fingers, bent oddly, or blurry. AI still struggles with hands.
- Eyes and blinking: unnatural staring, mismatched reflections in the two eyes, or strange blinking.
- Edges and skin: a too-smooth, waxy face, or a blurry line where the face meets the hair or neck.
- Teeth, ears and jewellery: melted or asymmetric details, earrings that do not match.
- Lip-sync: the mouth movements do not quite match the words, especially on consonants.
- Background: warped text, garbled signboards, or objects that bend strangely — a common giveaway in AI images.
Clues in AI-cloned voices
Voice scams are dangerous because we instinctively trust a familiar voice. Listen for these signs, especially during an unexpected or urgent call.
- A flat, even tone with little natural emotion, or strange pauses and breathing.
- Background that is too clean and silent, or oddly looping.
- The caller avoids spontaneous back-and-forth and sticks to a script.
- Pressure and urgency: 'send money now', 'don't tell anyone' — the classic scam pattern.
Habits that beat even a perfect fake
Because detection clues will eventually disappear as AI improves, your strongest defence is behaviour, not eyesight. These habits work no matter how realistic the fake is.
- Verify through a second channel: if a 'relative' calls for money, hang up and call them back on their known number.
- Agree a family safe-word in advance, so a real relative can prove who they are during an emergency call.
- Distrust unexpected urgency about money, OTPs, or secrecy — it is the signature of a scam.
- For public videos, check whether trusted news outlets and the person's official accounts confirm it.
- Slow down. Deepfake scams rely on you reacting fast and emotionally.
Tools and limits
Some platforms are starting to label AI-generated content, and reverse image search (such as Google Lens) can reveal where a picture really came from. Online 'deepfake detectors' exist but are not reliable enough to trust on their own — they produce both false positives and false negatives.
Treat any single tool as one opinion, not a verdict. The most reliable check is still context: does the claim make sense, who is the original source, and can you confirm it through an independent, trusted channel?
Key takeaways
- ✓A deepfake is AI-made video, photo or audio showing someone saying or doing something they never did.
- ✓Look for clues in hands, eyes, edges, lip-sync and warped background text — but never rely on looks alone.
- ✓Voice cloning needs only seconds of audio; an urgent call from a 'relative' asking for money is a major red flag.
- ✓The strongest defence is verifying through a second channel and using a family safe-word.
- ✓Online deepfake-detector tools are unreliable; context and independent confirmation matter more.
How to Detect AI Deepfakes — FAQ
Can a deepfake be made from just one photo of me?+
Increasingly, yes — modern tools can animate a single photo or clone a voice from a short clip. This is a good reason to be thoughtful about the videos and audio you post publicly, and to set up a family safe-word.
Are deepfake-detector apps reliable?+
Not reliable enough to trust alone. They can flag real videos as fake and miss good fakes. Use them only as one weak signal alongside context and independent verification.
A video shows a famous person promoting an investment — is it real?+
Be extremely cautious. Fake celebrity and leader endorsements for investment or crypto schemes are a very common deepfake scam. Verify on the person's official, verified accounts before believing or acting.
What if a deepfake is made of me or my family?+
Save the evidence, report it to the platform, and inform people in your circle so they are not fooled. For serious cases such as harassment or fraud, you can report cybercrime to the Nepal Police Cyber Bureau.
Sources & data note
These guides explain widely-accepted SEO, AEO and GEO practice as documented by Google Search Central, schema.org and current industry research. Search and AI systems evolve continually — treat specific thresholds (e.g. Core Web Vitals targets) as current guidance and verify against the latest official documentation. Examples are tailored to Nepal's market.