E-Paper
Sign In
Spot deepfakes: How to avoid AI scams and protect your money
AI-generated content is everywhere. Our guide reveals the red flags to spot fake videos and audio, helping you protect your finances and information
- PUBLISHED: Tue 5 May 2026, 8:00 AM
We live in an era where a voice message from your boss or a video call with a colleague could be entirely fabricated; the ability to distinguish real from fake has become a critical skill. With deepfake content on social media growing by 550 per cent between 2019 and 2023, knowing what to look for could be the difference between catching a scam and losing millions.
Maher Yamout, lead security researcher at Kaspersky, has outlined several tell-tale signs that even non-experts can use to identify manipulated content. These indicators work across text, photos, videos and audio recordings — the four formats most commonly weaponised by scammers.
Stay up to date with the latest news. Follow KT on WhatsApp Channels.
Recommended For You
"During periods of heightened tensions or major news events, misleading or manipulated content can circulate even more rapidly," Yamout explains. "It is becoming increasingly important for people to pause before sharing content and assess whether what they are reading, seeing or hearing is genuine."
Visual red flags that expose fake videos
Watch the lighting: Deepfakes often ignore basic physics. Shadows on a person's face may not match shadows in the background. A subject might appear to be standing outdoors or near a window, yet their face is lit as though they're in a studio. Similar inconsistencies appear in video calls when artificial backgrounds look unnatural or lighting doesn't match the environment.
Check the hairline: This is where deepfake software frequently fails. Look for blurring, flickering or unnatural colour transitions along the hairline. These glitches are caused by flaws in the software used to superimpose a cloned face onto original footage.
Count the blinks: A person typically blinks 10 to 20 times per minute. Some deepfakes blink too often, while others barely blink at all. Eyelid movements may appear abrupt, and blinking can be out of sync — one eye closing while the other stays open. "Glassy" or "dead-eye" stares are also common. In some cases, a pupil may twitch randomly due to a neural network glitch.
For static images, zoom in on the eyes and compare reflections in the irises. In real photos, these reflections tend to be identical. In deepfakes, they often differ.
Look for lip-sync issues: Even high-quality deepfakes struggle to synchronise speech with lip movements. A delay of just a hundred milliseconds can be noticeable to the human eye. Watch for unnatural lip shapes when the person pronounces sounds like "M," "F," or "T."
Examine the background: In AI-generated videos, backgrounds may look unrealistic — overly blurred, static even when the camera moves, or with elements that don't interact naturally with the person in the foreground.
Watch facial expressions: Deepfakes still struggle to convincingly reproduce emotion. Facial expressions may not change naturally during conversation. Smiles may appear frozen. The fine wrinkles and folds that normally appear when people express emotions may be missing entirely.
Spotting impossible architecture
AI-generated imagery often contains structural inconsistencies that would be impossible in the real world. While an image may appear convincing at first glance, closer inspection reveals physical impossibilities.
A staircase might lead nowhere, abruptly ending against a wall or floating mid-air. Doors and windows may be misaligned or oddly spaced, creating architectural layouts that would never exist in reality. These "impossible architecture" details are subtle but revealing.
"Deepfakes or manipulated visuals may also show mismatches between people's behaviour and the surrounding situation," Yamout notes. People may appear relaxed or smiling during what should be a serious or emergency situation. Their reactions may not match what is happening around them.
Audio deepfakes: Listen for what's missing
Flat, robotic tone: If a voice sounds unusually flat, lacking natural variation in intonation, or carries a slightly electronic quality, there's a high chance it's AI-generated. Real speech contains subtle tonal variations and natural imperfections.
No breathing sounds: Humans naturally take micropauses and breathe between phrases, especially in longer sentences. Small coughs, sniffs and breaths are part of normal speech. Synthetic voices often lack these nuances or place them unnaturally.
Abrupt cuts and unnatural pacing: The voice may abruptly cut off. Words may sound unnaturally "glued" together. The stress and intonation may not match how a familiar person usually speaks.
The uncanny valley of excessive detail
Some AI platforms generate photos and videos with excessive detail — more than what would be captured even with the highest-quality cameras. If an image looks hyper-realistic to the point where it seems unnaturally sharp or detailed beyond what the naked eye would normally see, it may be artificially generated.
Context is king
Beyond technical tells, context remains one of the strongest indicators. If a message creates urgency, bypasses normal approval processes, or asks you to do something slightly outside established patterns, pause. Verify through a second, trusted channel before acting.
"A few extra seconds of scrutiny can make the difference between stopping misinformation and spreading it further," Yamout says. "As deepfake technology becomes more sophisticated, critical thinking and careful inspection of online content are becoming essential digital skills."
The technology will only improve and the tells will become subtler. But the principle remains: when something feels off, it probably is. Trust your instincts, verify independently, and never let urgency override caution.
