Deepfakes: the defence is no longer 'seeing is believing' but verifying through another channel, because visual detectors have an expiry date

By Diario Mendoza · November 7, 2026.

Deepfakes (fake videos, images or audio generated with AI to make it look as if someone said or did something that never happened) have become an established scam tool, according to an explainer guide from Diario Mendoza. Typical cases include celebrities who "recommend" investments or cryptocurrencies they never endorsed, voice notes from a supposed relative urgently asking for money, and video calls with fake bosses.

The most striking case the outlet cites dates from 2024: an employee of a multinational in Hong Kong transferred more than $25 million after a video call in which every other participant, including the supposed chief financial officer, was a deepfake.

The term combines "deep learning" (an AI technique) and "fake," and became popular in 2017. According to the article, what has changed since then is access: creating them used to require technical skills and a lot of time; today, tools available to anyone can generate them in minutes.

The guide suggests warning signs. In video: unnatural blinking or gaze, blurry or "flickering" face edges, out-of-sync lips, lighting and shadows that don't match, and distorted hands and teeth. In audio: a flat or robotic voice, odd pauses and breathing, and sound that is too clean. And in context: urgency, a request for money or personal data, or someone asking for something out of character.

The protective measures it recommends are simple: verify through another channel (hang up and call the usual number), agree on a family code word, distrust "risk-free" investments, look for the original source, use reverse image search and reduce how much of your own material you post on social media. It adds that some tech companies offer detectors that identify content created with their own tools, though none is foolproof.

**Our take.** The most valuable part of this guide is not the list of visual flaws but what surrounds it. Technical tells have an expiration date: glitches in hands, teeth or blinking are artifacts of the current generation, and each new generation of models fixes them. The article itself admits as much with an "although the technology is improving, still…". Advice that relies on the forger being clumsy ages badly.

What doesn't expire is the scam pattern: urgency, money and pressure to bypass normal procedures. That is the robust signal, because it depends not on video quality but on social engineering (manipulating the person, not the machine). That is why the Hong Kong case is instructive: what failed was not the ability to spot pixels but the process. A payment order of that size should not hinge on whether a face on a screen looks real.

More broadly, in business this points to controls that don't depend on perception: dual authorization for transfers, confirmation through an independent channel and amount limits. At home, the code word is a homemade and surprisingly effective equivalent. These are verification measures, not paranoid distrust.

There is an underlying cost worth stating without drama. In the short term we face an uncomfortable transition in which voice and face stop being proof of identity, and the most exposed will be those with the fewest resources to verify: older people and small businesses without formal procedures. It is a real problem of inequality in security.

In the long term, the same technology has the potential to lead us to a better society, and sustaining that trust will require authenticity infrastructure: cryptographic signatures of origin, verified channels and social habits of checking. It is not a dark future; it is an adaptation, like the one we made with spam or phishing. The difference is that this time the deception arrives with the face and voice of someone we love. Learning to pause before acting is the best defense today.