MMOOZZGG.com

AI literacy

How to spot a deepfake and not fall for fraudsters

7 min read6 sources

To spot a deepfake, looking closely at a face on a screen is not enough: the fake is given away by an urgent request to send money. According to the International AI Safety Report 2026, in one study people took a cloned voice for a real person in 80% of cases. So the main rule of the Bank of Russia is simple: do not hurry, and call back yourself.

On this page · 6 sections

What a deepfake is

A deepfake is a video, voice or photo of a person made by a neural network to pass a fake off as real. More about the term and its history is in the glossary entry on deepfakes.

The Bank of Russia describes the scheme this way: fraudsters make a realistic video of a person and send it to friends and family through messengers or social networks. The voice in such a clip is sometimes hard to tell from the real one.

The material for the copy mostly comes from hacked accounts: photos, videos and voice recordings. The more of your clips and voice messages are public, the easier it is to make a fake.

A separate kind of fake is voice cloning: a program learns to speak in the voice of a specific person. In 2023, researchers behind the VALL-E model showed speech synthesis from a recording of only 3 seconds.

Signs of a fake video and voice

The Bank of Russia names three signs of a deepfake: unnatural, flat speech; defects in the sound and the video; and facial expressions the person does not normally have.

Speech. The voice sounds even, without live intonation, as if the person reads from a sheet of paper. People do not talk like that when something bad has happened.

Sound and picture. Breaks, noise, jerks, parts of the face that blur or freeze. One such defect can come from a bad connection, but together with a request for money it is a reason to stop.

Facial expressions. The person smiles, frowns or moves their lips in an unusual way. You know your family and colleagues, and an unfamiliar expression should put you on guard.

The signs help, but they prove nothing. The International AI Safety Report 2026 says that people are bad at spotting deepfakes and often guess no better than chance. If there are no signs, it does not mean you see a real person.

Bad: “The video looks like Mom, the voice is hers, so it is her.”

Good: “The video looks like Mom, but I will hang up and call her back myself.”

Typical fraud schemes

A call or a message “from a relative”. According to the Bank of Russia, a person you know tells about a problem in a short clip: an illness, a car accident, a lost job. Then they ask you to send money to a certain account.

A call “from the manager”. The Bank of Russia warns that fraudsters make deepfakes of employers. The International AI Safety Report 2026 cites cases where company executives approved transfers of millions to fraudsters.

A call “from the state”. The same tricks are used with images of government officials: a person in uniform or with an ID card on video is not yet a real official.

A video of a famous person. Fraudsters take the image of a celebrity from the field where the victim works. The message sounds like advice from an authority in your profession.

A fake page. Sometimes, instead of hacking an account, fraudsters create a new page with the person’s name and photo.

All of these are types of AI fraud, and they share one pattern: urgency, money and a channel chosen by the caller, not by you.

Rules of protection

Rule 1. Do not hurry to send money. The Bank of Russia says it directly: the fraudster needs urgency, and you always have time to check.

Rule 2. Call back yourself. Hang up and call the person in whose name the message came, on a number you already have. The US Federal Trade Commission (FTC) gives the same advice: if you cannot reach them, check with other relatives or friends.

Bad: “The number is unknown, but Mom explained that she lost her phone, so I will send the money.”

Good: “I call Mom on her old number, cannot reach her and call Dad.”

Rule 3. Ask a personal question. If you cannot call, ask in the chat something only the person you know can answer. This is exactly what the Bank of Russia advises.

Bad: “What is your name?” The answer is in any profile.

Good: “What was the name of our dog at the summer house?”

Rule 4. Remember that the account of someone you know may have been hacked. The message came from their page, but someone else may be writing.

Rule 5. Fewer public voice recordings. The FTC warns that a fraudster needs only a short clip from recordings a person posted online.

Other risks of neural networks for an ordinary person are on the page “Is AI dangerous”.

If you have already sent money

Act at once. On its page for people whose card was charged illegally, the Bank of Russia advises this order.

1. Call your bank as soon as possible on the number on the back of the card, report the fraudulent transaction and block the card. You can also block it in the bank’s app.

2. Go to a bank branch, ask for an account statement and write a statement disputing the transaction. Keep your copy of the statement with the bank’s stamp of receipt.

3. Report the theft to law enforcement, that is, to the police.

The bank will review the statement within 30 days, or within 60 days if the transaction was international.

Since 25 July 2024, a Russian bank must hold a transfer for two days if the recipient is in the Bank of Russia’s database of fraudulent transactions. If the bank did not do this, it will return the money within 30 calendar days.

For companies: the two-channel rule

The “call from the director” scheme hits accounting and finance. The AI safety report describes cases where fraudsters with a fake voice and face of an executive took millions out of companies.

The two-channel rule: a request for money or access that came through one channel is confirmed through another. If a video came in a messenger, the accountant calls the director on the work number from the company directory.

Bad: “The director asked me in a video call to pay an invoice urgently, so I pay.”

Good: “I call the director back on the number from the directory and pay only after confirmation.”

Write the rule into the invoice payment procedure and name the amount above which nobody pays without a second confirmation. Then the employee does not need to argue with the “director”: they just follow the rule.

How to train staff in these rules is covered in the AI security course.

Questions and answers

How do you check that the caller is your real manager?

Hang up and call the person back on a number you already have, for example from the company directory. If you cannot reach them, ask a question in the chat that only they can answer. The Bank of Russia, the country’s central bank, advises first calling the person in whose name the message came and checking the information again.

Where do you report a deepfake?

If you have sent money, call your bank at once on the number on the back of your card, then report the theft to the police. This is the Bank of Russia’s advice. If the fake came in the name of someone you know, call that person: their account may have been hacked.

Can a voice be faked from a short recording?

Yes. In 2023, researchers behind the VALL-E model showed speech synthesis from a 3-second recording. The FTC warns that fraudsters need only a short clip from recordings posted online. So a familiar voice on the phone proves nothing by itself.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Bank of Russia, 16 September 2024 (cbr.ru)
  2. Bank of Russia, 2026 (cbr.ru)
  3. Bank of Russia, 24 July 2024 (cbr.ru)
  4. International AI Safety Report, 3 February 2026 (internationalaisafetyreport.org)
  5. FTC, 20 March 2023 (consumer.ftc.gov)
  6. arXiv, Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers, 5 January 2023 (arxiv.org)

Found a mistake? Tell us

Publications

Mike Mozg

Anthropic IPO raises $81.34 billion and makes its compute costs public

The Anthropic IPO raised $81.34 billion, and trading on Nasdaq began on 19 November 2026, the Russian business daily RBC reported. That beats SpaceX's base IPO, but it becomes a record only once the agreed $26 billion private placement closes. I don't read this as market news: Anthropic's public filing shows what compute costs a frontier AI lab. In my view, it shows where prices, limits and access

News6 min
Contact us