Before you start
You need a phone and the people you would call in an emergency. The exercise at the end takes ten minutes with one of them.
What deepfakes and voice clones have already done
In May 2024 Arup confirmed it had been the victim of a fraud in Hong Kong. Police said a finance employee joined a video call with deepfake recreations of colleagues and then transferred money. The faces on that call were not real.
A cloned voice can start from a short clip. Wikipedia's article on audio deepfakes defines voice cloning as speech generated to mimic specific individuals. It reports a 2020 web app that produced high-quality speech from 15 seconds of training data, where earlier systems had needed tens of hours.
A cloned executive phones a UK energy firm
The voice of a parent company executive was imitated to order the firm's chief executive to transfer money.
A web app clones speech from a short sample
The demonstration needed far less training data than earlier systems, which took tens of hours.
Robocalls imitate a US president
Over 20,000 voters in the New Hampshire Democratic primary campaign got robocalls from an AI-imitated Joe Biden urging them not to vote. In February 2024 the FCC banned AI-faked voices in robocalls.
Arup confirms a deepfake video-call fraud
A finance employee in Hong Kong joined a call with deepfake recreations of colleagues and sent money in 15 transactions.
Source: Wikipedia: Audio deepfake (read November 6, 2026)
Five steps when a call or clip pushes you to act
- Notice the push. A call that sounds urgent or emotional and asks for money is the pattern the UN's crime office warns about in its post on X.
- Hang up. Whoever is on the line can keep you talking, so end the call. Then contact the person directly, on a number or in an app you already had. A number the caller gives you proves nothing.
- Ask for the code word or a private fact. A cloned voice copies how someone sounds. It has no way to know what only the two of you agreed. Wikipedia's deepfake article quotes one detection-company expert saying semantic passwords or secret questions can be used in important conversations.
- For a clip you did not receive in a call, find where it first appeared. Search a still frame as lesson 1 shows, and look for the same clip on the speaker's own channels or on news sites you trust.
- At work, require a second person and a second channel before money moves on a request that came by call. The Arup case shows why a convincing video call is not enough, and the same expert advises educating employees about deepfakes.
Why your eyes and ears are not the test
The 2021 MIT paper that Wikipedia's deepfake article cites found ordinary people 69 to 72% accurate on a random sample of 50 videos. The article adds that in natural viewing, people are not more likely to flag a deepfake than a real video. Cues such as blinking, lighting, lip sync and body language are real research topics, and the article says they are not always reliable.
Software does not close the gap. The Deepfake-Eval-2024 benchmark gathered about 45 hours of video, 56.5 hours of audio and 1,975 images that circulated online in 2024, from 88 websites in 52 languages. Open-source detectors lost 50% of their AUC on video and 48% on audio compared with earlier benchmarks. AUC is a score for how well a detector separates fake from real. Commercial and fine-tuned models did better and still fell short of deepfake forensic analysts.
The chart shows the best a model reached after the authors fine-tuned it on part of the data. The audio bar is the highest, and it still leaves about one answer in seven wrong. Wikipedia adds that audio research mostly targets English, and detectors trained mainly on English do worse on other languages.
Source: Hugging Face Papers: Deepfake-Eval-2024, a multi-modal in-the-wild benchmark of deepfakes circulated in 2024 (read November 6, 2026)
What the UN's crime office tells people to do
The UN Office on Drugs and Crime posted on X on 16 August 2026 that criminals use synthetic voices to impersonate loved ones in distress and demand money in fake kidnapping or emergency scams.
Its advice fits in one line, and it is the check that does not depend on how good the fake is. Steps 1 and 2 above are that line spelled out.
If you get a call that sounds urgent or emotional, verify before reacting, contact the person directly.
A screenshot of X's own embed of the post, taken on 6 November 2026, with the time in UTC. Select it to open the post on X.
How to set a family code word
A code word turns the check from judging a voice into asking for a fact. The rules below are our own, built on the semantic passwords and secret questions that Wikipedia quotes an expert recommending for the workplace. The video that follows is on the same idea. Its title calls it a safe phrase, and that title is all we rely on from it.
- Agree on it in person or on a call you started. Never by text or in a group chat.
- Pick something no one could guess from your posts. A pet's name, a school or a birthday fails that test.
- Make the caller say it first. If you say it, whoever is listening has it.
- Agree what a missing word means: you hang up and call back on a saved number.
- Change it if it was ever spoken on a call you did not start.
Common mistakes
- Calling back the number the caller gave you, or the one in the message. Use a number you already had.
- Judging a voice by how real it sounds. Realism says nothing about who is speaking.
- Trusting a video call because you can see faces. In the Arup case the faces on the call were recreations.
- Posting the code word, or saying it before the caller does.
Try it yourself
- Call one relative this week and agree a code word. Then save the numbers of the three people you would call back first, under their real names.
- Role-play it once. Ask a friend to phone you from a number you do not know with an urgent request, and notice whether you stop to verify.
- Take one video clip you have seen shared, find where it first appeared and write down the date.
Next: marks a tool can read
Every check so far depends on you. Lesson 3 covers marks that a tool can read, signed credentials and invisible watermarks, and what they can and cannot prove.
Check yourself
What did the 2021 MIT paper cited by Wikipedia find about ordinary people watching deepfake videos?
Show answer
D. They were 69 to 72% accurate on a sample of 50 videos Wikipedia reports that ordinary people were 69 to 72% accurate on a random sample of 50 videos, so about three in ten answers were wrong.
An urgent call from "your daughter" asks you for money. What is the first move?
Show answer
C. Hang up and contact her on a number you already had The UN's crime office says to verify before reacting and contact the person directly, and a number you already had keeps the caller out of it.
Why does a family code word work against a cloned voice?
Show answer
B. The clone copies how someone sounds but cannot know what you agreed in private The check asks for a fact and does not rely on judging the voice. Wikipedia quotes an expert recommending semantic passwords or secret questions for important conversations.
What did the Deepfake-Eval-2024 authors find about commercial detectors?
Show answer
A. They did better than open-source models but fell short of forensic analysts, and none reached 90% accuracy The paper says commercial models do not yet reach the accuracy of deepfake forensic analysts, and none of those evaluated had 90% accuracy or above.
Frequently asked questions
How common are voice-cloning scams?
One survey puts it at one in ten. Wikipedia reports a 2023 McAfee survey in which one in ten respondents said an AI voice cloning scam had targeted them. That is a single survey from a security vendor, so it says little about rates today.
Can a video call be faked?
Yes. In the case Arup confirmed in May 2024, a finance employee met deepfake recreations of colleagues on a video call. Wikipedia also notes that Zoom partnered with Tools for Humanity in April 2026 on a verification feature for video calls.
What is a deepfake?
Wikipedia defines deepfakes as images, videos or audio that have been edited or generated using artificial intelligence. They can show real or fictional people.
Do deepfake detectors work?
Only partly. On fakes that spread online in 2024, no commercial model the Deepfake-Eval-2024 authors evaluated reached 90% accuracy. Treat a tool's verdict as one clue, never the answer.
Does calling back always work?
It works when you dial a number you already had, because the caller is out of the loop. A number supplied by the caller hands the call back to them.
Sources
- Wikipedia: Deepfake https://en.wikipedia.org/wiki/Deepfake
- Wikipedia: Audio deepfake https://en.wikipedia.org/wiki/Audio_deepfake
- Hugging Face Papers: Deepfake-Eval-2024, a multi-modal in-the-wild benchmark of deepfakes circulated in 2024 https://huggingface.co/papers/2503.02857
- UN Office on Drugs and Crime on X, 16 August 2026 https://x.com/UNODC/status/2088913712354590785
- YouTube: WCCO CBS Minnesota, Create a safe phrase with your family to beat AI voice-cloning scams https://www.youtube.com/watch?v=1JN-QZAY7tQ