In February 2024, a finance employee at the global engineering firm Arup received what appeared to be a video conference call from his company’s CFO and several senior colleagues. The instructions were clear and the faces were familiar. He transferred $25 million to the accounts provided. None of the people on that call were real. Every face, every voice — including the CFO’s — was an AI-generated deepfake.
This was not an isolated incident. In 2025, deepfake-related fraud losses in the United States alone reached $1.1 billion, tripling from $360 million in 2024. Deepfake fraud attempts increased by 2,137% over the three years between 2022 and 2025. A deepfake attempt now occurs somewhere in the world every five minutes. According to a 2024 McAfee study, one in four adults has experienced an AI voice scam, and one in ten has been personally targeted by one.
The technology behind these attacks has crossed a threshold that most people have not yet recognized: it is no longer expensive, rare, or technically complex. It is cheap, accessible, and spreading fast.
What AI Voice Cloning and Deepfakes Actually Are
AI voice cloning is the process of creating a synthetic version of a real person’s voice using artificial intelligence. The clone can say anything the attacker types, in the victim’s own voice, with their cadence, accent, and emotional tone. According to security researchers, as little as three seconds of audio — captured from a social media video, a podcast appearance, a voicemail, or a YouTube recording — is enough to create a voice clone with an 85% match to the original speaker.
Deepfakes extend this capability to video. AI tools can now generate real-time video of a person saying and doing things they never said or did, with sufficient realism to deceive someone who knows them. A 2025 iProov study found that only 0.1% of participants correctly identified all fake and real media presented to them. Human detection accuracy for high-quality deepfakes can drop as low as 24.5% — meaning most people cannot reliably tell the difference.
These are not futuristic threats. The tools to create convincing voice clones and deepfake videos are freely or cheaply available online right now, with no specialized technical knowledge required. Consumer Reports tested six major AI voice cloning tools in 2024 and found that four of them lacked any meaningful safeguards against misuse.
How These Scams Actually Work
Understanding the mechanics of deepfake scams is the most effective defense, because the attacks follow predictable patterns that become recognizable once you know what to look for.
The family emergency scam is the most emotionally devastating form. An attacker finds a few seconds of audio from someone’s social media — a child, a grandchild, a sibling — and creates a voice clone. They then call a family member, using the cloned voice to claim they are in trouble: arrested, injured, stranded, in an accident. They beg for money to be transferred immediately and plead not to tell other family members. The emotional urgency of hearing a loved one’s voice in distress overrides the skepticism that would normally catch the deception. According to FTC data, imposter scams generated over $2.9 billion in reported consumer losses in 2024, with family emergency variations among the most common.
The executive impersonation scam targets employees with financial authority. A cloned voice — or, increasingly, a deepfake video call — impersonates a CEO, CFO, or senior manager instructing a finance employee to make an urgent wire transfer, often with explanations like a confidential acquisition, a regulatory requirement, or an emergency payment. The employee complies because the voice or face is genuinely convincing and the authority figure is real. This category of attack now costs organizations an average of $500,000 per incident, with banks reporting that over 10% of financial institutions have suffered deepfake vishing losses exceeding $1 million.
The celebrity endorsement scam floods social media with deepfake videos of well-known figures — tech executives, athletes, entertainers — appearing to endorse investment schemes, cryptocurrency giveaways, or products. In 2025, multiple deepfake videos of prominent figures circulated across major platforms promoting fraudulent cryptocurrency schemes before takedowns occurred. Victims sent funds believing they were responding to a legitimate promotion, only to discover the videos were fabricated.
The romance scam evolution involves AI chatbots sustaining months-long emotional relationships across messaging platforms, now augmented by deepfake video calls that “prove” the person’s identity. Romance scam losses topped $1.3 billion in combined 2024 and 2025 data from the FTC. Victims report participating in video calls with people who smiled, nodded, and responded in real time — none of whom were real.
Why These Attacks Are So Effective
The reason AI voice and deepfake scams succeed at such high rates comes down to a fundamental aspect of human psychology: we trust what we recognize. A voice that sounds like your daughter activates the same trust response as your actual daughter’s voice. A face that looks like your CEO generates the same compliance behavior as seeing your actual CEO. The emotional familiarity created by a realistic clone essentially bypasses the skepticism that would otherwise protect you.
This effect is amplified by the conditions under which these attacks are delivered: urgency, emotional pressure, a sense of secrecy, and the disruption of normal verification habits. When someone who sounds like your grandchild is crying and asking you not to call other family members while you arrange help, the natural impulse is to act rather than verify.
The 70% of people who told McAfee researchers they are not confident they can distinguish a real voice from a cloned one are being realistic, not overly cautious. The technology is genuinely difficult to detect by ear alone, particularly in emotionally charged situations.
How to Protect Yourself: Practical Defenses That Work
The defenses against AI voice and deepfake scams are not primarily technological — they are procedural. The most effective protection comes from habits that interrupt the attack’s reliance on urgency and emotional pressure.
Establish a family code word. This is one of the most practical and widely recommended defenses security experts offer. Choose a word that family members will use to verify identity in any emergency situation where money or sensitive information is being requested. The word should be something not inferable from social media and changed periodically. If someone claiming to be a family member cannot provide the code word, the call is suspicious regardless of how convincing the voice sounds.
Verify all urgent financial requests through a separate channel. If you receive a call, voicemail, or video message from anyone — including someone whose voice you recognize — requesting a wire transfer, gift card purchase, or any financial action, hang up and call that person back on a number you independently know. Do not use a callback number provided in the original message. This single habit would have prevented the $25 million Arup loss: the employee was instructed not to verify, which itself should have been a warning sign.
Be extremely skeptical of any urgent request that includes instructions not to tell others. Legitimate emergencies do not require secrecy. A grandchild genuinely in trouble would not instruct you to keep it from their parents. A CFO with a legitimate urgent payment would not instruct an employee to bypass normal authorization processes. Secrecy is a consistent feature of scam structure, not of real emergencies.
Reduce your voice’s availability online. AI voice cloning requires source audio. Extensive voice samples in public social media videos, podcasts, or online recordings provide richer material for cloning. This does not mean avoiding all video or audio content, but it is worth being aware that public audio is usable material for this attack. Private account settings limit who can access your recordings.
Be alert to specific technical tells in suspicious calls. Current AI voice cloning, while impressive, sometimes produces audio artifacts: slightly unnatural pacing at certain points, occasional pitch inconsistencies, breathing patterns that sound wrong, or responses that are slightly delayed. These tells are not always present and are becoming less common as the technology improves, but they remain worth listening for when something feels slightly off.
What to Do If You Suspect You Have Been Targeted
If you receive a suspicious call using what sounds like a family member’s voice: stay calm, do not send any money or provide any information, and immediately call the person supposedly on the line using their real number. If you cannot reach them, contact another family member who would know their actual situation.
If you have already transferred money: contact your bank immediately — some transfers can be reversed if reported within hours. Report the incident to the FTC at reportfraud.ftc.gov, and to the FBI’s Internet Crime Complaint Center at ic3.gov. File a police report. Do not feel shame about having been deceived — these attacks are specifically engineered to overcome normal skepticism, and they work on intelligent, careful people.
Frequently Asked Questions
Q: How can I tell if a voice I am hearing on a call is AI-generated?
A: There is no reliable real-time test you can perform during a call. The most practical defense is not attempting to detect the fake but refusing to act on any urgent financial or sensitive request without independent verification through a channel you control. Hang up and call back using a number you know independently.
Q: Can deepfake detection software help protect ordinary people?
A: Detection tools exist and are improving, but they are not reliably accurate enough for consumer use in real-time situations. They work better on recorded content than live calls. The procedural defenses — code words, independent verification, skepticism of urgency — are more consistently effective than relying on detection technology.
Q: Is it safe to post videos of myself or my family on social media?
A: Short clips for typical social sharing present limited risk for targeted attacks against ordinary individuals. The risk increases significantly for people in public-facing roles or with large audiences, as more source material improves clone quality. For children, limiting public video content is a reasonable precaution since their voices are particularly likely to be used in family emergency scams targeting grandparents.
Q: What should I do if a deepfake video of me is being used in a scam?
A: Report the video to the platform hosting it — most major platforms have specific policies and reporting pathways for synthetic media fraud. Report the scam to the FTC and FBI IC3. Contact an attorney if significant financial or reputational damage has occurred. Document everything, including screenshots and URLs, before content is removed.
