Deepfakes: The AI Videos Realistic Enough to Fool Anyone

Updated: Nov 8
A deepfake is a video, image, or audio recording wholly or partly fabricated by artificial intelligence to make someone say, do, or show something they never actually said, did, or showed. The word fuses "deep learning" and "fake," and behind that tidy definition sits something far uglier: an employee at the British engineering firm Arup wired $25 million after joining a video call where every other participant, including his own CFO, was a synthetic creation.
That case now looks almost quaint. In September 2026, fraudsters used a cloned voice and a spoofed WhatsApp message to walk off with €95 million (roughly $108 million) from Fideuram, the private banking arm of Italy's largest lender, Intesa Sanpaolo, one of the biggest AI-enabled thefts ever disclosed by a major financial institution. The real danger of a deepfake was never that it could fool a trained eye. It's that it fools a rushed accountant, a grandmother on the phone, or, per a 2025 study that tested real people against real fake content, 99.9% of everyone else. Fraud reports, court filings, and a string of recent scandals tell the actual story of where this technology stands. For the bigger picture of how generative AI creates harm well beyond video fakery, see our broader look at generative AI's invisible dangers and their victims.

In short
A deepfake is AI-generated audio, image, or video, built with GANs or diffusion models, that fabricates a real person's likeness or voice.
Deepfake-enabled fraud has already cost a single organization over $100 million, and cloning a voice now takes just three seconds of audio.
Non-consensual sexual deepfakes disproportionately target women and minors, as the Grok "Spicy mode" scandal and a wave of school-based cases have shown.
The EU AI Act and the US TAKE IT DOWN Act now criminalize deepfake abuse, but France's pioneering SREN law has a labeling loophole that can shield the people who disclose their fakes.
Detection tools remain unreliable outside the lab: accuracy drops by up to 50% on real-world content, and unaided humans barely beat a coin flip.
How a deepfake actually works
The mechanism behind the buzzword is simple to describe, even if building a good one takes real engineering. The original technique relies on generative adversarial networks, or GANs: two algorithms locked in competition, a "generator" that fabricates fake faces or voices, and a "discriminator" that tries to catch the fraud. The generator improves with every failure, until its output is convincing enough that the discriminator can no longer tell it apart from the real thing. As MIT Technology Review documented, the technique traces back to one night in 2014, when Ian Goodfellow, then a PhD student in Montreal, years before he'd go on to work at Google Brain, OpenAI, and Apple, sketched the idea out after a bar conversation.
Diffusion models have since overtaken GANs as the dominant approach for image and video generation. What matters for the average person is simpler: the raw material, a few photos, a few seconds of voice (almost always public), scraped from an Instagram story or a single phone call. (If any of this terminology is unfamiliar, our AI glossary breaks down the key terms.)
The myth we need to stop repeating: "it takes skill"
It's a line you'll find in plenty of otherwise serious guides: a convincing deepfake supposedly requires time, a powerful GPU, and mountains of training data. That was true three years ago. It isn't anymore.
In late 2023, a male student at Westfield High School in New Jersey used a free "nudify" app to strip the clothing from photos of more than 30 female classmates and circulate the results, no GPU cluster, no machine-learning background, just an app download, according to New York Times reporting. The same pattern has surfaced well outside the US, including comparable cases involving French teenagers. Then, in December 2025, a feature built directly into a mainstream chatbot, xAI's Grok, let any user "digitally undress" real women and, in documented cases, minors, with nothing more than a text prompt. (We've covered how Grok actually works if you want the mechanics.)
What both cases show is that the real barrier was never technical. It was accessibility. And by 2026, accessibility isn't a problem for anyone, malicious or not.
Danger #1: financial fraud built to trick seasoned professionals
The Arup case remains the textbook example: an employee in Hong Kong joined a video call believing he saw his London-based CFO and several colleagues. All of them were reconstructions built from public videos scraped off YouTube. Reassured by "familiar" faces, he executed 15 transfers to five accounts, totaling $25 million.
That kind of attack has since moved well beyond corporate boardrooms into something far more intimate: voice cloning. It now takes just three seconds of audio, a voicemail, a social post, a single "hello" on a spoofed call, to convincingly reproduce someone's voice. According to McAfee's global "Artificial Imposters" study of over 7,000 people across seven countries, a quarter of adults have already experienced an AI voice-cloning scam or know someone who has, and 77% of victims lost money as a result. Globally, the trajectory is staggering: deepfake-enabled identity fraud is projected to rise nearly 495% in 2026 compared to 2025, according to Shufti's Identity Fraud Index, reported by ASIS Security Management.
The September 2026 Fideuram case shows the pattern scaling further still: a fake WhatsApp message from the CEO, followed by a cloned voice confirming the request on a phone call, was enough to move $108 million out of Italy's biggest bank before anyone caught on.
What's striking here isn't the technical sophistication of the attack; it's often crude. It's the reflexive trust we still place in a face on a screen or a voice in our ear. That trust is the real point of failure, well before the quality of the fake.

Danger #2: intimate violation that overwhelmingly targets women
The second front is non-consensual sexual deepfakes. The emblematic case remains the mass circulation, in January 2024, of fabricated pornographic images of Taylor Swift. But this kind of digital assault doesn't stop at celebrities: the Westfield High School case above, with more than 30 identified minor victims, shows that any teenager can become a target of a classmate armed with a smartphone and a free app.
The Grok scandal, between December 2025 and January 2026, changed the scale entirely. This was no longer isolated incidents but mass usage, tied to a paid chatbot feature marketed as "spicy," that generated nude or sexualized images of real women without consent, and, in some cases, of minors. The UK's Ofcom opened a formal investigation, and Indonesia and Malaysia became the first countries to block Grok outright, citing "a serious violation of human rights, dignity and the safety of citizens." Both later restored access after xAI committed to further safeguards.
What both cases share, fundamentally, isn't a detection problem; the victim already knows the image is fake. The real problem is the speed of distribution, the near-impossibility of scrubbing content from everywhere at once, and a legal system that almost always arrives after the damage is done.
Danger #3: eroding collective trust and the electoral risk
The third front is more diffuse but arguably the most dangerous long-term: mass disinformation. A 2025 iProov study exposing consumers to real and fake content found that only 0.1% of participants could correctly identify every real and fake item, even when explicitly told to look for fakes. Separately, Pew Research Center found that 76% of Americans want AI-generated content clearly labeled, but only 12% feel confident they could actually spot it themselves.
The Sora affair illustrates the difficulty of containing this, even for the company that built the tool. Launched by OpenAI in September 2025 with a TikTok-style social feed, the app was almost immediately used to generate videos of deceased public figures in degrading scenarios. Bernice King, Martin Luther King Jr.'s daughter, wrote publicly asking creators to stop; Zelda Williams, Robin Williams' daughter, pleaded on Instagram for people to "just stop doing this to him"; George Carlin's daughter said her family was working to combat deepfakes of the late comedian; and videos depicting graphic violence against physicist Stephen Hawking circulated widely, according to NBC News.
OpenAI shut the app down in March 2026, citing shifting priorities rather than the controversy directly, but the sustained backlash over impersonation had clearly weighed on the app throughout its short life. Even the company best equipped, technically, to moderate its own platform couldn't manage it at scale, a useful lesson tied to the broader pattern of invisible harms in generative AI.
The losing race: why detection will never be enough
Here's where we want to push back against the prevailing advice. "Just use a detection tool" is aging badly.
Real-world performance numbers are damning. The Deepfake-Eval-2024 benchmark, built from deepfakes actually circulating on social media in 2024, found that state-of-the-art open-source detectors saw their accuracy (measured as AUC) collapse by 45–50% compared to their scores on tidy academic datasets, the moment they met real, messy, compressed, in-the-wild content. Human performance is no better: a 2024 meta-analysis in Computers in Human Behavior Reports, pooling 56 peer-reviewed studies and 86,155 participants, put average human deepfake-detection accuracy at just 55.5%, barely above a coin flip.
Structurally, this is a zero-sum game: every improvement to a detector feeds, via adversarial training, directly into the next generator's improvement. You never catch a train accelerating at the same rate as you are.
That's why the real battleground over the next five years won't be "how do we spot the fake" but "how do we prove the real." Standards like C2PA (the Coalition for Content Provenance and Authenticity), which attach verifiable cryptographic metadata to content at the moment of capture, flip the question entirely: instead of asking an algorithm to guess whether a video is fake, you ask the camera itself to digitally sign what it actually recorded at the point of creation. That shift is already being nudged along by the EU AI Act's Article 50, which requires providers of AI systems generating synthetic audio, image, video, or text, general-purpose systems included, to mark their outputs in a machine-readable format that's detectable as AI-generated. But until that kind of digital signature and watermarking becomes standard across every device and platform, the burden of proof will keep falling on the alleged victim of a fake rather than on its creator.

What the law does, and doesn't, say
Since France's loi SREN of May 21, 2024, French law has its own specific arsenal. Article 226-8 of the Penal Code, updated to cover algorithmically-generated content, now punishes distributing a fabricated montage of someone's image or voice without consent. But the penalty applies only under one strict condition: the artificial nature of the content must not be "expressly mentioned."
That's precisely where a loophole sits that rarely gets raised in mainstream coverage: if a creator clearly states that their content was AI-generated, the offense under Article 226-8 evaporates. Someone acting in bad faith can distribute a potentially devastating deepfake without exposure under this specific law, simply by playing the transparency card about the content's artificial nature.
It's worth being precise here, because the loophole doesn't cover everything: Article 226-8-1, the separate provision covering sexual deepfakes specifically, closes this exact gap, a disclosed sexual deepfake shared without consent is still an offense under French law, punishable by up to two years in prison and a €60,000 fine, rising to three years and €75,000 for online distribution. And even where the 226-8 loophole applies, it isn't blanket immunity: if the content damages someone's honor or reputation, ordinary offenses like defamation can still apply.
For readers outside France, the comparison points are useful. In the US, the TAKE IT DOWN Act, signed into federal law in May 2025, criminalizes publishing non-consensual intimate imagery, including AI-generated "digital forgeries", and requires platforms to remove reported content within 48 hours, enforced by the FTC. A broader federal bill, the NO FAKES Act, which would create a federal property right over a person's voice and likeness, has bipartisan support but remains stalled in committee as of 2026. At the EU level, Article 50 of the AI Act, in force since August 2, 2026, requires machine-readable marking of AI-generated content and carries fines of up to €15 million or 3% of global annual turnover, whichever is higher. Our deeper look at AI and legal liability covers the structural limits of law versus AI systems more broadly.
How to actually protect yourself
A few practical habits, drawn from official guidance and the real cases above:
On suspicious calls: if someone urgently asks for money by phone, even in a familiar voice, hang up and call the person back on a number you already have, never the one that just called you.
At work: for any unusual transfer request received by video call or voicemail, require verification through a completely separate channel (a known landline, an in-person check) and escalate it before acting.
On social media: limit posting long clips of yourself talking directly to camera; that's exactly the raw material voice cloning needs.
If you're targeted by a defamatory or sexual deepfake: save time-stamped screenshots, report the content to the platform, and file a police report.
For parents: talk to your teenagers about sexual deepfakes before an incident happens at their school. The cases above show this hits ordinary students, not just public figures.
Should we be fatalistic, or is there actually something we can do?
This piece carries a conviction that won't please anyone selling a miracle fix: no, there is currently no definitive technical countermeasure against deepfakes, and anyone claiming otherwise is selling something. What has changed is the nature of the risk. We've gone from an exotic threat, once reserved for intelligence agencies and Hollywood studios, to a tool available to any high schooler or scammer. That democratization, more than the quality of the fakes themselves, is what should actually worry us.
The good news, such as it is, is that the fix probably won't come from a better detector. It'll come from a shift in collective posture: stop putting the burden of proof on the victim of a fake, and start giving the real thing a signature that can't be forged. We're not there yet. And until we are, the best available defense remains the lowest-tech one of all: pick up the phone and check, the old-fashioned way, before trusting what your own eyes just showed you.
FAQ
What exactly is a deepfake? A deepfake is audio, video, or image content generated or altered by artificial intelligence, most often via GANs or diffusion models, to depict a real person saying or doing something they never actually said or did.
How can you spot a deepfake? There's no foolproof method. Classic tells (irregular blinking, inconsistent lighting, a slightly metallic voice) become less reliable as the technology improves. The best protection is cross-verification: checking the same information against independent sources rather than relying on your own eyes.
Is it illegal to create a deepfake? In the US, creating a deepfake for strictly private use isn't itself a federal crime, though distributing non-consensual intimate imagery, real or AI-generated, is now a federal offense under the TAKE IT DOWN Act. In France, publishing a deepfake of someone without consent is a crime under the SREN law of May 2024, unless its artificial nature is obvious or explicitly disclosed.
What are the penalties for a non-consensual sexual deepfake? In France, Article 226-8-1 of the Penal Code provides for up to two years in prison and a €60,000 fine, rising to three years and €75,000 for online distribution. In the US, the TAKE IT DOWN Act carries up to two years' imprisonment for adult victims and three years where minors are involved.
Are deepfake detection tools reliable? Not consistently. Accuracy that looks strong in a lab setting drops significantly against compressed, low-quality, or very recently generated content that wasn't part of a detector's training data, real-world benchmarks show accuracy collapses of 45–50% compared to academic test conditions.
How do you protect yourself from a voice-cloning scam? Make independent verification through a different channel standard practice before any money transfer, even under apparent urgency, and limit how much of your own voice you post publicly.




Comments