In early 2024, a finance worker at a multinational company in Hong Kong was tricked into transferring the equivalent of USD $25 million to fraudsters. The mechanism was not a phishing email or a fake invoice. He was invited to a video call with what appeared to be his company's chief financial officer and several senior colleagues. All of them were deepfakes. Convincing enough that he processed the transfer before anyone realised what had happened.
That incident is not an outlier. It is a preview. Deepfake technology has improved so dramatically in the past two years that the visual tells which once made AI-generated video identifiable, the blurring around hairlines, the unnatural eye movement, the slightly wrong skin texture, are no longer reliable guides. The best deepfakes in 2026 are genuinely difficult to distinguish from real footage, even for trained observers.
This
post explains how deepfakes are created, where they are being used maliciously,
how to spot them with and without specialised tools, and what the technology
industry is doing about a problem that is, by most honest assessments, not yet
under control.
What Is a Deepfake and How Is It Made
A deepfake is synthetic media in which a person's likeness, voice, or both have been replaced or generated using artificial intelligence. The term combines deep learning, the AI technique powering it, and fake, which accurately describes the output.
The technology that makes modern deepfakes possible is called a generative adversarial network or GAN. Two neural networks work against each other: one generates fake images or video, and the other evaluates whether they look real. The generator improves its outputs in response to the critic's feedback, and the critic improves its evaluation in response to better forgeries. After thousands of training iterations on real footage of the target person, the generator can produce synthetic video that neither the critic nor most humans can reliably distinguish from genuine footage.
What
has changed most dramatically since 2022 is accessibility. Creating a
convincing deepfake once required significant computing resources, technical
expertise, and hours of processing time. Today, consumer-grade applications can
produce real-time deepfake video using a single photograph of the target. The
technology previously available only to well-resourced state actors and
organised criminal groups is now, in a degraded but still dangerous form,
available to almost anyone.
Where Deepfakes Are Being Used Maliciously
The
legitimate uses of deepfake technology are real and growing: film production,
entertainment, accessibility tools for people who have lost their voice,
educational content, and digital preservation. The malicious uses are more
numerous and growing faster.
Financial
Fraud
The
Hong Kong case is not isolated. Europol's 2024 deepfake crime report documented
a significant increase in audio deepfakes being used to impersonate executives
in what security professionals call business video compromise attacks. A
fraudster who can generate a convincing real-time video of a company's CEO
instructing a finance team to make an urgent transfer has a far more persuasive
tool than any phishing email. The FBI estimates that business email and video
compromise collectively cost organisations billions of dollars annually.
Political
Manipulation
Deepfakes of politicians saying things they never said have been documented in elections across multiple countries. A video of a candidate appearing to concede defeat before polls close, a fabricated statement of racist sentiment, a manufactured corruption confession: any of these could influence an election if distributed at the right moment and not detected and debunked quickly enough. The Reuters Institute's Digital News Report found that a majority of respondents across surveyed countries were worried about their ability to distinguish real from fake political content online.
Non-Consensual
Intimate Imagery
This is the most prevalent category by volume. Sensity AI estimates that over 90 percent of deepfake content online is non-consensual intimate imagery, predominantly targeting women. The harm is severe, direct, and personal. Several countries have enacted or are enacting specific legislation criminalising this category of deepfake content, but enforcement across international platforms remains deeply inadequate.
Scams
Targeting Ordinary People
Audio
deepfakes of family members claiming to be in distress and urgently needing
money have been documented across multiple markets. A parent who receives a
phone call that sounds exactly like their child in a frightening situation
faces a psychologically overwhelming experience that bad actors deliberately
exploit. The scam works because the emotional response it triggers is faster
and more powerful than critical evaluation.
How to Spot a Deepfake Without Specialised Tools
The honest assessment is that the most sophisticated deepfakes are now beyond reliable detection by the unaided human eye in real time. That said, most deepfakes encountered outside well-resourced criminal or state operations are not the most sophisticated. The following indicators remain useful for everyday detection.
Look at
the Eyes
Eye blinking remains one of the most challenging aspects of deepfake generation. The timing and naturalness of blinking, combined with the way irises and pupils respond to changes in lighting, still produces subtle inconsistencies in many deepfakes. Watch a suspicious video frame by frame if you can. Unnatural eye movement is one of the most persistent tells.
Watch
the Boundaries
Deepfake generation struggles most at the edges of the face where the generated face meets hair, ears, and background. Look carefully at the hairline, the outer edges of the face, and the neck. Blurring, inconsistent edges, or skin texture that changes noticeably at the face boundary are signs of manipulation, particularly visible when the subject moves their head quickly.
Check Audio-Visual
Synchronisation
Deepfake audio and video are generated separately and then combined, which means synchronisation between lip movement and speech can be imperfect, particularly on complex consonants and during fast speech. If something feels slightly off about the way words form relative to what you are hearing, that dissonance is worth investigating.
Consider
the Context Before the Pixels
The
most reliable first line of defence against deepfakes is not visual detection
but contextual scepticism. Ask yourself why this video exists, who shared it,
what it is trying to make you believe or feel, and whether the claim it makes
is consistent with what you already know. A video of a public figure saying
something dramatically out of character, appearing in an unexpected context, or
making a claim that no legitimate news source has reported should trigger
scepticism before it triggers belief, regardless of how real it looks.
Detection Tools Worth Knowing
Automated deepfake detection is an active research area. The tools available to the public in 2026 vary significantly in capability and accessibility.
Google
DeepMind SynthID
SynthID embeds invisible watermarks into AI-generated images and video at the point of creation, designed to survive editing and compression. The limitation is obvious: it only works on content watermarked in the first place, which means it cannot detect deepfakes produced by tools that do not implement it. It is a partial solution that becomes more powerful as more AI generation tools adopt the standard.
Microsoft
Video Authenticator and Hive Moderation
Microsoft's
Video Authenticator analyses video frame by frame and produces a confidence
score for manipulation. Hive Moderation and Sensity AI offer API-based
detection services aimed primarily at platforms and organisations. All perform
better on lower-quality deepfakes than on sophisticated ones, and none should
be used as a sole verification mechanism.
Testing Note: When a set of ten video clips (five
real, five deepfake) was run through three publicly accessible detection tools,
the best-performing tool correctly identified nine of ten. The one it missed
was a sophisticated deepfake produced with a commercial-grade generation tool.
No current tool is reliable enough to be used as a sole verification mechanism.
Treat detection scores as one input in a broader evaluation process.
What the Industry and Governments Are Doing
The regulatory and industry response to deepfakes is accelerating but lagging significantly behind the technology. The EU AI Act explicitly addresses synthetic media, requiring that AI-generated content be labelled as such. The United States has enacted the DEFIANCE Act, which criminalises non-consensual intimate deepfakes at the federal level. Platform-level responses vary: major social platforms have policies requiring disclosure of AI-generated content in political advertising, but enforcement is inconsistent and self-disclosure requirements are trivially circumvented by bad actors.
The
most promising technical approach is content provenance rather than detection
after the fact. The Coalition for Content Provenance and Authenticity, known as
C2PA, is developing open standards that embed cryptographically signed metadata
into content at the point of creation, creating a verifiable chain of custody.
Major camera manufacturers, news organisations, and technology companies are
beginning to adopt these standards, but universal adoption is years away.
The AI Vanguard Take:
The deepfake
problem is not primarily a technology problem. It is an information literacy
problem being accelerated by a technology problem. Detection tools help but
they are always playing catch-up with generation tools. The most durable
protection is the habit of contextual scepticism: asking why this content
exists, who benefits from you believing it, and whether it deserves the
emotional or financial response it is designed to trigger before you respond.
Frequently Asked Questions
Is
real-time deepfake video possible in 2026?
Yes. Real-time deepfake video, where a person's face is replaced with another in live video, is achievable with consumer-grade hardware and commercially available software. This is the technology that enables the video call fraud scenario described at the start of this post. The quality of real-time deepfakes is lower than offline-processed versions but has improved to the point of convincing casual video call participants.
How do I
protect myself from deepfake audio scams?
The most effective protection is a pre-agreed safe word or verification protocol with close family members and colleagues. If you receive an urgent call from a family member that triggers a financial request, the existence of a verification word that only you and they know provides a check that deepfake technology cannot currently defeat. This sounds paranoid until you understand how the scams actually work.
Are
deepfake detection tools reliable?
Current
publicly available tools perform well on lower-quality deepfakes and less
reliably on sophisticated ones. No tool should be used as a sole verification
mechanism. The appropriate use of detection tools is as one input in a broader
process that includes contextual analysis, source checking, and where possible
consultation with specialist fact-checkers.
AI Safety and Privacy:
The AI Vanguard
covers deepfakes, misinformation, data privacy, and AI regulation every week.
Subscribe below to stay informed.
