Generated video, cloned voices and fabricated images are now cheap, fast, and good enough. This lesson is about what actually happens with them and what you can realistically do.
The real harm is not what the headlines say
The fear is a fake video of a president starting a war. The reality, measured repeatedly, is more ordinary and worse for the people involved.
By volume, the overwhelming majority of deepfake material online is non-consensual sexual imagery, and almost all of it targets women. Studies since 2019 have consistently put this above 90% of deepfake videos found online. Most victims are not famous. School students have found images of themselves circulating, made by classmates from a photo lifted from social media.
The second big category is fraud. In 2024 a finance employee at the Hong Kong office of an engineering firm transferred about US$25 million after a video call with what appeared to be the company's CFO and several colleagues. Every participant on the call except the victim was synthetic. Voice cloning fraud runs the same play at small scale: a call from a relative's voice, distressed, needing money now.
Third is targeted political material — often not a fake speech but a fake robocall or a fake audio clip released hours before a vote, when there is no time to debunk it.
Spotting them: an honest assessment
You will read lists of visual tells. Count the fingers, check the blinking, look at the ears and teeth, watch for hair edges and warped background text. These worked in 2020, work sometimes now, and will work less every year. Do not build your defence on them.
Detection software exists and is not reliable enough to trust on a single item. It fails on compressed and re-uploaded media, which is what actually circulates.
What holds up better is thinking about provenance rather than pixels:
Where did this come from? Not who forwarded it — who originally published it, and can you reach that source. A clip with no traceable origin is the strongest single warning sign.
Does anyone else have it? Real events are recorded by more than one person and reported by more than one outlet. A dramatic event that exists in exactly one video is suspicious in itself.
What does it want from you? Synthetic media is usually pointed at an action: send money, believe this about that person, share this before the vote. Urgency plus a request to transfer something is the pattern.
Can you check by another channel? This is the one that stops voice fraud. Hang up and call the person back on the number you already have. Agree a family code word now, before you need it. No cloned voice survives a callback.
Labelling, and why it only half works
The C2PA standard attaches signed provenance metadata to a file — what made it, what edited it. Camera makers and major AI tools are adopting it. Invisible watermarking embeds a signal in the pixels or audio.
Both are real progress and both are limited. Metadata is stripped by most social platforms on upload. Watermarks survive some transformations and not others, and an adversary who wants them gone can usually get them gone. And neither addresses the deeper asymmetry: the *absence* of a label proves nothing, since most genuine media carries no provenance either.
What you can do is label your own. If you publish something AI-generated or AI-edited, say so plainly in the visible caption, not only in metadata that will be stripped. It costs a sentence.
The liar's dividend
The most durable damage may not be false things believed. It is true things disbelieved. Once everyone knows video can be faked, anyone caught on camera has a ready defence — *that is a deepfake*. Real evidence of real wrongdoing gets waved away. That erosion needs no fake to be made at all.
Before you move on