RN
A guide to spotting deepfakes written in 2021 is now a historical document, and reading it as advice will get you the wrong answer three separate times.
The tells it taught you to look for have been engineered out. The sites it sent you to are mostly gone. And the legal position it described as a grey area with regional variation has hardened into criminal offences on both sides of the Atlantic — including, in the UK, an offence you commit by making the file at all, whether or not you ever show it to anyone.
That is a lot of change in five years. It is worth going through properly.
What the word covers now
The original meaning was narrow: a face swapped onto another body in video, produced by training a model on images of both people. That technique still exists and is now the least of it.
Current systems generate from a text description, animate a still photograph, clone a voice from a short sample, or alter an existing video in place. Some produce a person who does not exist. Others produce a specific real person doing something they did not do. The second category is the one that matters here, and the term "deepfake" is now doing so much work that it has stopped being precise. The useful distinction is not how something was made — it is whether a real, identifiable person is depicted, and whether they agreed.
Why the detection checklist stopped working
The 2021 advice was to watch the lips, watch the eyes, watch for mismatched body language and abrupt speed changes. That advice was reasonable for face-swap output produced by hobbyist tools of that era.
It has aged badly, and the reason is structural rather than incremental. Older methods worked by pasting a generated face into existing footage, so the seam was the weak point — hairlines, jaw edges, the boundary where lighting failed to match. The blinking tell that circulated widely around 2018 existed because training sets were full of photographs of people with their eyes open. Every one of those artefacts became a benchmark, and benchmarks get optimised against. They were the first things fixed.
Contemporary generation frequently produces the entire frame rather than patching one, so there is no seam to find. Voice cloning has closed the audio gap that once gave lip-sync mismatches away. What remains catchable by eye is inconsistent physics across a long take, hands and fingers under motion, text or logos in the background, and jewellery and glasses that change shape between frames — and these are all getting fixed too, on the same treadmill.
The honest position for a general reader in 2026: you cannot reliably tell by looking. Detection is now a provenance problem rather than a perception problem. Where did this file come from, who first posted it, does the account have a history, is any outlet that would be liable for getting it wrong carrying it? Those questions still work. Squinting at the eyes does not.
The claim that this mostly targets famous women, revisited
Deeptrace's 2019 report on the deepfake landscape found that the overwhelming majority of the videos it catalogued were pornographic and that these almost exclusively depicted women, mostly entertainers. That figure gets quoted constantly, and it is worth saying clearly that it describes a sample taken seven years ago from a landscape that has changed completely.
What replaced it is worse rather than better. The growth area is not celebrities. It is "nudify" apps — services that take an ordinary photograph of an ordinary person and return a fabricated nude, in seconds, for a small fee, with no technical skill required. The targets are classmates, colleagues, ex-partners and strangers whose social media is public. Schools in multiple countries have dealt with incidents involving pupils generating images of other pupils.
That shift dismantles the ethical argument the 2021 article rested on. Its position was that celebrities are public figures who should expect parodic and sexual uses of their image, and that restricting this amounts to telling people what art they may make. Set aside whether that was ever persuasive about famous people. It has no purchase at all on a sixteen-year-old whose classmate uploaded her school photograph to an app.
The enforcement response has followed the harm. Meta sued the developer behind the CrushAI nudify app in mid-2025 over advertising that repeatedly evaded its review process. The San Francisco City Attorney sued sixteen of the most-visited undressing sites, and in July 2026 sent cease-and-desist letters to Apple and Google naming thirteen apps still listed in their stores. The pressure is landing on distribution and payment rather than on individual creators, which is generally where pressure lands when the creators are numerous and anonymous.
Where the law actually stands
This is the section that has changed most, and the 2021 description of a patchwork with Scotland strict and England permissive is simply no longer the world.
United States. The TAKE IT DOWN Act passed Congress in April 2025 and was signed that May. It criminalises knowingly publishing intimate visual depictions of non-consenting adults and of minors, and it says explicitly that AI-generated "digital forgeries" count. Penalties include fines and up to three years' imprisonment. It also places a duty on covered platforms to operate a notice-and-removal process and to act on a valid request within 48 hours — the criminal provisions applied immediately, and platforms were given until May 2026 to have the removal process running. This sits on top of state laws that already existed in most of the country.
United Kingdom. Section 138 of the Data (Use and Access) Act 2025 came into force on 6 February 2026. It makes it an offence to intentionally create — or to ask someone else to create — a purported intimate image of an adult without their consent or a reasonable belief in it. Note what that covers: creation, not distribution. Sharing was already criminalised under earlier intimate-image provisions. The 2025 provision closes the gap where someone made the file and had committed no offence until they sent it to somebody.
Both changes point the same direction, and they were passed by legislatures that agree on very little else. Other jurisdictions are moving too, at different speeds and with different definitions, so treat everything above as the shape of the trend rather than as the law where you are. It is also new law, which means how far it reaches in practice will be worked out by prosecutors and courts over the next several years rather than being settled now.
The sites the old article recommended
It listed three. The largest of them, MrDeepFakes, shut down in May 2025 after a service provider terminated it, posting a notice saying data loss had made continuing impossible and that it would not relaunch — while warning that anything later claiming to be it would be an impostor. The closure came days after the TAKE IT DOWN Act cleared Congress and followed legal pressure reported from the UK and the Netherlands.
No replacement is offered here, and that is deliberate rather than an omission. An article cannot responsibly hand you a list of places to obtain sexual images of real people who did not agree to them, at a point when making such an image is a criminal offence in one of the two jurisdictions discussed above and publishing it is an offence in both.
If the underlying interest is fabricated adult imagery as such, generators that produce people who do not exist are a different proposition entirely, and the entire legal and ethical problem described on this page turns on the depiction of a real identifiable person.
The short version
Deepfakes stopped being a curiosity somewhere around the point the tools became free and required no skill. The technical race is not one an ordinary viewer can win by looking harder, so provenance is the only reliable check available. The victim profile has moved from famous women to anyone with a public photograph. And the legal position has moved from ambiguous to explicit, quickly, in a direction that treats making the file as the act rather than sharing it.
For how the volume problem developed, the companion piece on the scale of it covers that ground. And if any of this happened to you, the removal duty now placed on platforms is a real mechanism with a deadline attached, not a courtesy — the request is worth making.