Deepfakes and synthetic faces

The word deepfake appeared on Reddit in 2017 for face-swapped pornography and now covers any media generated or altered by a learning model. Synthetic faces, the kind served by sites such as thispersondoesnotexist.com from 2019, are a related but distinct thing: a face with no original. The distinction matters for face search, because a search can only find a face that has been published somewhere.

How are they made?

What happens when you face-search a synthetic face?

Usually nothing, and that is informative. A face generated by StyleGAN or a diffusion model has no source photo, so an engine finds no match unless the same generated image has been posted somewhere public. A dating profile whose photo returns zero results everywhere is therefore either new, private, or synthetic, and the search cannot tell which. The reverse case is the dangerous one: a deepfake built from a real person's photos contains that person's face geometry, so a face search on the deepfake can return the victim's real pages, which is exactly how victims of 'nudify' apps discover that their images were used.

What are the tells that still work?

• Accessories and symmetry: earrings that do not match, glasses frames that merge with skin, teeth with irregular counts.
• Backgrounds and edges: text that is almost legible, hair strands that dissolve into the background, mismatched lighting between face and scene.
• Hands and ears: still the parts models get wrong most often, though less with each generation.
• Consistency across photos: a real profile has the same face across angles and years; a generated set often drifts.
• Behaviour: a refusal to join a live, unscripted video call outweighs every visual tell.

These tells degrade with every model release, and detection tools trained on one generator fail on the next. No consumer tool detects synthetic images reliably in 2026; vendors quoting accuracy above 90% are reporting results on their own test sets.

How do platforms and the law treat them?

1. 2017-11

   Deepfakes subreddit

   Face-swapped pornography named; Reddit bans the community in February 2018.

2. 2019-02

   StyleGAN faces published

   thispersondoesnotexist.com shows photoreal synthetic faces to the public.

3. 2019-10

   California AB 602 and AB 730

   Civil remedy for sexual deepfakes; restrictions on political deepfakes near elections.

4. 2023-02

   NCMEC launches Take It Down

   Hash-based removal for intimate images of minors, including AI-generated ones.

5. 2024-08

   EU AI Act in force

   Transparency duties for deepfakes; prohibition on untargeted facial-image scraping applies from February 2025.

6. 2025-05

   US TAKE IT DOWN Act signed

   Criminalises publishing non-consensual intimate images, including AI-generated ones; platforms must remove within 48 hours of a valid request from May 2026.

7. 2026-05

   FTC begins enforcing TAKE IT DOWN

   Warning letters to twelve 'nudify' services; TakeItDown.ftc.gov complaint site opens.

The TAKE IT DOWN Act is the first US federal law to reach AI-generated intimate images directly. The EU AI Act requires deepfakes to be labelled and bans scraping faces to build recognition databases. The UK Online Safety Act made sharing intimate deepfakes an offence in 2024, with creation added in 2025. India relies on the IT Act and Rules and criminal provisions rather than a deepfake-specific statute.

What should a victim do?

1. 01 Keep evidence: URLs, screenshots with dates, the account that posted.
2. 02 Use the platform's non-consensual intimate image report; under the TAKE IT DOWN Act US platforms must act within 48 hours.
3. 03 Create a case at StopNCII.org (adults) or Take It Down (under 18) so participating platforms block re-uploads by hash.
4. 04 Run a face search to find further copies on public pages, then repeat the reports; the NCII removal directory lists routes by country.
5. 05 Report to police where the law criminalises the act, which it now does in the US, UK, Australia and parts of the EU.