What is a deepfake?
Short answer
A deepfake is media made or changed with AI so it appears to show a person, voice, object, place, or event that may not be real.
People often use deepfake to mean face-swap videos, but the problem is wider now. Synthetic media can include AI-generated images, fake profile photos, voice clones, edited screenshots, manipulated documents, and videos with altered speech.
Not every AI edit is harmful. A film effect, translation dub, or creative image may be fine. The risk appears when the media is used to mislead, impersonate, defraud, harass, or manufacture evidence.
What types of synthetic media should you know?
Different deepfake formats need different checks. A video detector will not automatically solve a fake invoice, and an image provenance check will not tell you whether a voice note is cloned.
| Type | What it is | Common risk |
|---|---|---|
| Generated image | A picture made by an AI image model. | Fake events, fake products, fake people. |
| Face swap | One face mapped onto another person's body or performance. | Impersonation or non-consensual imagery. |
| Lip-sync edit | Mouth movement changed to match different speech. | Fabricated statements. |
| Voice clone | Synthetic audio made to sound like a real person. | Payment fraud or impersonation. |
| AI inpainting | Only part of an image is replaced or generated. | Subtle evidence manipulation. |
| Synthetic document | Fake layout, text, signature, image, or attachment. | Fraudulent applications, claims, or onboarding. |
Why does deepfake detection matter?
The harm usually comes from a decision. Someone pays an invoice, approves an application, publishes a story, hires a candidate, shares a rumour, or believes a person said something they did not say.
That is why the goal is not only to ask is this AI? The better question is: what decision is this media trying to influence, and what evidence would we need before trusting it?
How does deepfake detection work?
A deepfake detector may examine faces, motion, audio patterns, file metadata, provenance credentials, pixel-level signals, or document structure. The useful method depends on the file you actually have.
| Media | Useful checks | Common complication |
|---|---|---|
| Image | Provenance, metadata, pixel signals, source search. | Screenshots and recompression. |
| Video | Frame consistency, motion, audio-video sync, encoding traces. | Low quality clips and platform transcodes. |
| Audio | Spectral patterns, voice behaviour, source confirmation. | Noise, codecs, and very short clips. |
| Document | Metadata, layout, fonts, arithmetic, entities, provenance. | Mixed genuine and fabricated parts. |
Detection, provenance, and verification are not the same
This distinction helps people avoid bad conclusions. Detection asks whether the file resembles synthetic or manipulated media. Provenance asks what origin or edit history is recorded. Verification asks whether the claim being made is true.
| Approach | Question it answers | Example |
|---|---|---|
| Detection | Does this look generated or manipulated? | A face swap detector flags a video. |
| Provenance | What origin or edits are recorded? | A file has signed AI-generation credentials. |
| Verification | Did the event or claim actually happen? | The person confirms through a known phone number. |
What should you do if you suspect a deepfake?
The first move is to pause. Do not send money, publish, accuse, approve, or reject based on a suspicious file alone.
| Step | What to do |
|---|---|
| Preserve evidence | Save the original file, URL, account, timestamps, and message context. |
| Use the right checker | Match the tool to the file type: image, document, video, or audio. |
| Check provenance | Look for origin records or missing/stripped evidence. |
| Verify another way | Contact the claimed person or organisation through a trusted channel. |
| Corroborate context | Look for earlier copies, reports, records, or independent witnesses. |
| Escalate carefully | Report fraud, abuse, or safety threats through the right channel. |
What can Stipple check today?
Stipple inspects uploaded images and documents for provenance, AI-generation indicators, editing and tampering traces, and, in supported document workflows, internal consistency.
Stipple does not present itself as a general video or audio deepfake detector. That narrower claim matters. A trustworthy product should tell you what it checks and what it does not.
Three regulatory blocs now require identifiable AI content
Short answer
China (Sep 2025), South Korea (Jan 2026) and the EU (Aug 2026) all legally require AI-generated content to be identifiable. Deepfake policy just moved from principle to enforcement.
Marking is no longer a voluntary best practice — it is law in the three largest regulatory blocs. Anthropic now watermarks Claude text globally and signs C2PA credentials onto Claude-generated files; other signatories are implementing their own marks.
In China, platforms carry a legal duty to check for labels, which makes label-checking a platform obligation rather than only a user tool. Expect more provenance signals to exist, and verify them where you can.
The core problem is unchanged, though. Absence of a mark still proves nothing, because most synthetic content in circulation carries none — and the blocs regulate producers far more effectively than they reach the open internet.
| Regime | In force | Core duty | Distinctive feature |
|---|---|---|---|
| China — Labelling Measures + GB 45438-2025 | 1 Sep 2025 | Explicit (visible) + implicit (machine-readable) labels | Platforms must verify labels; label removal is itself prohibited |
| South Korea — AI Basic Act | 22 Jan 2026 | Indicate outputs are AI-generated; label realistic synthetic content | Enforcement grace period of at least a year |
| EU — AI Act Article 50(2) | 2 Aug 2026 | Machine-readable marking, detectable as artificial | The Code of Practice Anthropic and ~190 others signed |
Frequently asked questions
What is the difference between a deepfake and synthetic media?
Synthetic media is the broader term. Deepfake usually refers to convincing AI-made or AI-altered media involving a person, voice, identity, or event.
Can deepfakes always be detected?
No. Detection tools age as generation methods improve. Provenance and independent verification reduce the risk of relying on one classifier.
Can Stipple detect video or audio deepfakes?
Not as a general video or audio service today. Stipple currently focuses on uploaded images and documents.
Sources and further reading
- 01C2PA technical specification
- 02Content Authenticity Initiative
- 03NIST Open Media Forensics Challenge
- 04Europol: law enforcement and deepfakes
Educational guidance, not a forensic certification. Detection technologies and standards change; review material decisions against current evidence.