Made by a model? We find out.
A few seconds of someone's voice is enough to clone it. A single photo is enough to put their face on a video call. We examine suspect media for the traces generative AI leaves, and tell you what can be shown about whether a model made it.
Every kind of generated media.
Generative models leave traces: in the signal, in how faces and voices are assembled, and in the provenance data a growing number of tools attach. We look for those traces, and only those. For classic audio, video and image forensics, TechFusion is the place.
If you've received a suspect call, video or file
- Don't act on the request. Verify through a channel you already trust.
- Keep the original file, message or recording. Don't screenshot it, re-record it or re-save it.
- Note where it came from: number, account, platform, time.
- Call us or start a case. Platforms delete things; preservation comes first.
Voice-clone detection
Calls, voicemails and voice notes in a familiar voice that ask for money, credentials or secrets. We test the speech for the artifacts text-to-speech and voice-conversion models leave, and for signs of real-time voice changers.
Face-swap and lip-sync detection
Video calls, clips and livestreams with a real person's face or a puppeted mouth. We look for the blending boundaries, temporal flicker and audio-to-lip mismatches that face-swap and lip-sync models produce.
AI image-generator identification
Images produced or inpainted by diffusion and other generative models. We test for generator signatures and, where the evidence allows, which family of model produced the image.
Content credentials and AI watermarks
A growing number of AI generators sign their output under the C2PA standard ("content credentials"), and some embed invisible watermarks. We read and verify them, check whether they have been stripped, and explain what their presence or absence does and doesn't prove.
AI-generated documents and text
Invoices, letters, chat screenshots and written statements produced by language or image models. We test for the statistical and visual signatures of generation, and say plainly how far current methods can go.
The selfie check was the target.
Remote onboarding trusts a camera and a document scan. Attackers now feed both with generated faces, injected video streams and AI-made IDs. Banks, fintechs, marketplaces and employers need to know how it was done, and how often.
Camera-injection attacks
Virtual cameras, emulators and hooked apps that feed a pre-made or live deepfake into the verification flow instead of a real camera. We examine session artifacts and video for the signs of injection.
Liveness and face-match spoofing
Generated faces, face swaps and replayed sessions used to pass liveness and face-match checks. We establish which technique was used and why the check accepted it.
AI-generated ID documents
Driver's licences, passports and utility bills produced with generative models. We test submitted images for generator traces and template reuse across applications.
Synthetic identity rings
Many applications, one operator. We link synthetic identities through shared generation traces, reused assets and session patterns, so a fraud team can see the whole ring.
Stated confidence. Stated limits.
| Finding | What it means | Typical basis |
|---|---|---|
| Consistent with AI generation | Several independent methods find the traces generative models leave | Synthesis artifacts, generator signatures, provenance data agree |
| Partly AI-generated | Specific regions or segments show generation, the rest doesn't | Localized generator traces, AI inpainting or spliced synthetic speech |
| No AI-generation traces found | None of the tested methods found signs of a generative model | Stated with its limits: new generators may not yet be detectable |
| Inconclusive | The material can't support a reliable conclusion either way | Heavy compression, too short, or only a copy of a copy available |