Deepfake
Deepfake defined: what a deepfake is and what it means for a claim file
A deepfake is media generated by artificial intelligence rather than recorded by a camera: an image, video, audio clip or document that depicts something which never happened. In an insurance context it means evidence with no real-world original behind it.
The term began with face-swapped video and now covers the whole synthetic family: fabricated damage photos, cloned voices, invented invoices and certificates, entire accident scenes that never took place.
Deepfake vs shallowfake: how they differ
The useful test is origin. If there is no real-world original behind the file, it is a deepfake or a generated image. If there is a real original that somebody altered with a phone app or Photoshop, it is a shallowfake, and no AI was needed to make it.
- Production: a deepfake comes out of a model. A shallowfake comes out of an editing tool anybody already has.
- Effort: convincing synthetic evidence still takes intent, prompting and iteration. Editing a genuine photo takes minutes and no learning curve.
- Volume: the edited real file is by far the more common submission today, precisely because the skill floor is on the ground.
- Detection: model fingerprints and impossible physics point at synthetic media. Edit residue, metadata gaps and image reuse point at shallowfakes. Tuning for one leaves you blind to the other.
Where deepfakes turn up in claim files
- Damage imagery: a generated photo of a vehicle, roof or appliance in a state it was never in.
- Scene fabrication: a whole collision context, complete with plausible weather, road surface and debris.
- Documents: receipts, valuations, medical letters and repair estimates produced to order from a text prompt.
- Voice and identity: cloned audio used on intake calls, and synthetic identity documents at policy inception.
- Ownership proof: images of high-value items that were never owned, generated to support a theft or loss claim.
Deepfake example: proof of an item that never existed
A theft claim arrives for a high-value watch. The supporting evidence is a clear photograph of the watch on a kitchen table and a purchase receipt from a retailer, both submitted as email attachments.
Neither file has an original. The image was generated from a prompt and the receipt was produced alongside it, with a plausible date and a plausible total. Admiral has reported finding exactly this shape of evidence, naming fake number plates, imaginary watches and exaggerated damage among the AI-generated material in its 2025 caseload.
What the reported numbers show
Admiral detected £86.8 million of fraudulent motor, home and travel claims in 2025, a 71% rise on the £50.9 million detected in 2024, and tied part of that increase to easier access to AI tools that alter images and produce documents that never existed.
Aviva reported more than 18,400 suspicious claims worth £233 million across its brands in 2025, and said a growing number were supported by AI-generated images and manipulated documents, mostly in motor.
Read both carefully. They are detection figures. A rise in detected fraud reflects more fraud and better detection at the same time, and no public number tells you the size of what nobody caught. The wider picture is in AI-generated insurance fraud.
Why detection scores do not settle a claim
Deepfake detectors return probabilities, and those probabilities move every time generation models change. A score is an input to a decision, not the decision. Three habits keep that honest:
- Treat signals as reasons to look: a flagged file earns a closer review, not an automatic decline. Genuine claimants trip technical signals for ordinary reasons.
- Keep the human decision-maker central: the investigator weighs the evidence, the context and the history. A model does not settle a claim.
- Attack provenance, not appearance: asking where a file came from and how it reached you is more durable than asking a classifier whether it looks generated. Generation quality improves every year. The question of origin does not get easier to dodge.
The practical defence is to reduce how much unverifiable material enters the file at all. Evidence recorded inside a controlled capture process, with a server-side receipt time and a cryptographic fingerprint, gives a reviewer something to check months later. An emailed attachment gives them a file and a hope. The standards that make visual evidence usable are set out in photo evidence in insurance claims.
One caveat worth repeating internally: a receipt timestamp proves when your organisation received the submission. It says nothing about when the loss occurred. Overstating that is a good way to lose an argument you should have won, and no amount of evidence tooling makes fraud impossible. It makes the material easier to question, earlier, with less argument about what you are looking at.
