BERKELEY × GOOGLE DEEPMIND HACKATHONWitnessed · built with Google Gemini

Verified photos for every auto claim.

Your customer photographs the car live through a link. AI checks every photo against the policy. Your adjuster decides.

Why now

AI can edit a photo in seconds. It can't edit the car in front of the camera.

36%

of US consumers would consider editing a claim photo or document

Verisk, 2026
99%

of insurers have already seen manipulated or AI-altered evidence

Verisk, 2026
300%

rise in claims using doctored images, in one year

Allianz UK
32%

of insurers are confident they can spot a deepfake

Verisk, 2026

Sources: Verisk State of Insurance Fraud study, March 2026 (1,000 US consumers, 300 claims professionals); Allianz UK, 2021-22 to 2022-23, reported by The Guardian, May 2024.

The product

One link. Capture, check, decide.

Replace "send us some photos" with a guided live capture — checked before it ever reaches your adjuster.

SESSION K7-42Photograph the license plate
✓ Whole car
✓ Damage close-up
Take photo
01

Capture

A link by text. Live camera only — no app, no uploads, retakes on the spot.

✓Plate matches policyMatch
✓Same car, same colorMatch
!Damage on another sideReview
02

Check

Gemini reads every photo. Each one is compared with the policy and the story.

Approve
New photos
Refer to SIU
THE ADJUSTER DECIDES
03

Decide

Your adjuster sees the facts first. A person makes every call.

Features

The busywork, done before the file opens.

01

Live capture

Guided photos of the car, the damage, the plate and the VIN. No gallery, no edits.

02

Read automatically

Plate, VIN, make, color and damage — read by Gemini, never retyped.

03

Checked against the policy

Six checks on every claim, each with the photo that explains it.

04

Recycled photos caught

A photo already used in another claim is flagged at once.

05

New photos in a click

Something unclear? Ask for one more photo instead of sending an appraiser.

06

Sealed claim files

Locked the moment they are sent. Anyone can verify nothing changed.

Human in the loop

AI does the reading. People make the calls.

  1. Customer

    Captures

    Live photos, guided step by step, from a link.

  2. Google Gemini

    Reads & checks

    Every photo, against the policy and the story.

  3. Adjuster

    Decides

    Approve, ask for new photos, or refer to SIU.

  4. Your team

    Improves

    Every flag is marked right or wrong, so you always know how the AI is doing.

If the AI is unsure, it says so. Unsure is never a pass — and the AI never decides fraud.

Principles

Four rules we never break.

i.

Live, not uploaded

Every photo is taken through the link, in the moment.

ii.

Checked, not guessed

What the AI can't read is marked unchecked, never passed.

iii.

Sealed, not editable

Any change to a claim file after it is sent shows.

iv.

Decided by people

No reject button. Nobody is turned down by a machine.

Live capture makes cheating much harder, not impossible. Early product; no SOC 2 yet.

Pilot program

Make every claim photo one you can trust.

We are looking for a few US auto insurers to pilot Witnessed on one photo lane.

No sign-up. Demo workspace — please don't photograph personal documents.