Industries / Finance and business services / Insurance / InsurTech
Fraud
Software that looks for false or inflated claims and applications, by pooling records across insurers and scoring each claim.
6
Milestones achieved, from 2000 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2000
Claims pooled across insurers
ISO ClaimSearch is the anti-fraud database that ISO, now Verisk, set up for car, property and liability claims. By July 2000 "more than 4,000 insurers, self-insured entities and third-party administrators" reported claims to it, and it held 147 million records, Insurance Journal reported.
Credit
Insurance Services Office (ISO), now Verisk
2014
A register of fraudsters
In September 2012 the Association of British Insurers (ABI) and the Insurance Fraud Bureau announced the Insurance Fraud Register. It records, for five years, people "detected acting fraudulently towards insurers". It went live in January 2014 with "more than 20% of ABI members already signed up", against a goal of 90% of the market within five years.
LessonA shared register is only as useful as the share of the market that feeds it, and filling one takes years.
Credit
ABI; Insurance Fraud Bureau
2016
Fraud scores for sale
Shift Technology, founded in 2014, sells insurers software that scores claims for signs of fraud. By May 2016 it had processed about 50 million claims and said its flags were right 75% of the time. By March 2019 it reported more than 70 clients in over 25 countries.
Credit
Shift Technology (Jeremy Jawish, co-founder and chief executive), by its own account
2022
Wrongly accused by software
Michigan's unemployment insurance agency used an automated system, MiDAS, to find benefit fraud. About 40,000 people were wrongly accused, and in October 2022 the state agreed to settle the case for $20 million. This was public unemployment insurance, not a private insurer.
LessonAutomated fraud findings that people cannot easily challenge can harm thousands before anyone checks them.
Credit
Michigan Public, for the report
2023
Fraud screening in court
In December 2022 Black homeowners sued State Farm, alleging that its claims-screening algorithms gave them more scrutiny. They cited a 2021 YouGov survey of 648 white and 151 Black homeowners in which Black policyholders were 39% more likely to have to submit extra paperwork. On 11 September 2023 an Illinois court let their disparate-impact claim proceed under the US Fair Housing Act. A disparate-impact claim says a practice harms one group more, whatever the intent. Discovery about "algorithmic tools used to screen out potentially fraudulent or complex claims" was still under way in February 2026.
LessonA fraud screen that falls harder on one group can be challenged under civil rights law. ).
Credit
the plaintiffs in Huskey v. State Farm
2025
Edited photos, more claims
Allianz said in May 2024 that incidents of edited car photos in claims "increased by 300% between 2021-22 and 2022-23". Admiral said it found "£86.8 million in fraudulent claims" in 2025, "up 71% from £50.9 million in 2024". Its examples included "the same damaged car edited with a different number plate to claim twice". In June 2026 Aviva reported "a growing number of claims supported by AI-generated images and manipulated documents".
LessonEasy photo editing means a photo is weaker evidence on its own.
Credit
Allianz, Admiral and Aviva, each for its own figures
Watch for
Screening for made-up images
SAS says a new tool lets insurers "quickly detect synthetically generated or manipulated images before they are used in claims decisions". It names no insurer and gives no accuracy figures.
Achieved whena named insurer or an outside test reports how many altered images are caught and how many genuine ones are wrongly flagged.
Claimed by
SAS, May 2026
Record fraud found by AI
Zurich UK says its "connected AI and advanced analytics identified 860 fraudulent claims worth £17.9 million", part of a record £109.3 million of fraud it found.
Achieved whenan outside party, such as the Insurance Fraud Bureau or a court, confirms those cases.
Claimed by
Zurich UK, November 2026
Open
01
Fraud flags tested for fairness
Open-source code and made-up claims to test whether a claim-fraud screen sends one group of customers to extra checks more often than another. The test data carry planted gaps of known size in three outcomes: referral to investigators, requests for more documents, and time to payment. These are the kinds of outcome alleged in Huskey v. State Farm, and the planted gaps let anyone confirm the code finds them.
The briefCredit · unclaimed02
Cloned voices put to test
A mock phone claims line, built and run by the team, that tests whether speaker-verification systems, which check a caller by their voice, accept cloned voices. Volunteers who consent record their voices, and the team clones them. The cloned and genuine calls go through a phone-line audio channel to the mock line, and the test counts how often each system accepts a clone. No real insurer's line is called.
The briefCredit · unclaimed
Open 01 · the brief
Fraud flags tested for fairness
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Open-source code and made-up claims to test whether a claim-fraud screen sends one group of customers to extra checks more often than another. The test data carry planted gaps of known size in three outcomes: referral to investigators, requests for more documents, and time to payment. These are the kinds of outcome alleged in Huskey v. State Farm, and the planted gaps let anyone confirm the code finds them.
Why it is still open
The nearest public benchmark, Feedzai's Bank Account Fraud dataset, plants bias in made-up data but covers opening bank accounts, not insurance claims. The small insurance projects found on GitHub measure flag rates by group but do not plant group gaps to test against.
Built with what exists
Python; Fairlearn for the group measures; an open-source race-estimation method such as BIFSG with US Census name and geography tables; a script that generates made-up claims.
First step
Write the three outcome measures as a one-page specification and generate 10,000 made-up claims with one planted gap.
Done looks like
On 100,000 made-up claims with planted gaps of 2%, 5% and 10% in each outcome, the code flags the 5% and 10% gaps in at least 95 of 100 runs, and the 2% gap in no more than 5 of 100. Code and data are published under an open license.
Open 02 · the brief
Cloned voices put to test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A mock phone claims line, built and run by the team, that tests whether speaker-verification systems, which check a caller by their voice, accept cloned voices. Volunteers who consent record their voices, and the team clones them. The cloned and genuine calls go through a phone-line audio channel to the mock line, and the test counts how often each system accepts a clone. No real insurer's line is called.
Why it is still open
The ASVspoof 5 challenge tests spoofed speech in general, with no insurance or call-centre set. A 2025 study found that playing audio through a phone channel changes how well fakes are detected. Pindrop, which sells call security, said synthetic voice attacks on insurers "rose by 475 per cent" in 2024, from more than 1.2 billion calls it analysed.
Built with what exists
Asterisk, an open-source phone system; SpeechBrain's open speaker-verification models; a voice-cloning service such as ElevenLabs, used only on volunteers who have consented and within its terms; the ASVspoof 5 data as a baseline.
First step
Set up Asterisk with one test line and record five genuine calls from one volunteer.
Done looks like
50 consenting volunteers, with 10 cloned and 10 genuine calls each, run against at least two verification systems. The share of cloned calls accepted at each system's default setting is published with the code.