AI-Generated Insurance Fraud: Statistics, Cases and Evidence

The findingHow common is AI-generated insurance fraud?

No reliable industry-wide figure is currently available. Across the 18 markets reviewed for this page, we found no published dataset measuring what share of insurance claim photographs are AI-generated or digitally manipulated, from any insurer, regulator, industry association or third-party researcher.

That is a statement about what is published, not about what exists. Individual insurers may hold internal measurements. What we can report is that none had been made public in the sources reviewed.

This is not because the question is being ignored. Insurers across Europe and North America have published warnings, case examples and survey results. Several national associations describe manipulated evidence as a growing problem in their most recent annual reporting. The concern is real and it is documented.

What we could not find is the number. The figure that appears in most coverage of this subject, a percentage of claims affected, rests on material we were unable to inspect. We went looking for the measurement behind it.

The investigationWhich AI insurance fraud statistics could not be verified?

A figure claiming that 20 to 30% of insurance claims contain AI-altered media is widely repeated in industry material, but we could not locate a published dataset, sample size, measurement period or methodology supporting it.

We followed it backwards. It appears in a software vendor's blog post, which attributes it to a third party. It appears in a consultancy post, which attributes it to the same third party. Both name Shift Technology, and both quote the same phrasing. The original publication is no longer available.

What survives the trip is a sentence, not a measurement. The wording is that claims "may now include altered images, fabricated documents, or synthetic medical reports". There is no sample. No period. No definition of what counts as altered. No description of how anything was counted.

A statistic became evidence while the measurement behind it stopped being inspectable.

That is the finding, and it is worth stating carefully. It does not establish that the figure is wrong, and it is not an accusation against anyone who has published or quoted it. Citation drift of this kind is ordinary. What it does establish is that a number repeated in industry discussion cannot presently be checked against the measurement it came from.

Three further figures behave in similar ways.

A claim that 25% of claims are manipulated, attributed to the vendor Attestiv, does not appear in that company's published material. What its own material does cite is a different and much broader industry figure, that as many as 1 in 10 property and casualty claims may be fraudulent, which is not a statement about images at all.

A 300% rise, frequently described as an increase in deepfake claims, traces to a real and attributable publication: Allianz UK, in May 2024. But the underlying statistic counts cases involving consumer photo editing apps used on genuine photographs, videos and documents, comparing 2021/22 with 2022/23. Editing apps and deepfakes are materially different techniques. The figure is real; the version now in circulation describes something its source did not measure.

And a figure that appears nowhere at all: the share of claim photographs with removed or altered metadata. No insurer, association, regulator or vendor in any of the 18 markets has published one. This appears to be an open question rather than a contested one.

The complete citation audit, with what each chain was followed to and where it broke, is in the reference section below.

The evidenceHow many insurance claim photos have actually been measured as manipulated?

One substantial measurement exists in the public domain. A 2023 study of 768,000 claim images found 1,967 duplicates, or 0.26% of the sample, linked to 1,475 claims carrying approximately $5.3 million in indemnity payments. One photograph appeared in 44 different claims.

That is the strongest observed number in this field, and it needs two qualifications attached to it every time it is quoted.

The first is what it measures. It counts image duplication: the same image appearing in more than one claim. It does not measure AI generation. It does not measure editing. It does not measure manipulation of any other kind, and it is not a fraud rate. Those things remain unmeasured.

The second is who published it. The study was released by Verisk, a commercial provider of claims data and image forensics services, in a press release. The sample was drawn from a cross-carrier claims database. The selection method was not published, and the measurement has not been independently replicated. It is the best available evidence, and it carries the same inspection problem this page has just described in others, which is why it is labelled rather than leaned on.

The strongest observed number measures something far narrower than the most widely repeated one.

These two figures are frequently encountered by the same reader, and they are not two estimates of one quantity. Setting them out with their full labels is the clearest way to prevent that mistake.

The first figure is not a manipulation rate, a fraud rate or an AI rate, and it is not the "real" version of the second. It counts how often the same image appeared in more than one claim, in one dataset, in one market, in one year.

The contextWhy did insurance become dependent on customer-supplied photographs?

Because the productivity difference was large enough to restructure claims operations around it. A field appraiser completes roughly three to four inspections a day. The same work done from submitted photographs runs to fifteen or twenty. The change was made on operational grounds, well before manipulated digital evidence was a live concern.

Allstate documented the transition unusually clearly, because it happened to fall inside an earnings call. By 2017 the company was assessing approximately half of all drivable vehicles from customer photographs. Handling time fell from five to seven days to twenty four hours. More than 500 field adjuster roles were replaced, and $52 million of restructuring expense was recorded, attributed largely to the rollout.

It was not an isolated decision. By 2025, photo estimating accounted for 26.4% of repairable motor claims in the United States. In a 2026 study of 5,093 property claimants, 49% had submitted their photographs through a digital channel.

The operational gains were substantial. But something else changed at the same time, and it was never recorded as a change, because it had never been recorded as a function.

Physical presence established certain facts that a submitted image does not establish by itself.

An adjuster standing beside a damaged vehicle observed, as a by-product of being there, that this damage was present, on this vehicle, at this location, at that moment. It was rarely described as a verification control, because it was not the purpose of the visit.

This is a narrow observation and it is worth keeping narrow. Physical inspection does not establish everything about a claim, and remote handling is not without controls: insurers apply triage rules, indicator sets and forensic tooling to submitted images. The point is only that those four facts came automatically with a visit, and do not come automatically with a file.

We found no published assessment, from any insurer in the 18 markets reviewed, of what that trade-off cost. The full evidence on the shift is in the reference section.

The realityWhat does manipulated insurance claim evidence actually look like?

In the cases published by insurers and associations between 2023 and 2026 that we located, the recurring techniques are altered dates and locations, photographs taken from the internet or social media, substituted licence plates, and damage added to genuine images. Fully synthetic imagery appears among them, but it is one category rather than the whole.

Case evidence describes what happens. It does not establish how often it happens, and the cases below should not be read as a prevalence measure.

Zurich UK's head of claims fraud described the method that best illustrates the point. Fraudsters locate written-off vehicles on salvage websites, then implant a different registration number onto the image. Nothing is generated. The photograph is real, the damage is real, and the car belongs to someone else.

Sweden's industry review found something similar at scale. In 2022, 5% of all rejected claims involved manipulated attachments: receipts, and images "where the date and place had been changed". Again, the images were genuine photographs of genuine damage. What had been altered was their history.

Allianz UK published an example in 2024 of an edited photograph of a van, taken from social media, submitted alongside an invoice for over £1,000.

Fully generated images do appear, and when they fail they fail conspicuously. A Norwegian insurer detected an AI-generated photograph of a pair of glasses in 2026. It had three arms. In Belgium, an AI-generated photograph of water damage contained an extra window. Germany's GDV reports fully generated smartphone photographs of vehicle damage, and AI-generated X-ray images submitted to a pet insurer.

Conspicuous failures are the ones that get published. What that implies about images that do not fail conspicuously is not something the available evidence settles.

Read together, the cases point somewhere broader than the heading of this page. The operational question raised by most of them is not only whether an image was generated by AI. It is what can reliably be established about where a piece of digital evidence came from and what it depicts. AI generation is one route into that question and currently the most discussed one; on the published evidence it is not the only one.

All the documented cases we located, with sources and dates, are in the reference section.

The gapWhat has the insurance industry not measured?

Six questions that sound basic remain unanswered in public. Several of them are straightforward to measure for any insurer that holds claim images.

• What percentage of claim photographs have missing or altered metadata? No published figure was located from any insurer, association, regulator or vendor in the 18 markets reviewed.
• What percentage are actually AI-generated? Surveys of perception exist. No published measurement was located.
• How many manipulated-image cases do major insurers detect, in absolute numbers? Percentage changes and individual examples are published. We located no underlying counts.
• What happens to claim frequency or claim value when the capture of an image can be checked? One vendor customer account exists, describing a 70% fall in claims at an extended warranty provider. We located no controlled comparison published by an insurer.
• What is the difference in inspection cost between physical and remote assessment, from insurer-published data? Not located, despite targeted searching. Nor were insurer-specific thresholds below which no inspection takes place.
• Has the 768,000-image duplicate study been replicated by a separate organisation? We located no replication. In the published record we reviewed, it stands alone.

These gaps show how little has been measured in public. They are not evidence that the underlying problem is small, and they are not evidence that no insurer has measured any of it internally.

Insurance has become steadily more dependent on digital evidence. What the industry has not yet published is a reliable measurement of how far that evidence can be relied on.

SummaryWhat this research found

1. No reliable industry-wide statistic was located for the share of insurance claim photographs that are AI-generated or digitally manipulated. Across the 18 markets reviewed we found no published dataset measuring this, from any insurer, regulator, industry association or third-party researcher. This describes the published record, not what may be held internally.
2. The most widely repeated figure, that 20 to 30% of claims contain AI-altered media, could not be traced to a published dataset or methodology. The trail ends at two consultancy and vendor blog posts attributing it to a third party whose original publication is no longer available.
3. What has been measured is smaller and more specific. In a study of 768,000 claim images published in 2023, 1,967 were duplicates, or 0.26%, linked to 1,475 claims carrying approximately $5.3 million in indemnity payments. This was published by a commercial detection provider in a press release; the methodology was not published and the measurement has not been independently replicated.
4. Insurance claims have become substantially dependent on customer-supplied photographs. Photo estimating accounted for 26.4% of repairable motor claims in the United States in 2025. In 2017 Allstate reported inspecting approximately half of all drivable vehicles this way, replacing more than 500 field adjuster roles.
5. Insurers report low confidence in detecting manipulated media, in their own words. In a March 2026 survey of 300 US claims professionals, 32% said they were very confident they could recognise a deepfake and 43% that they could assess media at scale. These are self-reported perceptions, not measured detection rates.
6. In the documented cases located, manipulation is more often mundane than cinematic. Cases published by insurers and associations in Sweden, the United Kingdom, Poland, Norway, Germany, Belgium and Canada involve altered dates and locations, images taken from the internet or social media, substituted licence plates and added damage, alongside fully generated images. Cases describe what happens; they do not establish how often it happens.
7. National fraud statistics measure fundamentally different things and should not be compared directly. Of 18 markets reviewed, 11 publish a countable figure and 10 publish enough methodology to be interpreted. A British figure counts the claimed value of refused claims; a Dutch figure counts payouts prevented; an Italian figure counts claims carrying a screening indicator.
8. The frequently cited "5 to 10% of claims are fraudulent" originates in interviews with claims adjusters conducted in the late 1980s. The Coalition Against Insurance Fraud's own 2022 report states that the body which produced the figure "did not specify exactly how they derived at this figure".

How to read this pageHow every figure here is classified

Each statistic on this page carries a label describing what kind of evidence it is. The labels are applied consistently and are the fastest way to judge whether a number can carry the weight being placed on it.

Three further terms recur throughout and are used in their technical sense, because national statistics differ precisely on this point.

ReferenceThe full citation audit

Each figure below was followed from the citing article to the named source, and that source to its own source, until either a primary publication or a break in the chain was reached. Where the chain broke, the break is described rather than filled.

How claims moved onto customer photographs

The figures below come from different publishers using different populations and do not form a single time series. Together they describe a shift that took place across roughly a decade, on operational grounds.

ReferenceEvery documented case we located

Cases published by an insurer, industry association or regulator, or reported in national media citing one, between 2023 and 2026. Individual cases carry no statistical weight and this is not a complete census.

What do insurers say they are seeing?

Consistently, that they are encountering manipulated documentation and that they are not confident in their ability to detect it at scale. The largest recent measurement of this is a March 2026 survey of 300 US claims professionals and 1,000 US consumers. Every figure in this section is self-reported perception. None of it establishes how often manipulation actually occurs.

One observation from the 2016 UK Insurance Fraud Taskforce remains the most useful framing of why photographic evidence is particularly exposed.

Consumers generally find exaggeration of a genuine claim to be more morally acceptable than out-and-out fabrication of loss. In some cases consumers may not even realise they are exaggerating a claim.HM Treasury, Insurance Fraud Taskforce final report, January 2016, paragraph 2.92

A photograph is the natural medium for exaggeration rather than invention. Adding a scratch to a real vehicle does not require staging a collision, and the person doing it may not classify the act as fraud.

What are the insurance fraud statistics by country?

Eleven of the 18 markets reviewed publish a figure that can be checked against a stated unit of counting. The important column is not the amount but what one unit represents, because the units differ fundamentally. Seven markets publish nothing currently usable.

Markets where no usable figure currently exists

Seven of the 18 markets publish nothing that can be used. Each entry below is the outcome of a specific search rather than an assumption, and in several cases the absence is confirmed by the body itself.

Question 08Why can insurance fraud statistics not be compared between countries?

Because the published figures count different objects. Two national numbers can both be arithmetically correct and still be inappropriate to place side by side. The differences fall into three groups: what evidentiary threshold was met, whether the money figure is exposure or avoided payout, and whether fraud at policy inception is included.

Three concrete pairs illustrate the problem better than a general statement.

Terminology: translation is not enough

The terms used in national statistics do not map onto one another, and rendering them all as "fraud" in English removes exactly the distinction that matters.

What is the difference between suspected, investigated and proven fraud?

Roughly four orders of magnitude, in the one market that publishes every stage. Italy's regulator reports 613,728 motor liability claims carrying a fraud risk indicator in 2024, 292,170 taken forward for investigation, 2,938 criminal complaints, and 29 convictions.

This is not a criticism of Italian practice. A screening indicator is designed to over-select, and a system that flagged only provable fraud would be failing at its job. The point is narrower and applies everywhere: a single document contains a number that reads like a fraud rate and a number that counts fraud convictions, and they differ by a factor of roughly twenty thousand. Which one a presentation quotes determines the argument it appears to support.

The same problem inside one country

Two Dutch figures published for the same year, 2022, show the effect without any cross-border comparison. Fraud was proven in 1 in 1,042 claims. And 34% of the investigations actually opened proved fraud, described by the association as the highest rate in seven years. The first describes how rare proven fraud is across all claims; the second describes how well files are selected once picked up. A critic quoting only the first and an advocate quoting only the second can both be accurate.

Question 10Where does the "5 to 10% of claims are fraudulent" statistic come from?

First: what percentage of insurance claims are fraudulent?

No single international percentage can responsibly answer this. The published national figures count different things: proven cases in the Netherlands work out at roughly 1 claim in 1,042 for 2022, while Germany classifies around 10% of claims as worth investigating and Italy flags 24.4% of motor liability claims with a risk indicator. These are three different measurements, not three estimates of one number.

From interviews with claims adjusters conducted in the United States in the late 1980s. The figure was applied to national loss totals in 1995, was not updated or adjusted for inflation for nearly three decades, and subsequently entered European industry publications without an independent method being established.

1. Late 1980s · OriginInsurance Information Institute

   The III interviews claims adjusters, who put fraud at approximately 10% of property and casualty losses. The Coalition Against Insurance Fraud's own 2022 report states that the III "did not specify exactly how they derived at this figure".

2. 1995 · AppliedCoalition Against Insurance Fraud

   An $80bn national figure is built on that percentage. The Coalition states the number was then "never updated or adjusted for inflation" until its 2022 revision.

3. 2013 and 2024 · Adopted in EuropeInsurance Europe

   Uses "up to 10% of all claims expenditure" without publishing a method of its own. A frequently cited European range of €8bn to €12bn traces to a survey in which almost 40% of respondents believed that 5 to 10% of paid claims were fraudulent.

4. 2017 to 2026 · LocalisedNational associations

   GDV applies 10% as a suspicion rate drawn from member expert estimates. Assuralia uses "5 to 10% of payouts". The Swiss figure traces to a 2017 publication whose own wording is "in der Branche wird geschätzt". The Austrian range of 7 to 11% is attributable to one named official, and the VVO confirms it publishes no figures of its own.

5. Counter-argumentDerrig, as quoted in Feinman's critique

   Derrig is quoted as putting the ratio of suspected to provable fraud at approximately 25 to 1, which would imply provable fraud below 0.5% rather than 10%. We obtained this through the critique rather than Derrig's original publication, and it is presented here as a counter-argument in the literature rather than as an established measurement.

MethodologyHow this research was conducted

Scope. Eighteen markets were reviewed: the Netherlands, United Kingdom, Germany, France, Belgium, Italy, Spain, Poland, Austria, Switzerland, Sweden, Norway, Denmark, Finland, the United States, Canada, Australia and New Zealand. The Middle East, Asia, Latin America and Africa were not covered, and no statement on this page should be read as global.

Source selection. Priority was given to primary publications: association factsheets, regulator reports, insurer disclosures, government reviews and peer-reviewed literature. Ninety-nine source PDFs were downloaded and read in full. Where a figure appeared only in secondary coverage, that is stated at the point of use. Where a primary source was unreachable, for example because a server returned an error or required membership, that is also stated.

Citation tracing. For statistics in wide circulation, the citing article was followed to the named source, and that source to its own source, until either a primary publication or a break in the chain was reached. Where the chain broke, the figure was classified as unverified and the break was described rather than filled.

Classification. Every figure carries one of six labels: measured, reported, survey, vendor, estimate or unverified, defined at the top of this page. Where a measurement was published by a party that also sells related services, both facts are stated. Figures we calculated from published counts are labelled as derived at the point of use.

What was excluded. Figures in circulation without an attributable source were not listed, on the basis that recording an anonymous claim tells a reader nothing. Figures from the academic literature that could not be checked against the original publication within this research were left out rather than reproduced from secondary summaries. No figure on this page was estimated, rounded for effect, or reconstructed where the source was unavailable.

An editorial choice worth stating. The phrase "insurance fraud is rising" does not appear as a finding. Detected fraud rose in most European markets in the most recent reporting year, and detection effort also rose. No source in this research separates the two, and both the British and Dutch decade-long series show that a single year's direction can reverse. Similarly, no survey result on this page is presented as a prevalence rate.

Cut-off. Research was completed on 6 September 2026. Several bodies publish annually, so figures for 2025 were still pending in some markets at that date; where that is the case it is noted in the relevant table.

Corrections. Corrections are welcome, particularly from anyone holding a primary source for a figure classified here as unverified.

ReferencesPrimary sources

Grouped by publisher. Figures used from each are cited at the point of use throughout the page.