Verisk's 2026 State of Insurance Fraud Study, published March 17, surveyed 1,000 US consumers and 300 claims professionals. Its central finding is that 36% of consumers would consider digitally altering a claim image or document, rising to 55% among Gen Z and falling to 12% among baby boomers. On the other side, 98% of insurers say AI editing tools are fuelling digital fraud and 32% are very confident they could identify a deepfake.

Key Takeaways

  • 36% of consumers would consider altering a claim image, with a 43-point spread from 55% of Gen Z to 12% of baby boomers.
  • 66% of insurers believe digital media fraud goes undetected often or very often, while nearly every respondent says they have already encountered manipulated documentation.
  • A 26-point detection confidence gap: 58% are very confident detecting edits to real photos, 32% for deepfakes, and 43% for assessing authenticity at scale.
  • 15% of consumers consider exaggerating damage acceptable and 13% consider fabricating damage that never occurred acceptable, against 52% for brightness adjustment.
  • Human detection of high-quality video deepfakes runs at 24.5%, with only 0.1% of participants correctly identifying every real and fake item shown.

What the Survey Measured

The generational spread is the finding with the longest reach. Willingness runs 55% for Gen Z, 49% for millennials, 28% for Gen X, and 12% for boomers, against 36% across all consumers. That is not a cultural observation; it is a predictable shift in the insured population's fraud propensity as older cohorts leave active policy years and younger ones enter their peak claims decades for auto, renters, and homeowners.

Social proof tracks the same gradient. 64% of Gen Z respondents report knowing someone who has used AI tools for financial gain including insurance claims, against 41% of all consumers, and 62% of consumers believe people already manipulate claim documents often or very often. Perceived normality, not capability, is what moves a behaviour from hypothetical to ambient.

The acceptability question separates the categories that matter. 52% consider brightness or contrast adjustment acceptable and 49% consider cropping out background elements acceptable, both arguably presentational. But 15% consider exaggerating damage acceptable and 13% consider creating images of damage that never occurred acceptable. Those last two are deliberate misrepresentation, and 44% of consumers who have used AI editing tools described the results as very realistic.

Verisk's Shane Riedman described the tools as arming the average Gen Zer to commit insurance fraud quickly, easily, and in their own minds somewhat innocently. That word marks a category the fraud taxonomy does not have. Hard fraud is a staged accident. Soft fraud is inflating a genuine claim.

What the study describes is digitally augmented soft fraud: a real claim with real damage, enhanced with modified evidence. It is harder to detect than hard fraud because the underlying loss is genuine, harder to prove than soft fraud because the edits may be invisible to a reviewer, and harder to deter than either because the person doing it does not think it is a crime.

The Contamination Is Already Inside the Triangle

Chain-ladder, Bornhuetter-Ferguson, and frequency-severity methods all project historical development patterns forward, and those patterns embed whatever fraud existed in the experience period. If the long-standing industry estimate of roughly 10% of claims carrying some misrepresentation held stable, the development factors carried that 10% forward and the method worked.

The 66% of insurers who believe digital media fraud goes undetected often is what breaks that. An AI-manipulated claim that is never flagged settles at an inflated amount and enters the data as a legitimate paid loss. It inflates link ratios and age-to-ultimate factors without appearing anywhere as fraud, which means the historical triangle already contains an unmeasured quantum of it and the method faithfully projects that contamination forward.

The direction of the error is knowable even where the magnitude is not.

Scenario Fraud Rate Avg. Severity Uplift Incremental Pure Premium per Exposure
Historical baseline 10% $2,500 $250
Moderate shift (2028) 13% $2,800 $364
Full generational shift (2031) 15% $3,100 $465

On this illustration the incremental pure premium attributable to manipulation moves from $250 to $465 per exposure, an 86% increase, on a fraud rate moving from 10% to 15% and average severity uplift from $2,500 to $3,100. The scenario is a projection rather than an observation, and the survey measures stated willingness rather than confirmed behaviour, so the levels should be treated as a bracket. What the arithmetic establishes is that a rate of change in the fraud parameter produces a pure premium effect large enough to matter at the filing level.

The rate filing treatment is where this bites. A fraud and abuse provision has historically been treated as roughly stationary: fraud exists, it costs a known share of premium, and the load carries forward with modest adjustment. Two features of the Verisk data break stationarity. Tool accessibility moved as a step function rather than a trend, from requiring Photoshop skill to requiring a free app in about two years. And the detection gap widens as generation improves, because the share of manipulations falling into the harder-to-detect category grows while the 32% deepfake confidence figure stays where it is.

Deterrence does not close it either. 69% of consumers believe fraud raises premiums for everyone, and that awareness coexists with the 36% willingness figure in the same survey population. The premium externality is understood and it is not changing behaviour, which removes the argument that rational self-interest holds fraud at a stable equilibrium.

Detection Investment Arrives Two to Four Accident Years Late

Carriers are moving. About 65% use third-party AI detection tools and 50% use internally developed ones, so many are running both. Verisk reported a sixth top-10 carrier signing on to its Digital Media Forensics platform on its Q1 2026 call, and vendors including Shift Technology, FRISS, and Reality Defender occupy adjacent parts of the stack.

The problem is that the human layer underneath is already saturated. iProov's 2025 threat intelligence report put human detection of high-quality video deepfakes at 24.5%, with only 0.1% of participants correctly identifying every real and fake item shown. An adjuster handling 15 to 25 claims a day is not a control against this, and the Verisk figures agree: 43% are very confident assessing authenticity at scale, against 58% on a single edited photo. Escalating one suspicious image to a special investigations unit works; screening every inbound image across millions of annual submissions requires tooling most carriers have not fully deployed.

The timing is what makes this a reserving problem rather than an operations problem. Detection investment made today reduces paid fraud losses in future periods, but the lag from deployment through training, calibration, and measurable loss reduction spans roughly two to four accident years. Through that window, fraud frequency is rising and the countermeasure is not yet effective, so the current-period reserve is understated on both sides at once.

Swiss Re's SONAR report projected deepfake-related incidents rising more than 160% on automated bot networks and improving generation quality, and a 2024 survey it cited found 92% of companies had already experienced financial losses from deepfake incidents with 10% reporting damages above $1 million. Those figures describe the environment before the consumer editing tools in the Verisk study reached general availability. The detection stack being funded now is calibrated against the manipulations that exist today, and the calibration period is long enough that the tools will meet a different problem than the one they were built for.

Further Reading