From analyzing insurer fraud detection capabilities across a dozen carrier earnings calls this year, the gap between AI-enabled fraud and AI-powered detection is widening faster than most reserve models reflect. Verisk’s 2026 State of Insurance Fraud Study, published March 17, 2026, puts hard numbers on that gap for the first time. The study surveyed 1,000 U.S. consumers and 300 insurance claims professionals between December 2025 and January 2026, producing data that should concern every actuary responsible for reserve adequacy opinions or rate filing support.

The headline finding: 36% of all consumers would consider digitally altering a claim image or document to strengthen their case. That figure rises to 55% among Gen Z respondents, 49% among millennials, 28% among Gen X, and just 12% among baby boomers. On the detection side, 98% of insurers acknowledge that AI editing tools fuel digital fraud, yet only 32% feel very confident they can identify a deepfake. Two-thirds of insurers (66%) believe digital media fraud goes undetected “often or very often” across the industry. This is not a technology problem waiting for a technology solution. It is a structural shift in policyholder behavior that requires actuarial repricing.

The Verisk Data in Full: A Behavioral Map of AI-Enabled Claims Manipulation

The Verisk study’s consumer survey probed both willingness to manipulate and perceptions of realism. Among consumers who have used AI editing tools, 44% described the results as “very realistic,” a critical data point for actuaries because realistic edits are harder to detect and therefore more likely to survive claims adjudication. The study also found that 62% of all consumers believe people already use AI tools to manipulate claim documents “often or very often,” suggesting that digital fabrication is perceived as normalized, not exceptional.

The generational breakdown is the most actuarially significant finding in the study. The 43-point spread between Gen Z (55%) and boomers (12%) is not merely a cultural curiosity; it represents a predictable shift in the insured population’s fraud propensity as older cohorts age out of active policy years and younger cohorts enter their peak claims-generating decades for auto, renters, and homeowners insurance.

Generation Would Alter a Claim Knows Someone Who Has
Gen Z 55% 64%
Millennials 49% 54%
Gen X 28% Not reported separately
Baby Boomers 12% Not reported separately
All consumers 36% 41%

The “knows someone who has” column matters as much as the willingness column. Research on insurance fraud behavior consistently shows that social normalization is a leading indicator of actual claims manipulation. When 64% of Gen Z respondents report knowing someone who has used AI tools for financial gain, including insurance claims, the behavior has moved from hypothetical to ambient.

Verisk also asked consumers about the perceived acceptability of specific edit types. Just over half (52%) said adjusting brightness or contrast to make damage more visible is acceptable. Nearly half (49%) considered cropping out unrelated background elements to be acceptable. Smaller but non-trivial shares said exaggerating damage (15%) and creating images of damage that never occurred (13%) are acceptable. The 15% figure for damage exaggeration is the most relevant for frequency and severity modeling: these are not gray-area adjustments but deliberate misrepresentation that inflates claim costs.

The Detection Confidence Gap: Insurers Know the Problem and Cannot Solve It

The study’s claims professional survey reveals an industry that understands the threat but lacks the tools to match it. Nearly all insurers (98%) agree that AI-powered editing tools are fueling an increase in digital fraud. Ninety-nine percent said they have encountered manipulated or AI-altered documentation. Three-quarters (76%) reported that AI-altered claim submissions have become “more sophisticated” over the past year.

Yet detection confidence is strikingly uneven. While 58% of insurers feel very confident detecting edits made to real photos or videos (brightness adjustments, cloning artifacts, metadata inconsistencies), that confidence drops sharply for more advanced manipulations. Only 32% of insurers are very confident they could identify a deepfake. Fewer than half (43%) feel very confident assessing the authenticity of digital media at scale. The scale qualifier is critical: a claims handler reviewing a single suspicious image can escalate to a special investigations unit, but systematic screening of all inbound claim media across millions of annual submissions requires automated tooling that most carriers have not fully deployed.

Shane Riedman, president of Anti-Fraud Analytics at Verisk, captured the dynamic plainly: “What we’re seeing is that the wide proliferation of AI tools, and the ready availability of those tools, is arming the average Gen Zer to be able to very quickly and easily, and in their minds, somewhat innocently, commit insurance fraud.”

The word “innocently” in Riedman’s quote points to the actuarial crux. Traditional fraud models distinguish between hard fraud (staged accidents, arson, phantom claims) and soft fraud (inflating legitimate claims, adding undamaged items to a theft report). AI-enabled manipulation creates a third category: digitally augmented soft fraud, where a real claim with real damage is enhanced with AI-generated or AI-modified evidence to increase the payout. This category is harder to detect than hard fraud (the underlying claim is genuine), harder to prove than traditional soft fraud (the edits may be invisible to human reviewers), and harder to deter than either (the perpetrator views the edits as a minor enhancement, not a crime).

The Arms Race: Third-Party Tools, In-House Builds, and the Evolving Detection Stack

Insurers are not standing still. About 65% of insurers in the Verisk survey use automated AI-based detection tools from third-party vendors. Half (50%) use internally developed AI tools. These numbers indicate that many carriers are running dual detection stacks, combining vendor solutions with proprietary models trained on their own claims data.

Verisk itself has expanded its Digital Media Forensics product line, and the company reported on its Q1 2026 earnings call that a sixth top-10 carrier had signed on to the platform. Other vendors in the detection ecosystem include Shift Technology, whose Claims Fraud Detection product reports detection rates 3x higher than rules-based systems; FRISS, which integrates Verisk’s ClaimSearch database of more than 1.7 billion P&C claims; and Reality Defender, which focuses specifically on deepfake and synthetic media detection.

Swiss Re’s SONAR 2025 emerging risk report flagged deepfakes and AI-generated disinformation as a growing insurance threat, projecting that deepfake-related incidents would rise more than 160% driven by automated bot networks and increasingly realistic image and video generation. A 2024 industry survey cited by Swiss Re found that 92% of companies had experienced financial losses from deepfake-related incidents, with 10% reporting damages exceeding $1 million.

Detection Metric Insurer Confidence / Adoption
Very confident detecting edits to real photos 58%
Very confident identifying deepfakes 32%
Very confident assessing authenticity at scale 43%
Use third-party AI detection tools 65%
Use in-house AI detection tools 50%
Believe digital fraud goes undetected often 66%

The detection challenge is compounded by the pace of offensive tool development. Consumer-grade AI editing tools that produce photorealistic results are now available as free mobile applications. The iProov 2025 Biometric Threat Intelligence report found that human detection rates for high-quality video deepfakes are just 24.5%, and only 0.1% of study participants correctly identified all fake and real media shown to them. For claims adjusters handling 15 to 25 claims per day, the expectation that human review can catch AI-manipulated evidence is increasingly unrealistic.

Actuarial Reserving Implications: Why Current IBNR Models Undercount AI-Enabled Fraud

The Verisk data exposes a structural problem in how actuaries estimate fraud-related reserve components. Standard reserving methodologies, whether chain-ladder, Bornhuetter-Ferguson, or frequency-severity models, rely on historical loss development patterns. Those patterns embed whatever level of fraud existed in the experience period. If the fraud rate was stable at, say, 10% of claims containing some element of misrepresentation (the long-standing industry estimate from the Coalition Against Insurance Fraud), then historical development factors implicitly carry that 10% forward.

The Verisk study suggests that assumption is no longer safe. The generational willingness data implies a steadily increasing fraud propensity as the insured population shifts younger. If 55% of Gen Z would consider altering a claim (versus 12% of boomers), and if Gen Z and millennials represent a growing share of new auto, renters, and homeowners policies, then the fraud load embedded in historical experience understates the fraud load in the current and prospective periods.

Consider a simplified illustrative example. Suppose a personal auto book has a historical base rate of fraudulent or inflated claims at 10% of all claims, with an average severity uplift of $2,500 per fraudulent claim. If the generational shift raises the effective manipulation rate from 10% to 15% over the next five years (a conservative extrapolation given that the Verisk survey measures willingness, not confirmed behavior), the pure premium impact is material:

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

In this scenario, the incremental pure premium attributable to AI-enabled fraud rises from $250 to $465 per exposure over a six-year horizon, an 86% increase. Because fraud-related severity inflation shows up in loss development triangles as higher-than-expected case reserve development on open claims, it distorts IBNR estimates in a direction that standard actuarial methods will not correct until several accident years of elevated experience have matured.

The IBNR blind spot is compounded by the detection confidence gap. If 66% of insurers believe digital media fraud goes undetected often, then a meaningful share of AI-manipulated claims are settling at inflated amounts without ever being flagged. These overpayments enter the loss experience as legitimate paid losses, inflating link ratios and age-to-ultimate factors in ways that are invisible to the reserving actuary. The result is that historical development patterns already contain an unknown quantum of AI-enabled fraud, and the chain-ladder method faithfully projects that unknown contamination forward.

Rate Filing Implications: Should Carriers Load for AI-Enabled Moral Hazard?

From tracking recent P&C rate filings and regulatory discussions, the question of whether to include an explicit AI-fraud load in rate indications is moving from theoretical to operational. Traditional actuarial rate indications include a fraud and abuse provision, typically derived from industry studies (the Coalition Against Insurance Fraud’s $308.6 billion annual estimate is the most commonly cited) or from the carrier’s own SIU referral and recovery data. That provision has been treated as roughly static: fraud exists, it costs the industry a known percentage of premium, and the load is carried forward with modest annual adjustments.

The Verisk data challenges that stationarity assumption. If a measurable and growing share of the insured population views digital fabrication as an acceptable claims strategy, the fraud load is not static; it is a trend variable that requires explicit projection, similar to how loss cost trends capture medical inflation or building material cost increases.

Several structural factors support the case for a distinct AI-enabled fraud frequency driver in rate filings:

Tool accessibility is a step function, not a gradual trend. Unlike traditional soft fraud techniques that require some knowledge of the claims process, AI editing tools require only a smartphone. The barrier to manipulating a claim photo dropped from “requires Photoshop skill” to “requires downloading a free app” in roughly two years. This step-function change in accessibility means that historical fraud frequency trends may significantly understate the prospective fraud rate.

Detection lags creation by a structural margin. The Verisk data shows a 26-point confidence gap between detecting real-photo edits (58% very confident) and identifying deepfakes (32% very confident). As AI generation tools improve, the share of claim manipulations that fall into the harder-to-detect category will grow, meaning the effective detection rate is declining even as carriers invest more in countermeasures.

Generational cohort effects are predictable and durable. Unlike cyclical economic factors that may increase fraud during recessions, the generational attitudes documented by Verisk reflect cultural norms around digital manipulation that are unlikely to reverse. A 22-year-old who views AI-enhanced claim photos as a minor optimization will carry that attitude into their 30s and 40s, when they hold more valuable policies with higher claim severity potential.

Consumer awareness of fraud’s premium impact does not deter behavior. The Verisk study found that 69% of consumers believe fraud raises premiums for all policyholders. Yet this awareness coexists with 36% willingness to manipulate claims. The deterrence effect of premium externality knowledge is weak, which undermines the economic logic that rational policyholder self-interest limits fraud to a stable equilibrium.

The Detection Arms Race: Carrier Investment Patterns and the Economics of Countermeasures

From reviewing insurer technology disclosures across Q1 2026 earnings calls, fraud detection investment is accelerating but remains unevenly distributed. Travelers disclosed deploying Anthropic-powered AI assistants to 10,000 staff, with claims fraud detection as one application area. AIG reported an 88% AI-adjuster fraud agreement rate on its Q1 2026 call, the first carrier to disclose a quantitative AI governance metric of this kind. Chubb has framed its global claims AI mandate under new Global Claims Officer Kevin Rampe as targeting 85% automation across 54 countries.

Verisk’s own investment trajectory is instructive. The company’s Q1 2026 earnings showed seven new AI modules shipped in a single quarter as part of its Core Lines Reimagine initiative, with Digital Media Forensics gaining a sixth top-10 carrier customer. The forensics module uses computer vision models trained to detect artifacts specific to AI-generated imagery, including inconsistencies in lighting angles, compression patterns, and metadata that consumer-grade editing tools cannot perfectly replicate.

The economics of detection investment create their own actuarial signal. If carriers are spending materially more on fraud detection technology, those costs will eventually need to appear in expense ratios or be offset by reduced loss payments. The NICB projects that identity theft-related insurance fraud alone will rise 49% by 2025, and the FBI estimates that fraud costs the average U.S. family between $4,000 and $7,000 in increased premiums over a decade. These figures predate the AI-editing tool proliferation that the Verisk study documents.

For the reserving actuary, the detection arms race creates a timing problem. Investment in detection tools today should reduce paid fraud losses in future periods, but the lag between deployment, training, calibration, and measurable loss reduction may span two to four accident years. During that lag, the combination of rising AI-enabled fraud frequency and not-yet-effective detection investment means that current-period reserves are likely understated.

ASOP Compliance Considerations: Documenting the Fraud Assumption

Actuaries preparing reserve opinions or rate indications that involve fraud-sensitive lines should consider how the Verisk data intersects with existing Actuarial Standards of Practice. ASOP No. 43 (Property/Casualty Unpaid Claim Estimates) requires actuaries to consider “significant influences that may affect the future development of claims,” which includes changes in claimant behavior and fraud patterns. ASOP No. 25 (Credibility Procedures) provides guidance on blending carrier-specific fraud experience with industry-wide data like the Verisk study.

ASOP No. 56 (Modeling) applies directly to actuaries using or validating AI-powered fraud detection models. The standard requires documentation of model limitations, data quality issues, and the sensitivity of results to key assumptions. Given that 66% of insurers in the Verisk survey believe fraud goes undetected often, any model that treats historical paid losses as a clean measure of true loss costs is making an implicit assumption about detection completeness that should be documented and stress-tested.

The practical question for appointed actuaries is whether the Verisk findings constitute a “known deficiency” in historical data that must be disclosed. If survey evidence shows that a significant share of the insured population views digital claim manipulation as acceptable, and carrier detection capabilities demonstrably lag the creation of manipulated media, the argument that fraud contamination in historical loss triangles is immaterial becomes harder to sustain.

The Social Normalization Feedback Loop

The most troubling finding in the Verisk study may be the social normalization data. When 41% of all consumers, and 64% of Gen Z, report knowing someone who has used AI tools to alter content for financial gain, the behavior has crossed from individual deviance to social convention. Research on moral hazard in insurance consistently shows that normalization within peer groups is the strongest predictor of escalating claims manipulation. The Verisk data suggests that for younger cohorts, AI-assisted claim manipulation is already peer-normalized.

This normalization creates a feedback loop that actuaries should model explicitly. As more claimants perceive AI manipulation as common and low-risk, more will attempt it. As more attempts enter the system, the base rate of manipulated claims rises. As the base rate rises, detection systems become saturated, reducing the per-claim detection probability. As detection probability falls, the perceived risk of manipulation decreases further, encouraging more attempts. This cycle does not reach a stable equilibrium under current detection investment levels; it reaches an equilibrium only when either detection capability catches up (requiring ongoing and accelerating investment) or when manipulation becomes so pervasive that insurers restructure the claims process itself.

Some carriers are already responding by shifting toward structured, app-based claims submission workflows that constrain what claimants can submit. Instead of accepting free-form photo uploads, these platforms capture images directly through the carrier’s app with embedded metadata and geolocation verification, real-time capture authentication, and tamper-evident image pipelines. This approach attacks the problem at the submission layer rather than the detection layer, reducing the opportunity for manipulation before it enters the claims system.

Why This Matters: The Five-Year Reserve and Pricing Outlook

The Verisk fraud study is not just a technology story or a cultural commentary. It is an actuarial data event that should influence reserve opinions, rate indications, and reinsurance treaty pricing across P&C personal lines. The specific implications for actuarial practice over the next five years include the following:

IBNR adequacy: Reserve estimates for current and recent accident years on fraud-sensitive lines (auto physical damage, homeowners contents, renters, small commercial property) should incorporate an explicit adjustment for the growing share of claims that contain undetected AI manipulation. The Verisk data provides a credible basis for that adjustment.

Pure premium trending: Frequency and severity trend selections for rate filings should consider AI-enabled fraud as a distinct driver, separate from traditional inflation-based severity trends and volume-driven frequency trends. The generational cohort data provides a demographic projection framework for estimating the pace of change.

Expense ratio implications: Carrier investment in AI fraud detection will increase loss adjustment expense ratios before reducing loss ratios. The lag between cost and benefit should be modeled in combined ratio projections, particularly for carriers in the early stages of detection deployment.

Reinsurance treaty pricing: Ceding companies that can demonstrate superior fraud detection capabilities may negotiate better terms on excess-of-loss treaties for fraud-sensitive lines. Conversely, cedents with weak detection infrastructure may face upward pressure on treaty pricing as reinsurers incorporate the Verisk findings into their own loss projections.

Regulatory scrutiny of fraud provisions: State regulators reviewing rate filings will eventually need to evaluate whether carriers’ fraud provisions adequately reflect the AI-enabled threat. The Verisk study provides regulators with independent, survey-based evidence to support (or challenge) carriers’ fraud load assumptions.

The industry’s $308.6 billion annual fraud cost estimate, maintained by the Coalition Against Insurance Fraud, has been widely cited for years. That figure was calculated before AI editing tools became consumer-grade and before a majority of younger policyholders expressed willingness to use them on insurance claims. Any actuarial assumption that treats the historical fraud cost baseline as stable is, at minimum, worth re-examining in light of the Verisk data.

Further Reading