Two studies published in the first half of 2026 paint a consistent picture of how generative AI is reshaping insurance fraud.

The first is Verisk’s State of Insurance Fraud study, based on surveys of 1,000 U.S. consumers and 300 insurance claims professionals. Its central finding is less about criminal fraud rings and more about a widening “ethics gap” among ordinary policyholders: 36% of consumers say they would at least somewhat consider digitally altering a claim image or document to strengthen their case, even knowing it would break insurer rules, and that number climbs to 55% among Generation Z respondents and 49% among Millennials, compared with just 28% of Generation X and 12% of Baby Boomers.

Verisk also found that 41% of consumers know someone who has used AI editing tools to alter a photo, video, or document for financial gain in some context, including insurance claims.

On the industry side, 98% of insurers agree AI-powered editing tools are driving a rise in digital media fraud and 99% say they’ve personally encountered manipulated or AI-altered documentation, yet confidence in detection lags well behind that awareness: just 32% of insurers say they’re very confident they could identify a deepfake, and only 43% feel very confident assessing the authenticity of digital media at scale. Two-thirds of insurers (66%) believe digital media fraud goes undetected often or very often industry-wide.

The second is the 2026 Anti-Fraud Technology Benchmarking Report, the fourth installment of a joint research series the Association of Certified Fraud Examiners and data-analytics firm SAS have run since 2019, based on a survey of 713 fraud fighters across eight world regions (not limited to insurance, but heavily represented by financial services and insurance professionals).

Its headline finding: just 7% of anti-fraud professionals say their organizations are more than moderately prepared to detect or prevent AI-fueled fraud. The same research series found, in a preview released for International Fraud Awareness Week in November 2025, that 77% of anti-fraud professionals had already seen an acceleration in deepfake-driven social engineering over the prior 24 months, and 83% expect that trend to continue accelerating over the next two years.

The 2026 report also flags a governance gap behind the detection gap: only 18% of organizations that use AI in fraud-fighting say they test those models for bias or fairness, and just 6% feel completely confident explaining how their own AI models reach their fraud decisions.

Separately, insurance-specific vendors have demonstrated the underlying mechanics driving these concerns. SAS’s own insurance fraud specialists have published public demonstrations showing how easily generative AI tools can fabricate a convincing vehicle crash scene or add plausible property damage to an ordinary photo in seconds, using tools accessible to anyone with a computer, and reinsurer Swiss Re’s 2025 SONAR emerging-risk report similarly flagged a rising, UK-documented increase in deepfake use in low-value claims fraud specifically.

None of these sources put a specific dollar figure or percentage breakdown on how much of current insurance fraud is AI-generated as opposed to conventional — and none of the credible sources reviewed for this report claim to. That, itself, may be the most useful finding for insurance professionals: the industry’s own major research bodies are documenting a fast-growing, poorly quantified threat and a real detection gap, rather than a fraud phenomenon anyone can point to with confident overall dollar figures yet.

What this means for insurers and claims organizations:

– – Close the detection-confidence gap before it becomes a liability gap. With only 7% of fraud fighters more than moderately prepared for AI-fueled fraud and just 32% of insurers confident they could spot a deepfake, detection capability is the most exposed weak point right now. Benchmark existing fraud-detection tools specifically against synthetic-document and deepfake scenarios, not just traditional image manipulation, and treat this as distinct from general fraud-analytics investment.
– – Govern the AI tools being deployed, not only the AI tools being defended against. With only 18% of organizations testing their own anti-fraud AI models for bias and just 6% able to confidently explain their models’ decisions, insurers face growing regulatory exposure on the deployment side as well as the detection side. Oklahoma’s own Bulletin 2024-11, aligned with the NAIC’s model AI bulletin, is one example of regulators formalizing expectations of fairness, accountability, and transparency in AI-supported claims decisions; model governance should be treated as a compliance requirement, not an afterthought.
– – Address the ethics gap at the point of submission. Verisk’s finding that roughly half of Gen Z and Millennial consumers would consider altering claim evidence suggests a meaningful share of this problem is casual policyholder behavior rather than organized fraud rings. That argues for consumer-facing friction at upload — clear rule disclosures, provenance and metadata checks on submitted images — rather than relying solely on forensic detection after the fact.
– – Invest in shared intelligence, not solo detection. Because the same AI-generated assets can be recycled across carriers, cross-industry data-sharing consortia and state fraud bureaus become more valuable as the cost of generating fraudulent evidence falls; no single insurer’s claims history is enough to catch a pattern designed to be reused.
– – Weigh vendor fraud statistics with some skepticism. Several point-solution vendors selling AI-detection products are themselves the source of the more alarming, less-sourced statistics circulating in this space. Procurement and budget decisions are best grounded in named, methodologically disclosed research — Verisk, ACFE/SAS, NICB, and similar bodies — rather than a vendor’s own unsourced claims about the scale of the problem.
– – Update SIU training for AI-specific tells. Traditional red flags — inconsistent metadata, repeat claimants, staged-looking photos — don’t reliably catch evidence that was specifically generated to pass casual visual review; investigator training should be refreshed to reflect that.