Admiral’s 2025 fraud report put a number on something we at VerifyPDF have been watching for two years: detected insurance fraud rose 71% in a single year, to £86.8 million from £50.9 million, with AI-manipulated evidence driving the jump. Not a gradual creep. A jump.
The distinction that matters here is AI-edited versus AI-generated. AI-generated fakes built from scratch are, counterintuitively, easier to catch.
What adjusters are actually seeing now are real photos and real documents with targeted edits: a repair cost nudged upward, a damage photo retouched to show worse conditions than existed, a police report with a few sentences quietly removed. These start from legitimate sources, which is exactly why they are so hard to spot under visual review.
The Verisk March 2026 study on insurance fraud confirmed what claims teams already suspected: 98% of insurers surveyed said AI editing tools were driving fraud growth in their pipeline. When 98% of an industry agrees on anything, you can stop calling it a fringe problem. Something fundamental changed about AI insurance fraud in the UK in 2025, and the processes for catching it have not kept up.
UK insurance fraud in 2025: what the 71% actually means
The Admiral figure refers to detected fraud, which is an important qualification. If detection rates are improving because insurers are getting better at finding fraud, a rising number does not necessarily mean the absolute problem is getting worse at the same rate. But that is not what is happening here.
According to Admiral’s 2025 fraud figures, the increase was driven almost entirely by opportunistic fraud enabled by accessible AI tools. These are not professional fraud rings working sophisticated schemes. They are ordinary policyholders who discovered that AI image editors can convincingly extend a crack in a wall, deepen water staining on a floor or add visible smoke damage to a photo. And that the person reviewing their claim cannot tell the difference.
The Association of British Insurers estimates that roughly as much fraud goes undetected each year as insurers actually catch. If that ratio holds and detected fraud jumped 71%, the real volume increase is at least as large as the headline suggests. The 71% rise is the floor, not the ceiling.
The other thing worth noting: as forensic detection improves, detected fraud numbers will rise even if absolute fraud volume stays flat. A better detection rate shows up as a higher fraud figure in annual reports. Insurers should not read that as failure. They should read it as evidence that their tools are working.
Why AI-edited evidence fools adjusters but not forensics
AI-generated documents fail forensic checks in predictable ways. There are no authentic metadata signatures, no editing history from a legitimate software tool, no proof of provenance. The document appears from nowhere, structurally.
But AI-edited documents are different. They start as real files.
Take a car repair invoice. A fraudster gets a genuine PDF from a body shop, opens it in a PDF editor, changes the labour cost from £850 to £2,400 and submits it to their insurer. The document’s metadata still shows it was created by the body shop’s accounting software. The fonts are correct because they were always correct. The layout is perfect because nothing structural changed. A human adjuster reviewing this document will see a routine invoice, nothing more. Would you have caught it? Be honest.
The edit lives in the content layer. And that is precisely where forensic tools look.
This is the same problem we covered when writing about AI-generated receipts fooling finance teams - the tools for producing convincing fake financial documents have become so accessible that there are tutorials on every platform showing users exactly how to do it. Insurance documents are not exempt from this trend. If anything, they are a more attractive target because claim payouts are larger and the submission process is designed to be frictionless.
The three edits that are slipping through right now
Based on what claims teams have shared with us and what we see in the documents processed through VerifyPDF, three categories account for the majority of AI-assisted insurance fraud in the UK right now.
-
Reshot receipts and retouched damage photos. A policyholder photographs genuinely damaged property, then uses an AI image editor to extend cracks, deepen water staining or exaggerate structural damage before submission. The metadata shows a real camera and a real date. But the pixel patterns at the edit boundaries tell a different story: inconsistent compression artifacts, lighting angles that do not match the rest of the frame, cloning patterns from the retouching tool where textures have been replicated.
-
Modified repair and replacement invoices. This is the most common route we encounter. A genuine invoice from a real supplier, with the cost figures edited upward before submission. Sometimes the fraudster generates a plausible-looking invoice from scratch using a template. As research into fake documents invisible to the human eye shows, around 90% of sophisticated forgeries pass visual inspection even by a trained reviewer. Invoice fraud sits squarely in that category.
-
Altered police and accident reports. Rarer but higher in value per claim. A claimant removes the section of a police report that assigns partial fault to them, turning a 50/50 liability split into a 100% third-party fault claim. The document is otherwise identical to the original. Manual review catches this only if the adjuster happens to cross-reference the reporting officer’s reference number against the police database, which most do not do as standard practice.
None of these edits require technical expertise. The tools are consumer-grade, widely available and cost less than ÂŁ20 per month. The knowledge required to use them is freely distributed online. This is why the volume numbers keep rising.
Why 98% of UK insurers say AI editing is the new normal
The Verisk finding is striking enough on its own. But the detail buried in that study is more troubling: most insurers that acknowledged the AI editing problem had not changed their core adjudication process. They knew AI-edited documents were circulating in their pipeline. They simply did not have a systematic way to detect them.
This is a process gap, not a technology gap. Forensic tools that detect pixel-level editing, metadata anomalies and compression artifacts have existed for years. The gap is that evidence documents are still reviewed visually, by humans, in isolation from each other. Fraudsters know this. And as we covered in our piece on document fraud red flags that compliance teams should recognise, the signs of manipulation that matter most are ones no human reviewer can see without specialist tools.
There is also a growing legal dimension to this. The UK Failure to Prevent Fraud legislation that came into force in September 2025 means companies can be criminally liable for fraud that slips through processes that were not reasonably robust. For insurance companies, “we reviewed the documents visually and they looked fine” is not going to hold up as a reasonable procedure defence, not with 98% industry acknowledgment that visual review is failing.
What forensic layers actually catch AI-edited insurance documents
Let me be specific about what document forensics detects, because the term gets used loosely. Here is what a proper check actually does.
Pixel-level analysis examines image content for signs of editing. AI retouching tools leave characteristic patterns: inconsistent JPEG block boundaries where the edit meets the original image, noise profiles that differ from the surrounding content, cloning artifacts where textures have been copied and pasted to extend a damaged area. These patterns are invisible to the human eye but statistically significant to an algorithm running against thousands of comparison examples.
Metadata inspection checks the document’s internal history: creation software, modification timestamps, the sequence of edits recorded in the PDF structure. A document that claims to have been generated by accounting software but shows signs of having been opened and re-saved in a different application is flagging a discrepancy that warrants investigation.
Compression artifact analysis looks at how image data within a PDF has been encoded. When a photo is embedded in a PDF, then extracted, edited and re-embedded, the compression signature changes. That layering is detectable as a distinct pattern in the file structure.
Cross-reference checks compare data points across multiple submitted documents. If a repair invoice shows a VAT number that does not match the supplier’s Companies House registration, or if the claimed repair date falls on a Sunday when the body shop’s own website shows it is closed, that inconsistency is an immediate red flag. These cross-reference checks are impossible to do manually at scale. They are trivial to automate.
These checks run in parallel, automatically, in seconds, before a single adjuster opens the file. The human reviewer sees a risk rating and a list of specific flags, not a raw document with no context.
How VerifyPDF flags AI-edited insurance claims documents
At VerifyPDF, we run pixel-level forensics and metadata checks against every document submitted to our API. For insurance claims, that typically means damage photos, repair invoices, police reports and medical certificates.
The analysis runs in under 5 seconds. The output is a risk rating: Trusted, Low risk, Needs attention or High risk, with specific flags explaining what triggered each finding. An adjuster reviewing a “Needs attention” document still uses their own judgment. The difference is they start from a specific, answerable question: was this image retouched at the boundary between the dented panel and the adjacent bodywork? That is a very different task than staring at a photo hoping something looks off.
In our experience, forensic screening earns its keep as triage. It tells adjusters which documents need closer scrutiny, so experienced human reviewers spend their time on the right 5% of the queue rather than reviewing everything at the same shallow level of attention. The humans stay in the loop. They just work on better information.
The math is not on manual review’s side
A major UK insurer processes tens of thousands of claims per month. Even if every adjuster were a trained document forensics specialist, the volume makes thorough manual review impossible at current staffing levels. You cannot give each document the forensic scrutiny that an automated system applies in under 5 seconds per file.
And remember who is driving the 71% rise in UK insurance fraud: ordinary people with a consumer-grade AI editing subscription and 20 minutes spare, not organised crime rings with classified technology. The tools to commit this fraud are cheaper and more accessible than the tools to detect it - unless you automate the detection.
Here is what I find genuinely frustrating: the solution is not new or expensive. The forensic techniques exist. The infrastructure to apply them at claims volume exists. The bottleneck is the decision to put systematic document forensics between the submission inbox and the adjuster queue, rather than treating it as an occasional tool for suspicious cases.
If you are running claims at a UK insurer and still relying on visual review as the primary filter for evidence documents, treat the Admiral numbers as a message addressed to you. Your process is being systematically exploited, at scale, right now.
Is this happening to your organisation? Almost certainly. The better question is how many of those claims have already been paid. Running every damage photo, repair invoice and police report through VerifyPDF before it reaches an adjuster closes that gap in seconds, without slowing a single legitimate claim.
Everybody wins… except the fraudsters.