Fraudsters have AI too: What rising AI-written dispute claims mean for card issuers

Ask an experienced dispute analyst how they spot a suspect claim, and they will often describe the same instinct: the story doesn't quite hold together, the phrasing is inconsistent, the timeline wobbles, and details are thin. For years, that instinct, built on thousands of poorly written complaints, has been one of the most reliable tools in an issuer's triage process.

Generative AI is changing that.

The tell that trained a decade of dispute teams

First-party fraud, sometimes called friendly fraud, has always presented issuers with a harder problem than third-party fraud. There is no compromised card to block and no unfamiliar device to flag. The cardholder is real, the transaction is real, and the only question is whether the claim itself is honest.

Historically, dishonest claims carried a signature. Opportunistic disputes tended to arrive with typos, contradictory details, and narratives that fell apart on a second reading. Dispute teams learned to read fraud the way an editor reads a manuscript. Language quality became a working proxy for plausibility, and for a long time it was a reasonable one.

What generative AI changes in the claims queue

A large language model writes a fluent, plausible dispute narrative in seconds. It keeps the timeline consistent, includes the right level of detail and strikes a credible tone. Image generation tools can produce a passable screenshot or receipt to accompany it.

The result is that claims arriving in issuers' queues now read the way honest ones do: clear, coherent, complete. A well-written complaint no longer signals a genuine one, and a team trained to triage mainly on language quality is triaging on a signal that is disappearing.

This matters most at scale. The effort that once limited fabricated claims, the time it took to construct a convincing story, has fallen to almost nothing. Polished, plausible claims can now be produced in volume.

First-party fraud is rising as the tell disappears

The volume trend is measurable. More than six in ten merchants reported an increase in first-party misuse over the preceding year, according to a 2025 report published by Verifi and Merchant Fraud Journal Report. Visa similarly describes first-party fraud as a rapidly growing challenge and a leading cause of chargebacks across the payments ecosystem.

Merchants experience the commercial impact through lost revenue, inventory losses, and chargeback-processing costs, while issuers must determine whether each claim reflects genuine third-party fraud, transaction confusion, or first-party misuse. That assessment is difficult when issuers and merchants lack the shared transaction information needed to distinguish legitimate purchases from unauthorised activity. As dispute volumes rise, that ambiguity, combined with scheme requirements and response deadlines, makes manual triage increasingly difficult to scale, according to Mastercard's First-Party Trust, Ethoca's first-party fraud, and Visa's dispute management.

Agentic commerce compounds the ambiguity

The claims queue is changing at the same time as the transactions themselves. Google's Universal Commerce Protocol, announced in January 2026, and the global rollout of Mastercard Agent Pay mean that a growing share of purchases will be initiated by AI shopping agents acting on behalf of cardholders.

That adds a second layer of ambiguity to what issuers must adjudicate. When a dispute arises, the question is no longer only whether the claim is honest, but who or what actually made the purchase and on what instruction. We explored that shift in detail in Agentic commerce needs a new dispute framework.

From language-pattern intuition to evidence-level intelligence

The instinctive response is to train agents to read or listen more carefully. However, when statements (written or verbal) are machine-generated, it’s more challenging to rely solely on language quality to detect first-party fraud.

The judgement has to move to where the evidence actually is. That means checking submitted documents against scheme requirements rather than against writing quality. It means looking at claim history and behavioural patterns across cases rather than at a single narrative in isolation. And it means structuring intake digitally, so every claim arrives with the data needed to assess it rather than as free text an analyst must interpret.

None of this is beyond issuers. But it is a different operating model from the one most dispute teams run today, and the gap between the two is where AI-written claims currently succeed.

What card issuers should do now

Three practical moves stand out:

1. Structure intake: Guided digital claim collection replaces the free-text narrative as the primary artefact with verifiable data points.
2. Assess evidence at the document level: Submitted receipts, screenshots, and supporting files should be evaluated against the scheme's requirements for the relevant reason code and checked for AI-generated content.
3. Look across claims, not just within them: Serial disputers are visible in history and behaviour, not in any single well-written complaint.

The issuers who adapt fastest will be those who stop asking how a claim reads and start asking whether the evidence behind it holds up.

Where Amiko fits

Amiko, Rivero's dispute management platform for card issuers, was built around evidence-level intelligence. Amiko Virtual Agent evaluates statements and evidence against scheme requirements and deflects or flags first-party or serial disputer patterns before a chargeback agent opens the case (with 80% friendly fraud deflection through cardholder education and eligibility validation).

To find out more about Amiko, watch the recorded demo of the product to see it in action.