Every dispute pipeline runs on a quiet assumption: that documents are expensive to fake. A receipt took effort. A delivery photo took a driver. A convincing dispute letter took a person who knew what a reason code was. That assumption held for fifty years of chargeback rules. It does not hold anymore.
1 · The assumption that broke
Card network rules were written for a world where evidence has a cost of production. Filing windows, compelling-evidence standards, pre-arbitration — the whole machinery prices in the idea that fabricating proof is harder than obtaining it honestly. Generative models invert that. The cost of a plausible document is now a prompt and four seconds. The rules have not noticed.
This is not a future problem. The synthetic material is already in the queue — I work the queue — and it does not arrive looking fake. That is the entire issue.
2 · Receipts, rendered
A generated invoice is not a blurry photocopy. It is typographically better than the real thing: a VAT-format number that validates, consistent kerning, arithmetic that adds up. The analyst's eye was trained on errors — the wrong logo, the misaligned column, the font that changes halfway down. The new fakes have none.
The tells have moved from the document to its metadata, and from the pixels to the provenance. Most dispute tooling still reads the pixels.
"The cost of a plausible document is now a prompt and four seconds. The rules have not noticed."
3 · The face that passes KYC
Onboarding asks for a face and a gesture. Models now render both in real time, injected into the camera feed so the liveness check never sees a screen. The check passes. The person does not exist. Every bad outcome downstream — the mule account, the bust-out, the dispute nobody can place — begins as a face that passed.
Verification that reads the image loses. Verification that reads the pipeline — sensor noise, device attestation, injection artefacts — still has a chance. That is a hardware conversation, and most policy documents have not had it yet.
4 · An identity with a five-year history
Synthetic identity fraud is patient fraud. Stitch a real National Insurance number to a fabricated name, open small credit lines, repay them for years, then max everything and vanish. The dispute queue only ever sees the explosion — never the five years of fuse. AI did not invent this pattern, but it industrialised the patience: a model will happily manage forty credit files at once and never get bored.
5 · The dispute letter that reads too well
This is the one I watch most closely. A cardholder who bought something and regretted it can now file a narrative that quotes the merchant's own returns policy, names the right evidence type, and describes the harm in exactly the vocabulary analysts are trained to reward. Friendly fraud used to fail on presentation — caps lock, no dates, no merchant contact. It no longer has to.
The uncomfortable part: a polished claim is not a false claim. Genuine cardholders also use these tools, and they should. Which means polish stops being a signal in either direction, and the evidentiary weight has to shift to things that cannot be prompted into existence — authentication logs, device telemetry, delivery to a verified address.
6 · Scripts at scale
Romance and investment scripts used to be broken English at scale; the errors were the filter that kept victim counts low. Models write flawless, localised, emotionally calibrated scripts in any language and adapt them mid-conversation. The awkward phrasing that used to save people is gone. When a victim finally reaches their bank, the coaching they received was better written than the bank's own scam warnings.
7 · Where the liability lands
When the attacker is a model, the liability still lands on humans. A customer coached by an LLM-scripted scammer has still, in legal terms, authorised the payment. In the UK, the PSR's reimbursement rules for APP scams on Faster Payments have been mandatory since 7 October 2024 — capped at £85,000, with costs split between the sending and receiving firms — but card-funded purchases sit in a different regime, and card-not-present fraud still runs through the old dispute rails[1]. Under PSD2 the payer holds refund rights for genuinely unauthorised transactions, while a passed SCA challenge shifts the evidential burden sharply onto them[2].
Detection has to move down a layer with it: from reading documents to verifying rails. The sentence is no longer evidence. The rail still is.
Interactive · Spot the Synthetic
Five pairs from a fictional dispute queue. One item in each pair was written or generated by a model. Pick it — the working is shown after every answer.
The Analyst's Takeaway
Every fake that reaches an analyst was graded by a person reading a document. That gate is closing. What survives is the material that cannot be prompted: SCA logs, device attestation, on-chain hashes, network-level provenance.
That is the bet behind every tool on this page — show the working, trust the rail, and treat a well-written document as a neutral fact rather than a good sign.
Sources
Luca Caranzano — Chargeback Analyst, Monzo Bank. Written for this portfolio to show what the AI threat landscape looks like from the dispute queue, not from a keynote. No statistics are invented in this piece; regulatory references are linked above. Not legal, fraud or operational advice.
