A woman in Milwaukee called Starr ordered chicken wings on DoorDash, ate one, and decided the rest weren't worth paying for. So she opened ChatGPT, edited a photo of the bitten wing until it looked raw and undercooked, sent it in as proof of a bad order, and collected $39.24 in credit. Then she filmed the whole process and posted it to TikTok, where it ran up 4.4 million views — one of which belonged to DoorDash, who replied underneath: “Oop should've blocked us!”

It's a good line. It's also the only reason she got caught. DoorDash didn't spot the fake because a fraud system flagged a doctored image. They spotted it because four million people watched her brag about how she'd done it. The next thousand customers who try the same trick won't be filming for an audience, and their photos will sail straight through the refund queue.

That's the part worth paying attention to. The stunt itself barely matters; what it exposed does. For about twenty years, every refunds and loyalty playbook has rested on one unspoken assumption: that the evidence a customer hands you is real. A photo of a crushed parcel, a screenshot of a payment that failed, a picture of a cold and congealed meal: these got treated as facts, because faking them convincingly used to be more trouble than the refund was worth. That assumption has expired. A photo isn't proof of anything any more. It's a claim, like any other claim, and now anyone with a phone and a chatbot can manufacture one in under a minute.

The direction of trust has reversed. The old default was to believe the customer's evidence unless something obvious gave them away. The new one has to run the other way: treat any single piece of evidence as unverified until the rest of the account backs it up. That's a small change to say and an awkward one to live with, because it cuts against everything a good customer-service operation has trained itself to do.

The playbook that just expired

Think about what the standard advice has been. Refund fast. Don't interrogate the customer. A few quid lost to the occasional chancer is cheaper than the goodwill you burn by making an honest person prove themselves. Amazon built a continent-sized business partly on no-questions-asked returns. Every customer-service consultant of the last two decades has told small operators the same thing: trust the customer, eat the small losses, protect the lifetime value. It was good advice, because the maths worked. The cost of fraud was bounded by how hard fraud was.

What's changed is that this kind of fraud, call it AI refund fraud, just became nearly free to commit, while the thing it attacks — your willingness to take evidence at face value — is exactly what you were told to rely on. The customer used to need a genuinely bad experience, or at least the patience to fabricate one. Now they need a free tool and a grievance. The DoorDash case is almost quaint because it involved real chicken. The version that should worry you is the order that arrived perfectly fine, photographed, then aged a week in software.

I've written before about how agentic commerce keeps colliding with a trust model that doesn't quite hold — customers nervous about letting a chatbot spend their money, retailers nervous about a platform sitting between them and their own buyers. That was a question about who controls the intelligence when a machine does the buying. This is the same problem from the other side of the counter: what happens when the customer is the one holding the synthetic evidence. We spent a year asking whether we could trust the AI doing the buying. The more awkward question is whether we can trust the human holding it.

Why this lands hardest on small businesses

DoorDash will be fine. It can throw machine-learning detection, image forensics and a fraud team at the problem, and absorb the losses it misses. The corner takeaway can't. Neither can the independent e-commerce seller shipping forty parcels a day, or the small subscription service with one person handling support.

And here's the bitter irony: the automated, one-photo-equals-instant-refund flow that's now the soft target is precisely what small operators adopted to compete. They couldn't staff a returns department, so they bought the friction-free policy as a feature. “Just send a photo, we'll sort it.” That was the promise that let a two-person shop feel as easy to deal with as Amazon. The same automation that made them competitive is now the door that's been left unlocked.

So the SMB sits in the worst spot. Too small to detect fakes at scale, too dependent on goodwill to start treating customers like suspects, and running the exact lightweight process that's easiest to game.

Picture the owner of a single chicken shop opening the tickets on a Monday morning: three refund claims over the weekend, each with a photo of something that looks underdone, two of them from accounts nobody recognises. Refund all three and you've handed an evening's margin to people who may well have eaten the lot. Challenge all three and you've called whichever one of them was honest a liar. There's no clerk in the back office to escalate it to, no fraud team, just you and a Monday. And whichever way it goes, the cost lands on the shop and on the driver who delivered a perfectly good meal, never on the platform that promised the easy refund in the first place.

The trap is overcorrecting

The wrong response is to swing the other way and assume everyone's lying. That destroys the very thing the refunds policy was built to protect. Most customers are honest. The person whose chicken really did turn up raw still exists, still deserves a fast refund and a kind word, and will tell ten people if they get treated like a fraudster instead. Tighten the screws on everyone and you've solved a fraud problem by manufacturing a loyalty one — which, for most small businesses, is the more expensive of the two.

This is the hard part. You're asked to stay warm and generous with the ninety-odd per cent who are telling the truth, while no longer taking their evidence at face value — two instincts that pull against each other. The way to hold both is to separate the customer from the claim: treat the person generously and the evidence sceptically, instead of collapsing the two into a single yes or no.

What changes in practice

None of this needs a fraud team you can't afford. It needs a refund policy that stops treating a single image as the end of the conversation, and three changes carry the bulk of it.

  1. Weight the account, not the photo. A customer of three years with twelve clean orders and one complaint is a different proposition from an account that opened on Tuesday. Most of the signal you need is in the history you already hold, not in the picture you've just been sent — and fraud at scale leaves a pattern, the same address or the same suspiciously photogenic disaster, long before any individual image looks wrong.
  2. Make the friction proportional to the value. A £6 dipping sauce isn't worth investigating; asking for the order number is plenty. A £600 claim earns a human glance and a follow-up question, because the person fabricating it has put real effort in. Scale the scrutiny to the stake, the way you'd scale any risk.
  3. Put a person in front of the money above a threshold. Not every refund — that drowns you — but the ones where the amount, the account or the pattern crosses a line you've set in advance. Automating the refund was always fine. What got automated along with it was the judgement that used to sit behind it, and that's the part to claw back.

I'm genuinely unsure where the equilibrium lands. Detection and fabrication tend to escalate together — better fakes provoke better detectors, which provoke better fakes — and I can't see why that race favours the defender, especially the small defender. What I do know is that the cheap, trusting refund process belonged to an era when forging evidence was hard, and that era is over. What replaces it hasn't settled yet, and it's being worked out by people with far bigger fraud teams than you or I have.

Where this leaves us

The customer who used to walk through your door, then browse your website, now increasingly arrives as something AI-shaped — and, as Starr demonstrated, sometimes arrives holding an AI of their own. The technology cuts both ways. It always does.

None of this means the small business is doomed, or that you should greet every complaint with a forensic audit. It just means the cheapest assumption you ran for twenty years — that a photo is a fact — has stopped being safe, and the policy built on top of it needs rewriting before the next clever customer finds the gap. Do that thoughtfully and most customers will never notice. Leave it as it stands, and the bill won't land on DoorDash or Amazon — it'll land on the small restaurant eating the cost, and the driver who gets blamed for a meal that was perfectly fine.