There is an old grifter’s truism: you can’t kid a kidder. The con artist is the one person who sees the con coming, because they know the moves from the inside. It is a comforting idea for those of us who work in fraud prevention. We like to believe that experience makes us harder to fool.

But here is a question that turns the idiom on its head. Forget whether you can fool a fraudster. Can you manufacture a fraud convincing enough to fool a machine? And more to the point, should you?

This is not an idle thought experiment. It sits at the centre of one of the most stubborn problems in financial crime detection, and new research has just given us a useful way to think about it.

The rare-event trap

Every fraud team lives with the same inconvenient arithmetic – fraud is rare. In a healthy portfolio, it might be a fraction of one per cent of all transactions. That rarity is good news for customers and terrible news for the models we train to catch it.

When fraud is rare, supervised learning quietly optimises for the wrong thing. The model learns that the safest bet, almost every time, is to call a transaction legitimate. It is rewarded for this. It posts a headline accuracy figure north of 99 per cent and a board paper that looks reassuring. Underneath that number, fraud recall can be quietly dismal. The model is accurate precisely because it has learned to ignore the thing we built it to find.

The usual fix is to give the model more data to learn from. Since we cannot wait for criminals to oblige, we generate synthetic fraud. Oversampling techniques like SMOTE have done this for years. The newer, more ambitious idea is to use generative models to manufacture fraud that looks and behaves like the real thing.

Which brings us back to the question. If you are going to fake a fraud, how do you know your fake is any good?

The paper, and the part that actually matters

The research in question, Q-SYNTH, proposes generating synthetic fraud using a hybrid quantum and classical approach. A quantum circuit produces the samples, a classical network judges them. It is genuinely interesting work, and the quantum framing will get the headlines. Ignore the quantum part. It is not the lesson.

The lesson is the evaluation discipline the authors insist on, and it applies to any synthetic data, generated by any method, on any laptop. They argue that fake fraud must be judged on two entirely separate axes, not one.

The first axis is fidelity. Does the synthetic fraud statistically resemble real fraud? Does it sit in the same part of the feature space, with the same shape? The second axis is utility. Does training on it actually make your detection better? More fraud caught, fewer cases missed. These are not the same question, and this is the insight worth carrying into every model risk conversation you have this year. A synthetic dataset can be beautifully realistic and useless for detection. It can also be statistically rough around the edges and still lift your recall. Realism and usefulness are different properties, and we conflate them at our peril.

The authors add a third, rather elegant check. They train a classifier to tell the real fraud from the fake fraud. If it cannot, the synthetic data has passed a kind of Turing test for fraud. Their method scored close to chance on this, which they treat as a good sign. Nobody, machine included, could reliably spot the forgery. So can you fake a fraud? Yes. But “good enough to look real” and “good enough to be useful” are two different verdicts, and you need both.

The trade-off nobody can engineer away

Here is where it gets operationally interesting. These axes pull against each other. In the comparisons, plain oversampling looked strongest on raw statistical similarity. That did not automatically make it the best choice for actual detection performance. Optimise hard for one axis and you can quietly pay for it on another. There is no free lunch, only a choice about which lunch you are buying.

This is exactly the kind of trade-off that should live in a model risk framework rather than in the quiet discretion of whoever wrote the training pipeline. When you decide how your synthetic fraud is generated, you are making a risk decision, whether or not anyone has framed it as one.

Treat the synthetic ratio as a lever, not a default

The most practical finding is almost buried. When the researchers increased the proportion of synthetic fraud in training from a quarter to a half, fraud recall and F1 both climbed. More manufactured fraud, more caught.

Read that as a governance statement, not a technical footnote. The amount of synthetic fraud you inject is a dial. Turn it one way and you accept more missed fraud. Turn it the other and you trade against other costs, including false positives and the customer friction they create. That dial is currently sitting inside your model, set to whatever value an engineer chose months ago, probably undocumented.

I would put it differently to a board. The synthetic ratio is a harm reduction lever. It belongs in a policy, owned by someone accountable, reviewed on a cadence. It does not belong in a config file nobody reads.

The honest caveats

Three, because precision matters and overselling synthetic data is its own quiet harm. This is a preprint and is not yet peer reviewed. It doesnt mean it wont be, it is just early research. Promising, not proven. It works in a heavily compressed feature space, far smaller than a real transaction record. It is a proof of concept, not a production system.

And the fidelity tests look at features one at a time. They do not yet capture how features interact, which is exactly where real fraud lives. A transaction is suspicious because of the relationship between its parts, not because any single field looks odd in isolation. The authors know this and flag it as the next problem to solve.

The line to hold

The authors are careful about something we should all repeat more often. Strong results on synthetic data are not evidence that real fraud data can be replaced. Manufactured fraud trains the model. It does not understand the adversary. The criminal still innovates faster than any generator can imagine, and the day we forget that is the day the kidder kids us after all.

So, can you fake a fraud? Yes, convincingly enough to fool a classifier. The harder and more useful question is whether your fake makes you better at catching the real thing, and whether you can prove it on both axes before it touches a production model.

You can’t kid a kidder. But you can, it turns out, manufacture one. The discipline lies in knowing exactly what you have built, and exactly what you have not.

The research: Innan, A., El Alami, M., Innan, N., Shafique, M., & Bennai, M. (2026). Q-SYNTH: Hybrid Quantum-Classical Adversarial Augmentation for Imbalanced Fraud Detection. arXiv:2605.21164v1 [cs.LG; quant-ph]. Published/submitted 20 May 2026. https://doi.org/10.48550/arXiv.2605.21164. Article type: arXiv preprint (not peer-reviewed).