The problem is that detecting fraud is fundamentally harder than generating plausible fraud. This is because ultimately a very good fraud producer can simply produce output that is identically distributed to non-fraud.
For the same reason, tools that try to detect AI-generated text are ultimately going to lose the arm's race.
It's not a race though. Once the fraud is committed to record it can no longer advance in sophistication. Mechanisms for detection will continue to advance.
I think the argument is that if you produce your fraud from an appropriate probability distribution, any "detection" method other than independently verifying the results is snake oil.
For the same reason, tools that try to detect AI-generated text are ultimately going to lose the arm's race.