3 ms·
The problem is that detecting fraud is fundamentally harder than generating plausible fraud. This is because ultimately a very good fraud producer can simply pr
by roca 2y ago
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.
- Lerc 2y agoIt'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.
- hansvm 2y agoI 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.