[throwaway010718]

"it is not very good at explaining the process it has gone through to reach such a conclusion".

The above statement is just too convenient and practically superstitious.

AI is data hungry. That implies that there are so many past incidences of corruption you can generate large training data sets. So perhaps even a random guess would be too efficient since it is right more often than not.

[nerdponx]

"it is not very good at explaining the process it has gone through to reach such a conclusion".

To be fair, generating human-understandable explanations of predictions in a complicated nonlinear model is difficult in general. You typically need to come up with some kind of simplification of the model (like LIME [0]), and it's far from perfect or generally-applicable.

[0]: https://homes.cs.washington.edu/~marcotcr/blog/lime/

[soraki_soladead]

SHAP is more recent and does work with (just about) any ML model: https://github.com/slundberg/shap

There's still more work to do in interpretability but models are rarely opaque black boxes with no way to see inside.

[logicchains]

def is_corrupt(data): return rand(0.0, 1.0) < 0.9