I like the idea and would support it. My thoughts.

* False-positives need a good way to handle via UI

*User should be able to turn the feature off

* You're algorithmic approach could probably be improved on. The current approach could be easily circumventer I suspect unlike more traditional approaches like Naive Fisher Classifiers and such.

* By only storing the last 10 messages its also easier to circumvent than a larger value. Perhaps make this configurable?