Anecdotally, some folks decided not to use after the bot took long (anything more than near-instantaneous is long) to get a response. Thus, we may have to re-engineer such that we always reply with machine category first, but then
Steps
Validate this first using data
Retrain ML model to NOT predict unsure but rely on proba thresholds instead
Figure out how to get an automated assessment of the 3 info reply categories (misleading, true, untrue). Probably agents.
Once above is in place, then implement.
Need to caveat properly. Conceptually, we are trading off accuracy for speed maximally. Interim reply can then be canned.
Anecdotally, some folks decided not to use after the bot took long (anything more than near-instantaneous is long) to get a response. Thus, we may have to re-engineer such that we always reply with machine category first, but then
Steps