Thanks @elsnoman for participating in that AMA!
As someone who evaluates the quality of AI solutions, two particular topics in the AMA got my attention:
- How can you assess confidence your users have in your AI powered software?
I interpret this as âhow do I know that users agree with the AIâs predictions?â or âhow do I know that my AI is giving the right answers?â
Yes, the key is observability â but that is a bit of a broad statement. The detail thatâs missing is that you need to find a way to associate consequential user activity to the AI prediction that spurred it.
âLikeâ, âdislikeâ, or âpick the best answerâ interactions like Carlos described are very obvious ways to get feedback, but I think their value is shallow and frankly suspect. Itâs akin to asking the user for their opinion, but as the old saying goes, âactions speak louder than words.â
Whatâs better is to find a way to tie AI predictions to user actions that have consequence. If a user makes âimportantâ decisions based on predictions, then I would feel a bit more confident about the quality of the AI predictions. Same if the AI predictions align with user actions (where perhaps there is no direct tie between prediction and action).
It reminds me a little of playing poker with friends where no money is involved. Risky betting has no consequences â you can just grab more chips. Players have no skin in the game.
- Train your models using high-quality data (try to avoid bias)
Reminder of the old computer science saying: GIGO - garbage in, garbage out
Your model is only as good as your training data. If your data has mistakes, has inherent bias, or doesnât really match the data that your users work with, then it will not give very good/relevant predictions.
Consider the concept of verification vs. validation: building the thing right vs. building the right thing
A model trained on poor/biased/inappropriate data may get very good statistical results for accuracy, F1, etc. but it may not actually satisfactorily âbuilt for purposeâ.