25
Just learned my AI training data was full of my own bad habits
A data scientist reviewed my chatbot's outputs and pointed out it was copying my overly cautious phrasing from a year of support tickets. I switched to cleaner example conversations from the top performers and response quality jumped 40% in a month. Has anyone else seen their own blind spots baked into their models?
3 comments
Log in to join the discussion
Log In3 Comments
janarivera3d ago
Oh man, THIS is SO my story! I had the exact same wakeup call when someone pointed out my bot was saying "just checking in" like ten times per conversation because I always say that in real life. I switched to modeling it after the coworker who gets straight to the point and my numbers shot up too, its crazy how blind you can be to your own patterns.
6
shanelee3d ago
What was the actual shift you made in how you programmed the bot? Like did you write a whole new set of canned responses or just tweak the existing ones to cut out all the filler? I'm curious because I keep trying to strip my bot's personality down but it still ends up sounding too much like me even when I delete the obvious stuff. Did you notice any awkward silences or weird transitions after you made the change?
10
faith_king2d ago
Right? Its like looking in a mirror you didn't know was there. I had the same thing with my bot saying "no worries" like 8 times in a row because I say it constantly without thinking. Modeling it after someone who just says "okay" and moves on was a game changer. You'd think it would feel rude but people actually liked it more. Guess my real personality was just adding a lot of unnecessary chatter. Did you have to retrain your brain to stop using the filler words in your own replies too or did you just let the bot be the opposite of you?
4