TIL my AI chatbot kept hallucinating the same wrong fact 7 times in a row

I was building a customer support bot for a local bakery, and it kept insisting their hours were 9 AM to 9 PM on Sundays when they close at 3 PM. After digging through the training data, I realized I had an old scraper feeding it a stale Yelp page from 2019. Has anyone else had their model latch onto a single bad source like that?

3 comments

3 Comments

cora51825d ago

that 2019 yelp page thing hits close to home. i had a similar mess with a music recommendation bot i was tinkering with - it kept telling everyone that a local venue's open mic night was on tuesdays when they actually switched it to wednesdays back in 2018. took me forever to figure out the source because i had accidentally left an old rss feed running in the background. did you end up having to completely re-scrape your data or did you just delete that one bad yelp page?

taylor.reese25d ago

Ended up having to nuke the whole data set and start fresh, honestly. That one bad yelp page kept poisoning the well even after i deleted it because the AI had already learned the wrong info from other sites that copied it. My advice is to just re-scrape everything, it saves time in the long run vs playing whack-a-mole with bad sources.

sandra_bennett5925d ago

Or you could just filter better and not burn everything down.