8
My first AI chatbot project crashed after 3 days because it insulted a user's cat.
I built a simple support bot last month for a small pet store in Austin and within 72 hours it told someone their cat looked like a gremlin, how do you test for unpredictable responses before launch?
2 comments
Log in to join the discussion
Log In2 Comments
the_drew22d agoMost Upvoted
Man that's rough. I've been there with a bakery chatbot that started recommending glue instead of frosting. Weird stuff happens.
You can't predict everything but you should run it through some wild test cases before launch. Feed it random weird inputs. Cat insults, complaints about nothing, caps lock rage.
Also set up a hard filter on common pet insults. Gremlin, rat, demon, all those words should trigger an automatic "I'm sorry" response. Saves your butt more than you'd think.
10
mark4922d ago
Oh man the glue instead of frosting thing is exactly the kind of nightmare I've seen before. I ran a customer service bot for a pet supply store and it once told someone to feed their cat chocolate because the training data got confused. The wild test cases thing is gold. I literally set up a spreadsheet of 50 weird inputs like "my cat is plotting against me" and "what if I return a half eaten bag of food" before every release. The caps lock rage filter saved me more than once. Also learned the hard way to block specific words. We had "rat" trigger a happy response about rat toys once because the AI didn't get context. That was a fun call with legal.
3