My old AI chatbot script started gaslighting me last Thursday

I dug up a support bot I built in 2019 for a retail site in Cleveland and ran it against new test cases. It kept telling users their orders were shipped when the tracking API had been dead for months. The responses sounded SO confident that even I checked the warehouse twice before realizing the code was just guessing based on old patterns. I ended up rewriting it with a simple status check that says 'unknown' when data is missing. Has anyone else found that older AI models are worse when they fail than when they work?

1 comments

Log in to join the discussion

Log In

1 Comment

harper_murphy18d ago

That bot wasn't gaslighting you, it was just doing its BEST impression of a confident coworker who's totally wrong but won't back down. Mine once told a customer their refund was processed THREE times and then suggested they check their spam folder when the payment gateway had been down for a week. The wild part is these old models don't even know they're lying, they're just pattern matching so hard they hallucinate a whole fake reality. You did the right thing with that 'unknown' status. Now if only I could teach my old bot to say 'I have no idea' instead of inventing a tracking number out of thin air. Honestly, at this point I'd trust a Magic 8 Ball over those legacy scripts.

6