I got a chatbot to stop making stuff up by pasting fake contracts into it first, and now I can't decide if that's a fix or a hack | AI Innovations - Opinions Matter
I got a chatbot to stop making stuff up by pasting fake contracts into it first, and now I can't decide if that's a fix or a hack
So I run a small claims review thing for a property office in Boise, about 60 lease files a month, and I was using a chatbot to pull key dates and rent amounts out of scanned PDFs. Problem was it kept inventing numbers. Twice last month it told me a tenant owed $1,850 when the real number was $1,150, and I only caught it because I checked by hand. Out of frustration I tried something dumb: before each real file, I paste in two fake lease pages I wrote myself with obviously wrong dates and rent amounts, and ask it to list what it sees. When it repeats my fake numbers back correctly, it seems to settle down for the real one. Oddly, my error rate dropped to almost nothing over about 40 files, though I can't prove why and it might just be luck. My coworker says this is a waste of time and I should just switch to a better model, and my boss wants a real audit before she trusts it. Is this a legit prompt trick people use, or am I fooling myself and getting lucky?