I can't prevent gpt 3.5 from using certain words

Hello everyone,
so I’ve struggling with this for some long hours by now:

There is a list of words I don’t want gpt to use in its output and it doesn’t matter how I state this in the prompt, it keeps using them.

I’ve tried things like:
“Do not use the following words: [“banana”, “apple”]”.
Then I learned gpt isn’t very good with negative requests… then I tried:
“Refrain from using the following words: “apple”, “banana””

… It didn’t help, it keeps including the forbidden words.

Anyone has a similar issue? Anyone knows how to solve it? Thanks!

Assuming you are using the api, Did you put it in the prompt or system command. Sadly 3.5 does not pay as much attention to instructions as 4 - my advice is to capture the output programmatically and if it contains your words, feed back and ask ti amend. I did a quick test just now and that seems to solve it for you

Unfortunately, there is such an issue in 3.5. I experienced a similar problem, but I solved it with GPT-4.

1 Like