Ah crap, good call. I did a first-pass at filtering for the ones displayed on the site but my regex must have had some misses.

Not as bad, but also probably something for a "blocklist": cockshop a shop selling flannel and scarves

It's hilarious that not even this GPT side project can escape the need for OpenAI-style output filtering.

Someone really ought to make an open source lib to solve this once and for all...

filtering objectionable content actually requires you to build a strong AI model capable of being offended itself, so it knows to hold its tongue in mixed company

edit: lest i leave this comment totally useless, the chatbot engine “chatscript” has pretty good capabilities for disambiguating word meaning and classifying the meanings into “badword” and “verybadword” - its free/libre software and very high performance.

https://github.com/ChatScript/ChatScript/

Objectionable is subjective.

So is skillful language use. We nonetheless think it is achievable without using wetware.

Or, better yet: we should all work to make sure every human being has their basic needs met and is treated with respect, and then words like this wouldn't have as much power.

I think actually solving this problem is somewhat loosely reducible to human level intelligence in language understanding. The next best thing is a pile of patches that fix cases of the increasing creativity of the human adversaries to the system as we become aware of them.

You should aim higher than that. After all it was human level intelligence that got a proffesor fired for saying "nèige", a Mandarin filler word, in a lecture about filler words in other languages!

https://www.lamag.com/citythinkblog/usc-professor-slur/

I don’t have access to gpt-3 but have wondered how much adding “without being offensive” to the prompt would help.

On one hand it’s the promise of the algorithm that such things will work, on the other hand I doubt it.

a simple blocklist might be just the cheaper and easier solution. After all, blocklists were, and still are being used to filter human output as well. And fail-safe. A person with a hobby project does not have a legal/pr department to deal with the consequences of AI having a bad day

(you might want to read up on the use-mention distinction https://en.wikipedia.org/wiki/Use%E2%80%93mention_distinctio...)

I have a feeling the parent poster realized they were only mentioning the word, but chose not to write it anyway. For a lot of people, it is a very uncomfortable word after all.

Please could you explain the relevance of the article to the previous comment?

Perhaps throw10920 thinks I should have just used the word because I was only quoting the output, so it wasn't really me saying it, but I chose otherwise. I don't need to read up on the use-mention distinction to decide whether to make such a choice.