[reaperducer]

it quickly became apparent that ChatGPT was exceedingly willing to outright lie to you.

I've sometimes thought that these "AI" chat systems might be better if they were taught the simple human phrase "I don't know."

I think it would improve trust in these system if people knew that they had limits, and were aware of them.

It would be even better if something like the one on Bing, for example, could respond "I don't know, but here are some links to places where you might find the answer…"

It's like when it starts to become apparent that the new kid at school is a compulsive liar. Eventually people stop listening to him.

[rsynnott]

That's not really how it works, though. The line `if(dont_know) {make_shit_up()}` does not appear in the source code. It doesn't know it's lying; it doesn't know anything!

You could train one to appear terribly uncertain, but it would still 'lie'.

[AnimalMuppet]

OK, but it's probability-based. It will pick lower probability answers depending on the temperature. But that means it knows what the probabilities are.

So what it needs is a line of the form 'if(highest_probability < threshold){dont_know()}.

[cyanydeez]

That'd take a whole different model. Remember, these are billions of vectors.