Hacker News Reader

I think they can be useful without saying you "highly trust" them. If you moderately trust them, you can still let them do grunt work that you'll know you have to look over before committing.

Use and trust are different things. I use Google Search, I don't trust the first answer it gives me until I confirm it. I use co-pilot, but I don't accept every request. I don't trust them to be right, or even commonly right. Even the code generation benchmarks for language models are sitting between 30-50% correctness for the absolute best models. Would you trust a lawyer that is right 30-50% of the time?

I think the key word is "highly".

I won't argue the metrics but while I generally trust Copilot and ChatGPT, I would not say I "highly trust" them.