[deely3]

Can I just say that Im actually become scared reading your comment? Personally I would never ask chatGPT these questions because for me these questions are hard to verify, and knowing how often AI likes to hallucinate.. I just can't trust it.

You mentioned 50/50 correctness in domain questions. I can't be sure that other hard to verify questions do not follow these percentage..

[qrybam]

It IS dangerous. You must apply critical thinking to what’s in front of you. You can’t blindly believe what this thing generates! Much like heavy machinery, it’s a game changer when used correctly, and likewise it can be extremely damaging if you use it without appropriate care.

[vertis]

Quantum computing has a similar problem, in that the error rate is high. As does untrained data entry. You can put things in place to help counter this once you know it's happening.

[JeremyNT]

I'm reluctant for the same reasons.

Google search might uncover BS too, but I'm already calibrated to expect it, and there are plenty of sources right alongside whatever I pulled the result from where I can go immediately get a second opinion.

With the LLMs, maybe they're spot on 95% of the time, but the 5% or whatever is bullshit, but it's all said in the same "voice" with the same apparent degree of confidence and presented without citations. It becomes both more difficult to verify a specific claim (because there's not one canonical source for it) as well as it involves more cognitive load (in that I specifically have to context switch to another tool to check it).

Babysitting a tool that's exceptionally good at creating plausible bullshit every now and then means a new way of working that I don't think I'm willing to adopt.