On 2/2/2023 2:24 AM, Arnt Gulbrandsen wrote:
Hi,
This is precisely why ChaptGTP is banned on Stack Overflow. The ban announcement says “its answers are frequently bad but look plausible” or words to that effect.
Arnt, yes, I was aware of that. But plausibility is not the only metric we use when we make decisions whether to rely on advice by human experts. While none are perfect, they are still pretty decent heuristics. (see comment added below). A./
Arnt
*From: *UA-EAI <ua-eai-bounces@icann.org> on behalf of Asmus Freytag via UA-EAI <ua-eai@icann.org> *Reply to: *Asmus Freytag <asmusf@ix.netcom.com> *Date: *Thursday, 2 February 2023 at 02:05 *To: *"ua-eai@icann.org" <ua-eai@icann.org> *Subject: *Re: [UA-EAI] Stackoverflow and ChatGPT
Marc,
you've identified two of the main problems with AI based answers. One, the impossibility to fix misperceptions embedded in the training data. The other, the difficulty for a human user to "rate" the level of expertise behind an AI generated answer. You can't use plausibility, because all AI generated answers are plausible. You can't use other "tells" of inexperience, because the AI generated texts don't have them. And you can't "rate" the AI like you can human experts; nor can you rate the true experts that it plagiarizes.
What is meant is that with human experts you can look at their track record and form some qualified opinion about what level and type of expertise to expect. Part of that depends on associating individuals with a specialization. The AI is not specialized in a knowledge domain and one result doesn't predict the next.
If, as can be expected, significant levels of AI generated output are then being published on various sites in ways that even an AI can't be sure it's AI generated, then you not only get garbage in, garbage out, but fully recycled garbage.
Misinformation of the human generated kind is harmless by comparison.
A./