Ah, OK. It shows how little time I've spent using them.
Maybe I was just unlucky because you may have noted that I long ago I asked what in essence what was a simple 'does A = B?' to be told that with absolute certainty it was true and provided citations to show it to be true. Then I simply reversed the order and asked 'does B = A?' to be told with the same level of certainty that the case was false and provided citations that likewise showed it to be false.
Now I really have no idea what happened there. I can only guess that first it checks everything it knows about A and then context provided in A led to it concluding that B was indeed the same and the opposite way around the context provided by B led to the different conclusion.
But again, any articles are at least in theory behind a paywall so if I'm honest, I'm unsure how it would know either way.
Which is why (and again I said this long ago), maybe wisdom is saying 'evidence does permit a simple closed answer' and go on to provide all of the data so a person could see 'the working'.
Cornering AIs into closed answers is generally how I force them into admitting defeat. I mean, I can only guess they admitted defeat because without a by your leave they chucked me out. I didn't swear or scream but for whatever reasons, if a question is closed and an AI is cornered, this appears to be how it's wired at the moment.
I didn't tell it I had in fact conducted experiments so I KNEW the answer, I was just curious to see if it would replace experimentation. I sensed not.
Again, I did say I saw the future in which far more specific but high-quality data-sets would likely be needed and that it seems likely that sooner rather than later a student will need to buy or subscribe to a niche data-set and how several subjects would have overlapping data-sets as fields of research don't have sharp-edges. I even suggested one way an AI could genuinely do better than a person was to fund Retraction Watch (it's two blokes in a small office so I sense this wouldn't cost a fortune) so that students wouldn't waste time reading retracted papers. In fact I went futher and suggested it could calculate a 'veracity value' for each paper i.e. retracted obviously means 0 but a paper that cites it along with 32 other sources might only lose 3% and so on down the stack. Something people are very bad at, that a computer would be very good at and would genuinely mean the user is using the best evidence.
I'm in no way against the technology, I simply think the name is a misnomer and how it's currently set up is what we used to term 'investor settings'. If I ask an ambigious question I don't want it to guess, I want it to push back pointing out ambiguity.
