Initial AI was trained on massive amounts of uncontrolled inputs--which resulted in large volumes of irrational outputs (literally, non-existent material being created by the AI). The problem is simple: The newer/better (? LOL) versions of previous AI(s) were based on flawed/failed models.

Sounds like what happens to a person who receives their education from Fox News (just kidding :-)

Let's assume (may, or may not be, true) that all AI's are not created equal and, possibly just as important, are not trained with the same information. Just as a person has to "weigh" new data to see if it is correct, there should be a way for an AI to also make this determination (possibly by the curating of the data it's being fed. One does not let our kids read anything they want and there is no reason to let our AIs be exposed to invalid information if it's avoidable.

I suspect the difference between a "computer" and an AI has more to do with pattern recognition more than any other factor. That coupled to their ability to learn "how to learn". Humans sometimes take drugs which cause them to hallucinate (and a few can hallucinate without the aid of drugs). They may not realize at the time that they are acting irrationally, but it is obvious to an outsider. That means that there is an opportunity to have a built-in "outsider" embedded in an AI as a governor to reel it back to reality if it drifts. (Maybe a specialized second AI?).

AIs are still in their infancy. Heck, they can't even drive an autonomous car - but, give them time.

I remember buying my first calculator, my first PC, my first fax machine, my first cell phone, etc., etc. Technology has leapt ahead in quantum leaps during my single lifespan. My wristwatch probably has more computing power than the mainframe I fed punch-cards to. The amount of resources being poured into AI leads me to believe that their advances will be exponential until mankind will be an inconvenience they have to put up with.

Jeff