I have several problems with “Artificial Intelligence” and none of them are “main”. They are instead a morass of stupidity that makes the whole field rather suspect both academically and socially.

1. The terminology

There is no universally agreed-upon definition of what “intelligence” even is. Even in the very specific and most tightly-focused relevant field, cognitive science, there is no agreement on this. If you expand to include related fields like philosophy and biology, there is even less agreement.

But sure, a bunch of computer nerds are going to make an artificial version of the thing we can’t even define yet.

2. The grift

I get it. In the first wave of “AI”, the progenitors of the field got a bit big for their britches and called what they were looking into “Artificial Intelligence” because they had no idea how much work had already been done in the field of intelligence, consciousness, etc. (being computer nerds and not philosophers and biologists) and they done fucked up. It happens. Call it “aspirational naming”. Minsky and co. weren’t grifters. That happened later.

But when that field collapsed as it became increasingly clear that “symbolic logic” was not how our brains worked and was not moving us with any appreciable speed toward “intelligence” (whatever the fuck that is, c.f. #1 above), that should have been the end of it.

There was a second generation. Who referred to their little numerical parlour tricks as “neurons” placed into “nets”. Ask any biologist who specializes in brain structures if a “neural net” with its software (or sometimes hardware) “neurons” bears any resemblance to an actual neuron and how those operate and connect. (Just get ready to cover your ears because the laughter can get very sudden and very loud…) And here’s the thing: I’m positive the people in this generation knew damned well that what they were doing had nothing whatsoever to do with real neurons, but knew that if they called them what they actually were (“Differentiable Computational Graph” or “Universal Function Approximator”) nobody with academic purse strings would open them widely enough for them to get the hookers and blow they needed. I submit that the term “neural network” was specifically coined to disguise the truth of what they were in a fraudulent attempt to get more funding by wowing the non-technical (like me) into responding “Oh, my! They’re building actual brains! MOAR MUNNEEE!”

And there was continued histories of this in wave after winter after wave after winter. Genetic algorithms don’t work like genes. Ant colony optimizations don’t work like ant colonies. Even machine learning isn’t learning in any meaningful sense. Are they useful? HELL YES! They’re useful in a wide variety of niches. But none of them have anything to do with “intelligence” (whatever the fuck that is, c.f. #1 above) and the terminology is, in my considered opinion, very much a sales snowjob for the credulously bureaucratic.

So now we have “large language models”, which are probably the most honestly-named technologies (along with the various things like “diffusion” techniques for pictures, etc.) of anything in “artificial intelligence”, but which use a different grifts to sucker the credulous: the grift of conflating fluency with intellect and knowledge. (The reason for this conflation is obvious: we’ve never had examples before of fluency detached from intellect so we’re wired to confuse the two.)

3. The point you bring up

A whole lot of money and effort has been expended on a pipe dream: making an artificial version of something we can’t even define. (C.f. #1 above.) That money and effort would have been far better spent on fighting climate change, education, health care, and a bewildering variety of other things that would actually do humanity some good.

But grifters gotta grift.