In a short article at Nature, Eddy Keming Chen, Mikhail Belkin, Leon Bergen, and Dave Danks argue that current AI has general intelligence.1 I think that they’re argument goes wrong in some of the details, but they deal with some issues very quickly.2 On reflection, I think my bigger problem is not with their argument but with their question itself.
They start by going back to Alan Turing’s 1950 paper in which he proposes what’s now called the Turing test, arguing that Turing would be satisfied that current AI are intelligent. AI can do lots of the smart things that humans can do, and the best AI systems aren’t constrained to just one kind of task or problem area.
Yet they slip from talking about general intelligence to using the term Artificial General Intelligence (AGI). AGI was adopted recently to be something distinct from and beyond ordinary AI. That linguistic maneuver was only necessary because it has became routine to refer to all sorts of machine learning as AI. Algorithms like neural nets, which we didn’t call “AI” back in the 90s, get branded as AI for marketing reasons. So it’s not coincidental that “AGI” is often used interchangeably with “superintelligence”, the kind of technology which was supposed to bring on the singularity. Even though CBBD distinguish general intelligence from superintelligence, talk of AGI carries with it implications of C3PO, Commander Data, or at least Gort.
Even general intelligence itself is a problematic notion. The history of intelligence testing in the 20th century illustrates how tempting it is to reify intelligence, as if it were one sort of thing. That simplifying assumption leads to all sorts of confusions.
The headline puts the point in terms of human-level intelligence, posing the title question: “Does AI already have human-level intelligence?” And they write in the article that AI has “core cognitive abilities at levels comparable to human-level general intelligence.” But human-level intelligence is a complicated grab bag of abilities, best thought of as a tangle of related things rather than an unambiguous thing. So the answer to their title question may be yes in a sense— but also no in a sense and it’s complicated.