An article in the October issue of IEEE Spectrum investigates the status of artificial-intelligence IQ tests, and speculates on when and whether we will see the arrival of so-called artificial general intelligence (AGI), which author Matthew Hutson says is "AI technology that can match the abilities of humans at most tasks." But mostly unasked in the article is an even more basic question: what do we mean by human intelligence?
To be fair, Hutson has done a good job of surveying several popular benchmark tests for AI intelligence. One of the more popular tests is something called the Abstraction and Reasoning Corpus (ARC for short), whose developer François Chollet has made something of a go-to standard, as charts in the article show the scores of over a dozen different AI programs on various versions of Chollets tests. Engineers like numbers, and standardizing a test is a good thing as long as the test measures what you want to know. But does ARC do that?
The version of the ARC test described in the article consists largely of coming up with patterns of colored figures that follow rules abstracted from examples. Human beings can score higher than AI systems on these tests, although the systems are improving. But it's an open question as to whether abstracting patterns from geometric shapes has a lot to do with being generally intelligent.
Chollet feels that his test measures the ability to acquire new abilities easily, which he thinks is the prime measure of intelligence. Whether it actually does that is debatable, but it looks like ARC is to the AI world what the old Stanford-Binet IQ test is for people. That IQ test was developed over a century ago and is now in its fifth edition.
Hutson comes close to the problem when he admits that "notions of intelligence vary across place and time." The Stanford-Binet test is mainly used to identify people who don't fit in well with the public-school system, which is mainly designed to produce worker bees for the modern economy. As the modern economy is shifting all the time, what counts as intelligence does too.
And even if we could perfectly track these shifts, the admittedly infinite array of "tasks" that people perform present an almost insurmountable problem to anyone who wants not only to define, but to evaluate something that could justifiably be called artificial general intelligence.
Geoffrey Hinton, a recent Nobel Prize winner in AI, is quoted in the article as saying that if an AI robot could successfully do household plumbing, that would be a milestone in AGI, and he thinks it's still about ten years off. I hope I'm around in ten years to check this prediction, which I personally feel is optimistic. For one thing, humanoid robots will have to get a lot cheaper before people even consider using one to fix a toilet.
All these approaches to AGI ignore a distinction in the field of human psychology which was first pointed out by Aristotle. The distinction has been described in various ways, but the most succinct is to differentiate between perceptual thought and conceptual thought.
Perceptual thought, which humans share with other animals and machines, consists in perceiving, remembering, imagining, and making associations among perceptions and memories, broadly speaking. Inanimate material objects like computers can display perceptual thought, and in crawling the Internet for raw material by which to answer queries, all AI chatbots and similar systems use perceptual thought, which ultimately has to do with concrete individual things.
On the other hand, conceptual thought involves the consideration of universals: freedom, for example, or the color blue, or triangularity as a property of a geometric figure, as opposed to considering any individual triangle. There are good reasons to believe that no strictly material system (and this includes all AI) can engage in truly conceptual thought. With suitable programming by humans, a computing system may provide a good simulation of conceptual thought, as a movie provides a good simulation of human beings walking around and even engaging in conceptual thought. But a movie is just a sequence of images and sounds, and can't respond to its environment in an intelligent way.
Neither can an AI program engage in conceptual thought, although by finding examples of such thought in its training, it can provide a convincing simulation of it. While having a robot do plumbing is all very well, the real goal sought by those who want to achieve AGI is human-likeness in every significant respect. And a human incapable of conceptual thought would at the least be considered severely disabled, though still worthy of respect as a member of the human community.
The vital and provable distinction between perceptual and conceptual thought has been all but forgotten by AI researchers and the wider culture. But if we ignore it, and allow AI to take over more and more tasks formerly done by humans, we will surround ourselves with concept-free entities. This will be dangerous.
A good example of a powerful concept-free entity is a tiger. If you walk into the cage of a hungry tiger, all it sees in you is a specific perception: here's dinner. There is no reasoning over abstractions with a tiger, just a power struggle in which the human has a distinct disadvantage.
Aristotle restricted the term "intellect" to mean that part of the human mind capable of dealing with concepts. It is what distinguishes us from the other animals, and from every AI system as well. Try as they might, AI researchers will not be able to develop anything that can entertain concepts. And attempts to replace humans in jobs where concepts are important, such as almost any occupation that involves dealing with humans as one ethical being to another, can easily turn into the kind of hungry-tiger encounter that humans generally lose. Anyone who has struggled with an AI-powered phone-answering system to gain the privilege of talking with an actual human being will know what I mean.
ARC may become the default IQ test for new AI prototypes vying for the title of AGI. But the concept symbolized by the acronym AGI is itself incomprehensible by AI. As long as there are humans left, we will be the ones awarding the titles, not the AI bots. But only if they let us.
Sources: Matthew Hutson's article "Can We Build a Better IQ Test for AI?" appears on pp. 34-39 of the October 2025 issue of IEEE Spectrum. I also referred to the Stanford Encyclopedia of Philosophy article on Aristotle. For a more detailed argument about why AI cannot perform conceptual thought, see the article in AI & Society "Artificial Intelligence and Its Natural Limits," by Karl Stephan and Gyula Klima, pp. 9-18, vol. 36 (2021).
No comments:
Post a Comment