The General Intelligence of Robots
“Artificial intelligence” researchers mean by the phrase “general intelligence” two things. One is simply a machine learning system that can, like humans, use its knowledge for many different tasks. This is not something that has currently been built; language models produce language, stable diffusion models produce images, and neither seem to have any concepts of the underlying realities these words and images reflect. The second is the “general intelligence,” g, derived from intelligence tests, which is believed by some to be a unifying reality behind intelligence test scores. In a paper on GPT-4 from Microsoft a Wall Street Journal Editorial by Linda Gottfredson, Mainstream Science on Intelligence is cited as providing a definition of general intelligence. It’s an appalling piece, repeating debunked claims about racial differences in intelligence and claiming as scientific consensus hypotheses that are at best debatable and at worst outright false.
So what is this “general intelligence” that is discussed in Gottfredson’s editorial? It’s an idea from pioneering psychologist and statistician Charles Spearman1. He did a series of tests of mental abilities and, using the techniques he developed, analyzed the results of those tests in a way which made the main unifying factor of those tests a factor he named general intelligence. To Spearman, English and very much committed to the intellectual ranking of individuals, it was plain this represented a sort of mental energy. This was criticized by L. L. Thurstone, who, relying on the same data, argued for a model of intelligence based on seven primary mental abilities and, again, to Thurstone, American from Chicago, this seemed an obvious model.
Unfortunately, for any given set of data, factor analysis can be used to construct arbitrarily many models. To know which, if any, model is accurately reflects the underlying system one needs to study the system the data is derived from. All this research was done long before modern neuroscience existed, before CT and PET scans – both Spearman and Thurstone were arguing ahead of their facts. Thurstone at least was aware of the problem. He commented 2:
The exploratory nature of factor analysis is often not understood. Factor analysis has its principal usefulness at the borderline of science… Factor analysis is useful, especially in those domains where basic and fruitful concepts are essentially lacking and where crucial experiments have been difficult to conceive. The new methods have a humble role. They enable us to make only the crudest first map of a new domain.
The idea of general intelligence is much-loved by racists, who very badly want to prove that whites are superior to blacks, and point at intelligence tests results to “prove” this, ranking the average g of whites above that of blacks. Never mind that the numbers are averages and there’s overlap between low-score whites and high-score blacks. Never mind that intelligence test raw scores have been rising over time, and blacks have closed the gap3. The rise in raw scores, known as the Flynn effect, indicates that intelligence test scores are heavily influenced by environmental factors, which makes the whole idea of inherent racial (or any other group) superiority nonsense.
It also makes g useless in evaluating an artificial intelligence. If one is going to measure the general intelligence (if this even exists) of an artificial intelligence, which year’s g does one compare it to? What environmental factors should be controlled for? The environmental factors that influence the g of an artificial intelligence cannot be the same as those that influence the g of a human intelligence.
The idea of a single factor that determines most of intelligence is twisting artificial intelligence research out of shape. We do know, based on neuroscience, that different brain regions have different functions and language is localized in particular regions. Large language and stable diffusion models replicate the function of parts of the brain but, if you believe that intelligence is unitary as believers in general intelligence do, and you have a persuasive large language model, it is easy to believe that expanding these models will replicate all the functions of a human brain. But human intelligence is demonstrably, physically, not unitary. Different regions of the brain have different functions.
Artificial intelligence researchers, if they are serious about their goal, need to study neuroscience. They also need to understand both the psychological and neurological fields better. People who know intelligence measurement know that Gottfredson’s editorial is part of a long-running dispute over the nature of intelligence and that her side of the dispute is racist. They need to pay attention to their own mathematical knowledge as well. It is well understood that arbitrarily many models may be constructed from a given set of data and mathematically trained computer scientists ought to know this, probably would be able to explain this if asked about it in the abstract.
The AI hypemeisters are selling hype.
To AI promoters and researchers: check out your citations. You really don’t want to be citing a scientific racist or a discredited model of intelligence that has no grounding in actual psychology and neuroscience. Second, learn from the psychologists and neuroscientists. Third, cool the hype. You haven’t created intelligence. You’ve replicated perhaps a part of human cognition, you haven’t copied the whole. Don’t let confirmation bias and greed make your decisions for you.
That’s for the AI hypemeisters and researchers. For the rest of us, don’t believe the hype and don’t let machine learning technologies destroy the lives and careers of humans. Especially, don’t let it destroy culture. Current machine learning technology is not creative except by accident. If it were creative, the basis of its experience would not be human experience. If popular art and literature is taken over by machine learning technology, there will be no new ideas. It is also inherently dishonest, since it does not know what truth is. We do not want to see massive amounts of fake journalism, ungrounded in reality. A future in which art and literature are dominated by machine learning models would be stifling. I hope it does not come to pass.
Some AI researchers seem to have fallen down the rathole sometimes physicists sometimes fall down: they sincerely believe they know more than everyone else and make grandiose claims based on that belief. So you get people like physicist Richard Muller, who started by saying that climate scientists were all wrong and he knew more than all of them and ended up agreeing with them. Add to this the hype of computing technology and you get – what? A lot of money and actual physical energy spent on something which seems ultimately to be unproductive if not anti-productive. Surely these resources could be used to better purposes?
This whole matter is discussed in depth in Gould, S. J. (1996). The Mismeasure of Man (Rev. and expanded). Norton. ↩︎
Cited by Gould, Gould, S. J. (1996). The Real Error of Cyril Burt: Thurstone on the Uses of Factor Analysis. In The Mismeasure of Man (Rev. and expanded). Norton. (Ebook, physical page number not known.) ↩︎
Neisser, U. (Ed.). (1998). The Rising Curve: Long-Term Gains In IQ And Related Measures (1st ed.). American Psychological Association. ↩︎
Comments
Post a Comment