Think AIs Are More Intelligent Than You? You Might Be Overestimating Them
Many will likely think that artificial intelligence is smarter than human intelligence. This is because artificial minds tend to be more precise, more organized, and capable of processing more information in a more efficient manner. However, you will probably be surprised to learn that one AI system did not do impressively well in the famous IQ test designed to assess human intelligence. Artificial Intelligence systems might be excellent in playing chess, handling mathematical problems, or doing pattern recognition but they are still far from being as intelligent as humans.
If your idea of artificial intelligence is the likes of Aya of the DC Comics, Sonny of I, Robot, Auto of Wall-E, or Skynet of the Terminator films, consider throttling your expectations for now. Humans have not achieved those levels of artificial intelligences yet. In fact, one AI intelligence test study published on Cornell University LIbrary’s ArXiv shows that some AI systems have IQs that are considerably lower than those of human adults.
IQ Testing Artificial Intelligence
A team of researchers from the University of Illinois at Chicago and a group of artificial intelligence researchers in Hungary conducted a standard IQ test on a specific artificial intelligence system to determine the extent of the capabilities of AI in relation to human intelligence. The study was presented in a paper that is now posted on arXiv entitled “Measuring an Artificial Intelligence System’s Performance on a Verbal IQ Test for Young Children.” Team members include Stellan Ohlsson, Gyorgy Turan, Robert Sloan, and Aaron Urasky who are experts in computer science, statistics, and psychology./
The team administered the VIQ (Verbal IQ) part of the Wechsler Preschool and Primary Scale of Intelligence (WPPSI-III) to the AI system called ConceptNet. The test questions were translated to inputs comprehendible to ConceptNet using a combination of the simple natural language processing tools of ConceptNet and short Python programs written by the team. The setup made use of a question answering system that employed a version of ConceptNet based on spectral methods.
Before proceeding to the results of the study, let’s get to know the AI system called ConceptNet. The specific version used in the AI test is ConceptNet 4, not the latest version but still a viable option for AI IQ testing. ConceptNet 4 is an artificial intelligence system developed at the Massachusetts Institute of Technology. Its foundation is a semantic network that contains a huge store of information tapped to “educate” the system about concepts. It creates the ability to process the relationship between and among information.
The results of the IQ test conducted on ConceptNet 4 showed that its intelligence can be compared to that of average IQ of a toddler around four years of age or the intelligence of a 5 to 7-year-old with below average IQ. In the exact words of the researchers, “the performance of the system fell when compared to older children, and it compared poorly to seven year olds.”
ConceptNet 4 scored well on the vocabulary and similarities sub-tests. It had an intermediate score for the information sub-test. Its results for the comprehension and word reasoning sub-tests, however, were on the low side. This is rather disappointing because the word reasoning and comprehension sub-tests are considered as the primary determinants of common sense. In short, ConceptNet appears to have some passable verbal skills (for a toddler) but is apparently lacking when it comes to common sense.
However, the researchers clearly asserted in their paper that the WPPSI-III VIQ results they obtained, with emphasis on the large variations in the scores for the different sub-tests, DO NOT SHOW that ConceptNet 4 has the verbal skills of a four-year-old child. The researchers want to emphasize that they are using IQ tests for children simply to have an objective basis as they evaluate and compare artificial intelligence systems. It’s virtually impossible to find a four-year-old with the IQ results similar to that of ConceptNet 4’s. Based on the IQ test results, ConceptNet 4 is like an odd little child with excellent memory but without the ability to make rational judgments or the capacity to draw sensible conclusions based on the vast amounts of information it has in its mind. It’s like a child who impresses with his knowledge on the freezing temperature but without the common sense to realize that ice is cold.
If it’s going to make you feel any better, yes the MIT-created AI system ConceptNet 4 is remotely close to your level of intelligence. However, this test is just one test. Back in June this year, a similar effort to test the IQ of an AI system was included in the ArXiv database. In that test the AI outscored humans. This test, however, has not been published in an authoritative peer-reviewed scientific journal yet. There will surely be endless debates as to whether or not machines have already become more intelligent than humans but one thing is certain – IQ is not the be-all measure of intelligence.
IQ tests have been widely believed as a credible source for intelligence metrics. However, it has to be pointed out that IQ tests can be hacked or tricked to make an artificial intelligence system excel in them. The aforementioned AI system that outpointed humans in the IQ test, for instance, was specifically built to handle the verbal questions of IQ tests using the approach called deep learning. It is not necessarily more intelligent than humans but it scored well in the IQ test because it is exactly what it is designed to do.
Until a definitive metric for intelligence is established, it is safe to say that most people are overestimating the intelligence of AI systems. Humans can still claim to be more intelligent than machines unless someone can offer a clear evidence to prove the contrary.