Can an AI program have an IQ? In other words, can the IQ of an AI program be measured? Like how humans can do an IQ test.
-
1$\begingroup$ If by IQ you mean pattern-finding (which is generally what the measurement of IQ is interpreted as), then computer programs can find patterns but it usually takes quite a while as they have to be evolved to find said patterns, but I am not sure if there exist AIs that can pattern find 'within one lifetime' (like humans can), there might be someone w ill have to update me and the asker on that. Also note that IQ tests were designed for humans to find them easy, so a machine might have a hard time even understanding what it 'sees'. $\endgroup$– god of llamasCommented Aug 2, 2016 at 16:14
4 Answers
Short answer: No.
Longer answer: It depends on what IQ exactly is, and when the question is asked compared to ongoing development. The topic you're referring to is actually more commonly described as AGI, or Artificial General Intelligence, as opposed to AI, which could be any narrow problem solving capability represented in software/hardware.
Intelligence quotient is a rough estimate of how well humans are able to generally answer questions they have not previously encountered, but as a predictor it is somewhat flawed, and has many criticisms and detractors.
Currently (2016), no known programs have the ability to generalize, or apply learning from one domain to solving problems in an arbitrarily different domain through an abstract understanding. (However there are programs which can effectively analyze, or break down some information domains into simpler representations.) This seems likely to change as time goes on and both hardware and software techniques are developed toward this goal. Experts widely disagree as to the likely timing and approach of these developments, as well as to the most probable outcomes.
It's also worth noting that there seems to be a large deficit of understanding as to what exactly consciousness is, and disagreement over whether there is ever likely to be anything in the field of artificial intelligence that compares to it.
The other answers are correct that machine IQ test results are currently not indicative of machine intelligence. One of the surprising facts of human intelligence is that performance on almost all cognitive tasks are correlated with each other; that is, there is such a thing as 'general smartness' and IQ tests attempt to measure that thing.
People have built programs that take IQ tests, however, and some of them perform quite well. Raven's Progressive Matrices, a visual pattern recognition IQ test, is an easy target for AI (see this paper as representative) and another group has constructed an AI that performs about as well as a 4 year old on the verbal intelligence portion of a standard childhood IQ test.
It all depends of what your A.I. can do. Even humans cannot do everything.
If your AI program is so smart, ask it to take the general IQ tests for humans. Because the real IQ tests are made of several questions from different areas, so in that way you can measure IQ of your AI.
This is because the IQ means the tests which are designed to assess human intelligence.
An intelligence quotient (IQ) is a total score derived from one of several standardized tests designed to assess human intelligence.wiki
So there is no any other way of measuring IQ without taking IQ test, otherwise it won't be IQ (very logical).
If your program is not so smart, you should look for specific tests related to the expertise or problem being solved. Ideally let it compete with humans who has the same expertise in that area, but it's important make the test on the same ground/level.
For example the intelligence of Deep Blue project was measured by playing chess with Kasparov. Then if world champion cannot win the game, who will?
If you're writing program to play a game, make it play with compete with humans and measure the intelligence in terms of score.
The equivalent of IQ for AI is a Turing Test (like MIST and other), see:
-
$\begingroup$ Good answer, generally AI is designed for a specific purpose and measured in a specific way. $\endgroup$ Commented Aug 3, 2016 at 9:15
All these questions are fully covered in the book The Measure of All Minds Evaluating Natural and Artificial Intelligence (Orallo, 2017)
An excerpt from the description give a good overview:
By replacing the dominant anthropocentric stance with a universal perspective where living organisms are considered as a special case, long-standing questions in the evaluation of behavior can be addressed in a wider landscape. Can we derive task difficulty intrinsically? Is a universal g factor - a common general component for all abilities - theoretically possible? Using algorithmic information theory as a foundation, the book elaborates on the evaluation of perceptual, developmental, social, verbal and collective features,