Why AI needs to drop the word ‘intelligence’
The Turing test arrived in the brilliant Computing Machinery & Intelligence (1950), along with its nine defences, still an astounding paper that sets the bar on whether machines can think and be intelligent. But it’s unfortunate that the title includes the word ‘intelligence’ , as it is never mentioned in this sense in the paper itself. It is also unfortunate that the phrase AI (Artificial Intelligence) invented by John McCarthy in 1956 (the year of my birth), at Dartmouth (where I studied), has become a misleading distraction.
Binet, who was responsible for inventing the IQ (intelligence quotient) test, warned against it being seen as a sound measure for individual intelligence or that it should be seen as ‘fixed’. His warnings were not heeded as education itself became fixated with the search and definition of a single measure of intelligence – IQ. The main protagonist being Eysenck and it led to fraudulent policies, such as the 11+ in the UK, promoted on the back of fraudulent research by Cyril Burt. Stephen Jay Gould’s 1981 book The Mismeasure of Man is only one of many that have criticised IQ research as narrow, subject to reification (turns abstract concepts into concrete realities) and linear ranking, when cognition is, in fact, a complex phenomenon. IQ research has also been criticised for repeatedly confusing correlation with cause, not only in heritability, where it is difficult to untangle nature from nurture, but also when comparing scores in tests with future achievement. Class, culture and gender may also play a role and the tests are not adjusted for these variables. The focus on IQ, a search for a single, unitary measure of the mind, is now seen by many as narrow and misleading. Most modern theories of mind have moved on to more sophisticated views of the mind as with different but interrelated cognitive abilities. Gardener tried to widen its definition into Multiple Intelligences (1983) but this is weak science and lacks any real vigour. It still suffers from a form of academic essentialism. More importantly, it distorts the filed of what is known as Artificial Intelligence.
Drop word ‘intelligence’
We would do well to abandon the word ‘intelligence’, as it carries with it so much bad theory and practice. Indeed AI has, in my view, already transcended the term, as it gained competences across a much wider sets of competences (previously intelligences), such as perception, translation, search, natural language processing, speech, sentiment analysis, memory, retrieval and other many other domains.
Turing interestingly anticipated machine learning in AI, seeing the computer as something that could be taught like a child. This complicated the use of the word ‘intelligence’ further, as machines in this sense operate dynamically in their environments, growing and gaining in competence. Machine learning has led to successes all sorts of domains beyond the traditional field of IQ and human ‘intelligences’. In many ways it is showing us the way, going back to a wider set of competences that includes both ‘knowing that’ (cogntitive) and ‘knowing how’ (robotics) to do things. This was seen by Turing as a real possibility and it frees us from the fixed notion of intelligence that got so locked down into human genetics and capabilities.
Other formulations of capabilities may be found if we do not focus on the anthropomorphic view of intelligence and learning but rather competences. The word ‘intelligence’ has too much human import and baggage. It makes man the measure of all things, whereas, it is clear that computer power has already transcended our brain in some areas, in terms of storage, exact recall, uploading, downloading, mathematical calculations, chess, driving cars and so on. The whole point of Google search, is not to be like us. It's better than us, with a greter memory, better serch and faster recall. Self-driving cars do not drive like us, they drive better than us. The whole point of the self-driving car is NOT to drive like us, as we kill 1.5 mllion people a yer while driving.
The great Turing did in fact explain that his test was an attempt to transcend the religious idea of man as the measure of all things but his test remains rooted in ’human-all-too-human’ abilities. Searle (1980) rightly criticised Turing’s approach with his Chinese Room argument, where the executor of the scripts in a translation task need know nothing at all about the actual meaning of Chinese or English, yet pass the test i.e. they do not ‘think’. Although this critique, in my view, makes a fundamental error.
Haugland (1997) questions the very idea that you need meaningful understanding of the meaning of Chinese and English at all. Searle seems to be demanding a human, gold-standard of understanding, self-awareness and intelligence. If we free ourselves from the tyranny of human ‘intelligence’ to general problem solving, the problem is no longer a problem.
Let’s take this idea further. Koch (2014) claimed that ALL networks are, to some degree ‘intelligent’. As the boundary for consciousness and intelligence changed over time to include animals, indeed anything with a network of neurons, he argues that intelligence is a property that can be applied to any communicating network. As we have evidence that intelligence is related to networked activity, whether these are brains or computers, could intelligence be a function of this networking, so that all networked entities are, to some degree, intelligent? Clark and Chalmers (1998) in The Extended Mind, laid out the philosophical basis for this approach. This opens up the field for definitions of ‘intelligence’ that are not benchmarked against human capabilities or speciesism. If we consider the idea of competences residing in other forms of chemistry and substrates, and see algorithms and their productive capabilities, as being independent of the base materials in which they arise, then we can cut the ties with the word ‘intelligence’ and focus on capabilities or competences.
Wonderful as the brain may be, as the organ that named itself and created all that we are discussing, it is a notoriously odd thing. It takes over 20 years of solid educational instruction to turn it into a remotely useful employee or member of society. It is famously inattentive, forgets most of what you teach it (Ebbinghaus -1908), is sexist, racist, full of cognitive baises (Kahneman -2011), sleeps 8 hours a day, can’t network, can’t upload, can’t download and, here’s the fatal objection - it dies. This should not be the gold standard for intelligence, as it is an idiosyncratic organ that evolved for circumstances other than those we find ourselves in.
We may even be able to move away from that other anthropomorphic obsession - consciousness. Daniel Dennett (1995) in Consciousness Explained, saw it as an epiphenomenon, not necessary for the explanation of actual competence and action. If we can drop the ghost in the machine, the machine itself can be seen as being capable, in a non-anthropomorphic sense. Psychologically, Kahneman (2011) in Thinking Fast and Slow distinguishes between the deliberate, slow rational and logical System 2 and the fast and instinctive System 1. He links unconscious intuitions with conscious decision making but the most interesting facet of his work is that we put too much faith in in human intelligence and judgements. Incidentally, he sees both systems, as being full of cognitive biases. Far from AI offering a world of bias it may be that AI can free us from the world of human biases towards more objective problem solving. The claim that all algorithms are biased, reduces statistics to a useless slogan. Sure there may be bias, but the brain is intrinsically biased and there’s little we can do about this.
If we move beyond brains, beyond inorganic versus inorganic, we can move as Harari (2016) in Homo Deux recommends, towards intelligence as fundamentally algorithmic and uncouple intelligence from humans and consciousness. His argument is that natural selection itself is algorithmic and gave rise to our species and brains, but these algorithms are independent of the substances in which they reside. We must therefore readjust our thinking around intelligence and learning to include wider definitions.
Beyond singular intelligences
Networks, such as the internet, have provided a new substrate, where collective intelligence is also possible, moving the concept of intelligences and capabilities on further. One brain cannot download directly from another or replicate knowledge and skills perfectly, in a fraction of a second. This is already happening in AI. In ‘Cloud robotics’ robots learn from experience but can also learn from each other, as they are networked. Experience and learning are therefore shared across the networked robots. Researchers in Google have already been teaching robots to learn skills in these three ways:
A. Learn motion skills (from direct experience)
B. Learn internal models (physics etc.)
C. Learn skills (human assisted)
In all three cases the learning is faster than one robot on its own with more variation in the learning experiences. So deep learning becomes not only possible from a mix of experience, models and being taught, but also by being pooled.
This concept of pooled learning and competences is something that drops the idea of a brain, human benchmark, or ‘subject’. Technology has become largely networked, which avoids our all too human tendency to hang on to ‘essentialism’. When it comes to our human abilities and what we regard as unique, we often invoke qualities such as ‘intuition’ or ‘thinking’ and ‘consciousness’. Turing opened up the possibility of “imaginable digital computers“ that would perform astonishing feats of what we would call intelligence or learning without recourse to a brain, soul or irreducible quality such as consciousness. That is becoming a reality. AI is now challenging what it is to be human, intelligent, competent, to think, to learn. The challenges we face are not to mimic humans but to find solutions that we are incapable of thinking of and executing. When Deepmind played its Atari game, it shot round the edge of the blocks and attacked from above, something humans had never thought of. We don’t want flawed human performance. We want performance beyond our capabilities, that is better than human. Humans crash cars and aircraft, kill patients through misdiagnosis and wrong prescribing. We need technology that is better than us. We didn’t go faster by copying the legs of a cheetah, we invented the wheel.
Beyond cognitive computing
Turing clearly foresaw rapid advances in the power of computers and, in the long term, was visionary in his understanding of their potential capabilities. Remarkably, he successfully predicted that computers would have 1Gig of storage by 2000. However, the Turing test itself has been critiqued and is still a contentious area. He was wrong in assuming that “at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted”. Many underestimated the recent advances in technology, machine learning, reinforcement learning and deep learning, that have since allowed AI to do the things that were regarded as unlikely, if not impossible.
Beyond IBMs misleading marketing
But the category mistake is to measure all of this in terms of human performance. This is why IBM's Cognitive Computing is simply misleading. There’s little that’s ‘Cognitive’ about it. That’s not to say that Watson is not useful. I use it myself in WildFire. AI has become much more powerful, with considerable advances in machine learning and in the practical application of such advances. This is partly to do with advances in AI techniques but also technical advances, which Turing predicted, and the rise of the internet and massive data sets. Few dispute the impact of AI on tasks, not just in the automation of manufacturing through robotics but also its impact in what has been seen as the cognitive domain. We just don’t need the language of human cognition to make progress - as that's a lie.
Few would argue that AI has progressed faster than expected, with self-driving cars and significant advances in machine learning, deep learning and reinforcement learning. In some cases the practical applications clearly transcend human capabilities and competences. We don’t need to see ‘intelligence’ at the centre of this solar system. The Copernican move is to remove this term and replace it with competences and look to problems that can be solved. The means to ends are always means, it is the ends that matter. What is wonderful here is the opening up of philosophical issues around agency, autonomy and morality. We are far from the existential risk to our species that many foresee but there are many more near-term issues to be considered. Ditching old psychological relics is one. Artificial smartness is with us it need not be called 'intelligent'.
Bloom (1956). Bloom's Taxonomy of the Cognitive Domain.
Dennett, D. (1995). Consciousness Explained.
Clark and Chalmers (1998) The Extended Mind
Dreyfus, H., & Dreyfus, S. (1997). Why Computers May Never Think Like People. Knowledge Management Tools, 31-50.
Ebbinghaus, H. (1908). Psychology: An elementary textbook. New York: Arno Press.
Gardner, H. (1983) Frames of mind: The theory of multiple intelligences, New York: Basic Books.
Frey B.C. Osborne M.A. (2013). The Future of Employment, Oxford Martin School.
Harari, Y.N. (2016). Homo Deus: A Brief History of Tomorrow. Harvill Secker, London.
Haugland, J. (1997). Mind design II: Philosophy, psychology, artificial intelligence. Cambridge, MA: MIT Press.
Kahneman, D. (2011). Thinking, fast and slow. New York: Farrar, Straus and Giroux.
Koch (2014). "Is Consciousness Universal". Scientific American Mind.
Searle, J. (1980). Minds, Brains and Programs. The Behavioral and Brain Sciences.3, pp. 417–424. (1980)
Susskind, R., & Susskind, D. (2015). The Future of the Professions. Oxford: Oxford University Press.
Turing, A. M. (1950). I.—Computing Machinery And Intelligence. Mind, LIX(236), 433-460.