Marcus du Sautoy is the kind of science writer who cares more about questions than answers. In his books he tackles “unsolved problems”, “number mysteries” and “the great unknown”, topics at the edge of human understanding. They are subtitled with words such as “odyssey”, “exploration” and “journey”. But Du Sautoy is a flaneur: his trips are not motivated by destinations. This is both the main strength and flaw of The Creativity Code, a wide-ranging and fact-packed tour d’horizon of current applications of artificial intelligence in mathematics and the arts.
So much of what is written about AI is either hype or scepticism about the hype. Ada Lovelace, the first computer programmer and perhaps the first AI-hype sceptic, wrote that “it is desirable to guard against the possibility of exaggerated ideas… the [computer] has no pretensions whatever to originate anything”. Her disclaimer was well taken in 1842. But, Du Sautoy asks, can computers be creative in today’s age of machine learning? To make his question sharp, he proposes a test for machine creativity, “the Lovelace test”: a machine must produce something “new, surprising and valuable”, and do so in such a way that its programmers are unable to explain the feat. As he sees it, the creative heavy lifting is often being done by the programmer, and not the program itself Du Sautoy’s discussion of computer creativity in the arts is fascinating but the computer art itself is underwhelming. An attempt to create a Rembrandt-esque portrait through data analytics effectively produces a forgery done without a paintbrush, and is, it turns out, a Microsoft PR gimmick. A superior algorithm can improvise a continuation of a jazz riff. A jazz musician who tested the algorithm is impressed, admitting that “it is years ahead of me, yet everything it plays is unquestionably me”. But the program is based on a “memoryless” random process (a “Markov process”) and so its improvisations lack any global structure. In the weeks since Du Sautoy went to print, Elon Musk-backed OpenAI has developed a cutting-edge storytelling algorithm that goes a step change beyond the text-generation programs in the book. Its release has been accompanied by a predictable hype cycle, and reminds us that progress in AI is rapid and continuous. But beneath the froth is a statistical model that of course lacks any understanding and therefore mentions “fires happening underwater”. While most of the chapters are given to a discussion of computer art, The Creativity Code is at its best when Du Sautoy, a mathematics professor at Oxford University, discusses his own subject. His analysis goes deep when he forges a rich analogy between the game of Go and mathematical proof. He is justifiably impressed with Google DeepMind’s superhuman Go engine and is intrigued by the potential of its algorithm to be repurposed as an automated theorem prover. The analogy interestingly breaks down because in mathematics, unlike in Go, “valuable” is immeasurably harder to code than “new” or “surprising”. In Go, any winning state of the board is valuable, while in mathematics valuable theorems – for example, those that are aesthetically pleasing or have scientific applications – are rare among the infinitude of all possible theorems. For Du Sautoy, today’s maths AIs are idiot savants, capable of generating original proofs but lacking the taste to discriminate which are important. Refreshingly, Du Sautoy is open about his doubts when evaluating the new AI. He asks: “Why do I still feel that anything to match human creativity is still way beyond the reach even of these amazing new tools?” At times, he casts AI as “the ventriloquist’s dummy providing the mouthpiece for our urge to express ourselves”. As he sees it, the creative heavy lifting is often being done by the programmer or the audience, and not the program itself. A strength of his analysis is that he manages to find a middle way between hype and anti-hype. Alan Turing asked: can machines think? He then replaced that question with his Turing test, before adding that it was “too meaningless to deserve discussion”. Perhaps he felt, as Dijkstra did, that asking if machines can think is like asking if submarines can swim: the answer rests on how far we choose to extend a metaphor from the biological to the artificial world. Unlike Turing, Du Sautoy is captivated by his original question (can machines be creative?) and seems to mistrust his own test (the Lovelace test). If he reaches a conclusion it is that consciousness is necessary for creativity, and so the behavioural Lovelace test is insufficient. This is not in itself an unreasonable claim, but it naively intrudes on a prominent body of psychology that attempts to carefully isolate what function, if any, consciousness fulfils. It is also a conclusion that was available to Du Sautoy from the armchair, before he embarked on his enjoyable, circuitous journey: it turns out he didn’t need AI to reach his destination after all. • The Creativity Code by Marcus du Sautoy is published by 4th Estate (£20). To order a copy go to guardianbookshop.com or call 0330 333 6846. Free UK p&p over £15, online orders only. Phone orders min p&p of £1.99