Why robots are so dof at making art and music
Even without the awful prospect of a robot-governed dystopia, can machines ever truly learn to be creative?
Can a machine ever be genuinely creative? Or is a pale imitation of true creativity, achieved by following a few rules cleverly spiced with random surprises, really the best technology can hope to achieve?
It’s a question that prompts fear as much as fascination. If a computer broke free of its programmer and showed evidence of creativity, we might wonder whether it could conceive plans and projects of its own – and, if so, why should it be content just to follow human orders? But even without the awful prospect of a robot-governed dystopia, the idea that machines could be creative is an affront to our vanity. Creativity is, after all, the mark of the human, a trait shared with no other species.
That self-image may be about to take a severe blow. Artificial Intelligence (AI) has made giant strides, as ever-increasing computer power has been yoked to astonishing advances in other sciences, such as robotics. One of the most unnerving results will be unveiled next month.
It’s an exhibition of artworks by a very humanoid robot named Ai-Da, created by UK-based Engineered Arts. With a pencil gripped in her bionic fingers, she’s a dab hand at drawing her human subjects from life. According to Salaheldin Al-Abd and Ziad Abass, two of the lead AI engineers, the program “starts by analysing the person in front of Ai-Da, and moving on to create a virtual path for Ai-Da’s arm, [which] is fed into a path execution algorithm that gives real-space coordinates, enabling the arm to produce the actual sketches”.
This seems frankly no more than imitating an outline caught by a robot eye, but other recent attempts to get AI to paint appear to be more genuinely creative. In October, Christie’s made its first-ever sale of an artwork created with AI, a composite of various historical portraits made by a Paris-based collective called Obvious. Entitled Portrait of Edmond Belamy, it shows an oddly blurry 19th-century gent with a high forehead. It fetched the astonishing price of $423,500 (R5.9m), nearly 45 times its estimate – which proves, if nothing else, that algorithms rival humans in earning power.
In 2016 scientists at Microsoft and Delft University created a program that learnt to imitate, through pixels on a screen, Rembrandt’s genius for creating portraits that in some mysterious way capture the soul of their subjects. In total, 346 paintings were analysed by a “deep learning” algorithm, which sifted them to isolate the archetypal Rembrandt subject: a middle-aged man with facial hair, white shirt and a hat.
In this case, the machine, having educated itself via the algorithm, was left to complete the task. But what if the learning process were interactive, with humans and machine placed in constantly evolving dialogue? That is the idea behind artist Mario Klingemann’s piece Circuit Training, one of numerous fascinating examples of AI creativity at the forthcoming exhibition AI: More than Human at the Barbican Centre in London. The spectators, we’re told, “help to create the data set by allowing the AI to capture their image, then select from the visuals produced by the network, to teach it what they find interesting”. As the machine learns more about people’s preferences, so the artwork evolves. The AI setup has already produced streams of richly coloured and curiously distorted faces.
A different sort of AI creativity was recently revealed by the Chinese telecoms company Huawei, which involved teaching one of its smartphones how to compose well enough to complete the unwritten third and fourth movements of Schubert’s Unfinished Symphony. As Walter Ji, president of Huawei Western Europe, put it: “By analysing the timbre, pitch and metre of the existing first and second movements of the symphony, the AI predicted similar melodies based upon what it had learnt and was able to generate the melody for the missing music.” With a bit of help from film composer Lucas Cantor, the phone “composed” a full score, and the results were premiered in February.
On a technical level, all these things are a marvel, and a tribute to human ingenuity. But are the results actually any good?
The drawings of Ai-Da are careful but lifeless, as if the outline of the subject has simply been traced mechanically. The “Rembrandt” produced by the machine was certainly Rembrandt-like at a glance, but not on closer inspection. The “painting” is exactly what you’d expect it to be, a composite of all the Rembrandt portraits you’ve ever seen, with zero individuality and a peculiar deadness about the eyes.
As for the Schubert symphony completed by the smartphone, the results were lamentable. In the new third movement, the music lurches from one idea to the next, like a random walk, with weirdly awkward transitions. In the fourth movement, things go from bad to worse. Any sense of Schubert’s style is left behind. The orchestration swells to an anachronistic Tchaikovskian grandeur, and weird echoes of Smetana’s Vltava flit across the music – as well as constant near-quotations from the symphony’s first two movements, which is something Schubert is unlikely to have countenanced.
We might cheer at this point, and say that all this proves human creativity can never be recreated by a machine slavishly sticking to a set of rules. Ah, but AI isn’t just about following a set of rules, says Marcus du Sautoy. His brilliant and thought-provoking new book The Creativity Code is full of fascinating – and at times scary – examples of just how quick machines now are at mastering complex mental skills, which require exactly the kind of “thinking outside the box” that is the hallmark of creativity.
Du Sautoy tells the riveting story of how computers first learnt to play the deceptively tricky Chinese game Go. The grandmasters of the game sneered at the idea they could ever be beaten by a mere computer, but were soon humbled. Mathematics is another area where computers have already been able, via a “bottom-up” process of learning from what human mathematicians have achieved, to hit on new theorems that are both novel and genuinely interesting.
The essential message of du Sautoy’s book is that computers no longer have to stick to following a program. They can now learn from their mistakes, change tactics, and even take a bold leap into the unknown to try out an entirely new strategy – which is what allowed them beat the best human players of Go.
Yet it was just such a “learning” algorithm that was devised by the Huawei team to compose the missing movements of Schubert's Unfinished Symphony – so why was it such a failure? In this case, my feeling is that one cannot blame the machine. The weakness, as so often with AI failures, lies with the humans. To get an adequate grasp of Schubert’s style, the phone needed a much larger data set than the first two movements of the symphony alone could provide. It should have scanned at least two or three more symphonies, and possibly the late piano sonatas.
The strange thing about the Huawei symphony is that it was so much worse than other attempts to get AI to compose music that date back 20 or even 30 years, when the technology was far less sophisticated. I can remember back in the 1980s listening to really quite decent chorales harmonised in the style of J S Bach, composed by a program developed by AI scientist Kemal Ebcioglu. But then, a chorale is far less complex than a symphony, and much easier to compose by following a set of rules.
A pattern is starting to emerge. In those forms of creativity where a human’s sensuous apprehension of the world can plausibly be reduced to data, AI does (fairly) well. The data can be scanned for recurring patterns or formulas, which are after all the defining features of many art forms, from the sonnet to the symphony. But what about those aspects of creativity where that sort of reductionism is utterly implausible, such as coining a vivid metaphor, or writing a novel?
The reasons why such tasks are the ultimate challenge for computers take us to the heart of why genuine creativity has thus far eluded AI, and I suspect always will. Computers can only generate one side of language, the side we call “syntax” – the thing that structures words into sentences. The other side, “semantics” – the thought of “chair”, for instance, that arises in our minds when we read the word “chair” – will always elude it, because there is no “mind” in a computer.
This is why machine attempts to create literary works are really just disguised pattern shuffling. AI is quite good at creating short poems in the sort of modernist idiom where meaning frankly doesn’t matter. The Cybernetic Poet created by futurologist Ray Kurzweil can churn out endless examples of obscurity, in any combination of styles. Here, for example, is a haiku in the style of Walt Whitman: “Ages and pink in Sex, Offspring of the voices of all my Body.” But anything requiring a determinable relationship to the real world soon defeats it.
The inability of computers to grasp semantics is just another way of saying that machines lack the essential attribute to be creative, namely consciousness. Some, like Du Sautoy, hope this “ghost in the machine” will be recreated. I think this is a vain hope, because of something even more fundamental that we possess but machines don’t.
Human beings are not just “brains in a box”. They are also living creatures, with urgent needs for food, shelter and sex. They live in constant interaction with other similar creatures, in a fabulously complex, socially constructed world. All this gives them a rich inner life, swarming with desires, feelings, thoughts, plans and dreams.
A computer, not being an organism, has none of this. It has no desires, no inner life, no regrets for the past, no hopes for the future. It therefore has no need for art, or indeed creativity of any kind.
So rather than wear ourselves out in the fruitless quest to recreate creativity in machines – a project in which machines will only ever be our passive slaves, never our eager collaborators – why not cherish and support it in the only place where it can ever be truly found – in those fallible, imperfect, but wonderfully inventive and surprising things called human beings.– © Telegraph Media Group Limited (2019)