Zen and the Art of Artificial Intelligence
Can an AI ask the right questions to help us be truly creative?
This is a rough transcript of my talk Zen and the Art of Artificial Intelligence at ‘I’ll Be Back South West’ on 17 September 2019. The slides are embedded at the end of the article.
I’m going to spend the next 15 minutes talking to you about Zen And The Art of Artificial Intelligence. I’d love to tell you this talk came from a deep appreciation of Zen Buddhism, detailed knowledge of AI and machine learning, and a healthy interest in creativity. However, the truth is a little more prosaic: After I’d agreed to the talk tonight, the title just randomly popped into my head a month or so ago, and since then I’ve been desperately trying to crowbar some kind of talk behind it.
Hopefully, I’ve succeeded.
But not only that, hopefully by asking myself a seemingly abstract question - ‘can I find something interesting from looking at AI through the framework of Zen Buddhism?’ - I’ve been forced to come up with a POV that’s new, surprising and valuable, the very definition of creativity.
Which leads nicely into the subject - asking abstract questions.
Most general understanding of Zen includes the concept of Zen koans, short questions of stories which are difficult, sometimes impossible, to answer. What is the sound of one hand clapping? Can a submarine swim? That kind of thing.
The purpose of these questions is not just to confuse the recipient, but to force you to think in a different way, and to be able to move forward without necessary having a precise destination.
Now, that seems to me like a good technique for creativity in business. Many business and marketing books are written on the importance of trying to reveal the underlying question or challenge a brand has to solve. I’m sure most of you would agree with me if I said it’s critically important to know the problem you’re solving before figuring out what the solution should be.
On top of that, we all know a lot of great businesses of the last 10 years, Uber, AirBNB, et al. have all become successful by finding a new question to solve in an existing category.
If this is so important, can AI help us here? Can AI ask questions to prompt us to think in a different way? Could an AI open our minds with a Zen koan that makes us think more creatively (useful, new and surprising). The benefits would be great.
As far as I can see, the answer is yes and no. So far, so Zen.
One of the early successes in AI was a programme called ELIZA, developed by Joseph Weizenbaum in 1966. Weizenbaum was concerned with passing the Turing Test, i.e. convincing a human judge that they were talking to a human, not a machine, and he realised an easier way to do this, influenced by therapy, was to always ask open questions in response to any input.
So he programmed ELIZA to always respond with an open question:
Client: Men are all alike.
ELIZA: IN WHAT WAY?
Client: They’re always bugging us about something or other.
ELIZA: CAN YOU THINK OF A SPECIFIC EXAMPLE?
Client: Well, my boyfriend made me come here.
ELIZA: YOUR BOYFRIEND MADE YOU COME HERE?
Client: He says I’m depressed much of the time.
ELIZA: DO YOU THINK YOU ARE DEPRESSED MUCH OF THE TIME?
So, from a simple set of ‘if… then…’ rules, Weizenbaum created a programme which asked questions, was initially perceived by some users as a human, and even led to various stories of visitors to the lab getting angry as they thought they were talking to a scientist who was being purposefully abstract. (You can try ELIZA yourself here.)
However, as great as ELIZA was, it didn’t take too much time with it to understand that you were in a limited conversation and there was no real benefit to the questions.
On this front, things haven’t improved much since.
In the ‘90s, the Loebner Prize was set-up, partially due to ELIZA, to award a cash prize to a chatbot which could pass the Turing Test. A silver prize and $25,000 would be awarded to any programme which convinced half the judges it was human, and a gold medal and $100,000 for any that could convince all the judges. No-one has yet won a silver medal, let alone get close to gold.
So we can establish that AI, or in this case an algorithm, could ask probing questions, but they were too formulaic and therefore not that useful.
But what if we broaden our definition of questions? What if we extend to the idea that questions don’t have to be verbal and you can ‘question’ situations with your actions - i.e. testing the water - without having to specifically ask it?
Well, it turns out that AI has had a good crack at this too.
You’re probably all aware that for sometime Chess programmes can now beat Grand Masters and World Champions at Chess.
And given the title of this evening, I’m also guessing that a significant proportion of you know that more recently an AI system, AlphaGo, has beaten the leading players in Go, a game considered to be far more complex than Chess and less likely to be beaten by brute force.
The connection with questioning comes in in how AlphaGo did it. It asked questions via the medium of Go. In move 37 of the second game of four against Go World Champion Lee Sedol, AlphaGo placed a black stone on the line five steps in from the edge of the board. Everyone was shocked.
It turns out in Go, that in the early stages of the game, you only play stones on the outer 4 lines. Playing on the 5th is considered suboptimal and was ridiculed by the commentators at the time. But, you can guess what happened, right? AlphaGo’s unorthodox move set-up a strategic play that left Sedol open in the latter stages of that same game. AlphaGo won and went on to win the series 4-1, setting a new standard for AI.
What AlphaGo had done was to question the orthodoxy of the game. Up until that point, Go players and culture had found a ‘local maxima’ - what they believed was the high point of Go based on a set of rules and principles leading to optimum performance.
What AlphaGo had discovered was that there was in fact a higher level of performance, a higher peak, unlocked by what was considered a bad move. It exploited this and beat Sedol.
This one act has changed Go since, with players now routinely using the 5th line in the early stages of the game. No-one knows if it is the true peak, or there’s even another higher point, but for now, by questioning the standard ways of behaving, an AI has changed the world of Go.
‘Humanity has played Go for thousands of years, and yet, as AI has shown us, we have not yet even scratched the surface. The union of human and computer players will usher in a new era.’
Ke Jie, Chinese Go Champion
So, does that prove it? Can AI truly ask probing and creative questions, verbal or non-verbal, that can change the state of play within a game, or more importantly for us, for a brand?
Well, yes. But, no as well. Still Zen!
AlphaGo did ask an open question, but it was very much within the concrete frame of reference of the game of Go. This is far more rigid than most situations an organisation would find itself in, even if you include the rules of a category such as travel or finance.
Therefore, for AI to be useful to the creative process, it has to be able to operate in environments without specific rules or parameters to create something truly new. Can AI ask powerful questions to increase creativity - koans - in an unstructured environment?
For us to answer this, it’s useful to borrow some existing thinking on creativity as it pertains to structured and unstructured contexts.
Psychologist Margaret Boden has enhanced the accepted definition of creativity to take context into account, rather than one form of creativity, she suggests there are three types of creativity, COMBINATORIAL, EXPLORATORY and TRANSFORMATIONAL.
1. The first type is COMBINATORIAL, unfamiliar combinations of familiar ideas. The kind of creativity we often see in a lot of collage-based artwork, poetry, a lot of advertising. This is the kind of creativity we see demonstrated by our friends at Tiny Giant with their cupcakes and cocktails and curation. A human has asked - ‘can I use AI to use familiar elements to create an unfamiliar result?’ and machine learning or AI has answered positively. The human asks, the machine responds. Great stuff and very useful, but it doesn’t feature AI posed questions.
2. The second type of creativity is EXPLORATORY, where the existing styles of conventions are used to generate novel structures or ideas, whose possibility may or may not have been realised before the exploration took place. I.e. you stick within the rules, but push the limits as much as possible and go beyond the local maximum. This is the kind of creativity AlphaGo demonstrated when it beat Lee Sedol. It ‘questioned’ what was right and possible within the game of Go, and discovered a new peak of achievement.
Exploratory creativity is not to be sneezed at, most scientists, artists, advertising creatives etc. will only ever produce creative work of this type, and will achieve great things. Think of what has been achieved and will be achieved merely with the constraints of paint and paper and you can understand the scope of this.
Here, there is great scope for AI, not only in terms of answering questions set by humans but as we’ve seen with AlphaGo, posing new questions within the framework of a particular medium or context.
But is that all? Not quite.
3. The third and final type of creativity is TRANSFORMATIONAL where some deep dimension of the conceptual space is altered so that ideas and concepts can now be generated which could not be generated before, and which are not "all of a piece" with the previous style. Imagine a game of chess where the pawns can jump over pieces, or, shock-horror, a taxi company without any taxis like Uber, or a hotel chain without any hotels like Airbnb.
This is the domain of the most powerful ‘What if?’ questions, questions the answers to which fundamentally have shaken up categories and built and destroyed brands. Can AI help you here?
Unfortunately, there’s little evidence they can. As far as I can see, most examples of creative AI focus on exploratory creativity and don’t break the rules like this.
Even those on the inside of the AI revolution, like Judea Pearl, computer scientist and philosopher, best known for championing the probabilistic approach to artificial intelligence, think with our current techniques we’re not close to cracking this:
"Current machine-learning systems operate almost exclusively in a statistical, or model-blind, mode, which is analogous in many ways to fitting functions to a cloud of data points. Such systems cannot reason about 'What if?' questions and, therefore, cannot serve as the basis for Strong AI"
Judea Pearl, The Limitations of Opaque Learning Machines
And given there is a long history of the most powerful ideas being rejected at first; even if we did have an AI capable of producing them, how would the operator be expected to understand the value in the question? There’s something about the persistence and craziness of a human questioner and inventor which at the moment still seems to be the main foundation of this kind of thinking.
So, from our journey from Zen koans, we’ve seen that AI can ask questions, and that sometimes those questions can be useful and surprising, but only if their based within the constraints of a particular way of thinking.
For now, the only place we’re going to get the truly transformational creativity that results from the open questioning of things like Zen koans, is from ourselves and our peers, asking better questions far before we concern ourselves with creating better answers.
So, like the stereotypical Zen master who answers every question with another question of their own, the next time you begin thinking about how you solve a problem at work, maybe you should step back, look at yourself and your brand, and try to come up with a better question first.
As Prof. Luciano Floridi, puts it: ‘"Data do not speak by themselves, we need smart questioners".
(To see the slides in Google Slides click here.)
Creativity: How does it work?, Margaret Boden, 2007
Possible Minds - 25 Ways of Looking at AI, John Brockman (ed.), 2018
The Fourth Revolution: How the Infosphere is Reshaping Human Reality, Luciano Floridi, 2014
The Creativity Code - How AI is learning to write, paint and think, Marcus du Sautoy, 2019