[audience clapping]
Well, it was talked about the idea of fear.
And I think in society AI has produced a lot of fear,
especially the idea of,
are we gonna have a job in the future?
AI seems to be able to do so many things,
but I think there's one thing that we do regard
as uniquely our human domain.
And it's something that is expressing what it means
to be human and that's creativity.
But I think already the exhibition
that we're seeing here at the Barbican,
there are hints
that AI may be even able to to creep into our world
of the creative world.
And I think that for me, I really saw a sea change
that happened in that very famous story a couple
of years ago now when AlphaGo managed
to beat the world's best player at this game,
ancient game of Go.
Developed by Deep Mind, it forms the centerpiece
of the exhibition here at the Barbican
looking at the interaction of AI on the arts.
This piece of code learnt on way humans played the game,
then played itself, played, made synthetic games
that it learned from and was able to achieve
such a high level that it beat Lee Sedol
four games to one.
Lee Sedol regards that one game
that he did win as the most valuable game
of his whole career.
Now maybe not so extraordinary
that we've got computers doing things at such a high level.
We've already seen chess played
by computers in the nineties,
but I think something significantly different happened
with the AI playing this game.
The game of Go is one that requires
a lot of intuition to play.
When you talk to a Go player,
they're often not very clear
why they're making particular moves.
There's a lot of pattern recognition
as the stones build up on the board,
something the human brain is very good at reading
into patterns, but not perhaps articulating why
they are seeing patterns there.
It's a game which involves a lot of creativity.
And in the past, computer science regarded this
as a game which you could not code up.
But because this idea of not having
to do code from a bottom up manner,
but building the code from the bottom, from the bottom,
being able to learn how to play the game,
failing and from failure learning to do new things,
this code developed into this very powerful machine.
But there was something very significant, I think,
which happened here, which marked a phase change in the way
AI is approaching kind of problem solving.
And that happened in game two on move 37.
Move 36, Lee Sedol placed a white stone on the board,
went to the top of the hotel
that they were playing in Korea.
He needed cigarette breaks.
AI currently doesn't need nicotine for stimulation.
The AI sat there, thought for a while,
and then asked the human player,
this wasn't an exercise in robotics,
we still find it quite difficult
to make something which will pick a stone up
and put it very nicely on the board.
Asked the player,
the human player
to place a stone on the fifth row in.
All the commentators that were watching this gasped.
Because traditionally your Go master teaches you
that early on in the game,
you only play on the first four rows in,
that that's where the early competition goes
on for territory,
on the edges and on the kind of starting
to creep into the board.
And it's regarded,
if you play on the fifth row this early on
as an incredibly weak move.
And I remember watching these obsessively on YouTube.
I was going through a bit of an existential crisis
'cause I regard this game as very similar
to doing mathematics.
And if code could play this game,
maybe my job was going to be under threat.
And I remember the moment the commentators
just gasped and said, Well it's made a huge mistake.
Lee Sedol will be able to win this game.
And when Lee Sedol came back down from the roof,
he too looked at this move
and just couldn't understand why it had played
such a weak move.
Yet, as the game went on
and more territory built up
from the bottom right hand corner,
it turned out that that move 37 was crucial
in AlphaGo winning this game.
It controlled the territory
because of that early play of that black stone.
And for me, this passed three qualities
that I think we should be looking for
in something that we will regard as creative.
I was on a committee, the Royal Society,
looking at the impact of machine learning on society.
Demis Hassabis, the mind behind AlphaGo was there,
but there was also a philosopher, Margaret Boden,
and she's been thinking a lot about what the computers
or what she calls a tin cans might be able to achieve.
And she has a nice working definition
of what she regards as creative.
And I think it's quite useful to have in mind
as we go forward thinking about creativity this afternoon.
Creativity should be something which is novel,
new or computers can easily make new things
and that's quite objective.
We can judge that.
But it's these other two qualities,
surprise and value,
and of course surprise
and value are much more subjective.
Value for one person, I might write a poem
and value it very highly,
but nobody else values it.
But machine learning is gonna enable an AI to actually learn
what we find surprising and what we value.
And I think in that game too, move 37, we saw something new.
We saw something that surprised the commentators.
In the confines of a game,
it's very easy to judge something that has value.
And I think what's exciting is we've seen already
how AI can push our creativity into new realms.
And this is certainly what's happened with the game of Go.
We were playing this game in a very,
what we thought was an optimal way,
only play on the first four rows in early on in the game.
But the AI has shown us
that there are new ways to play this game.
What we thought was the optimal way
turns out to just be what we call in mathematics
a local maximum.
If you take a risk and explore other territory,
you might find an even better way to play the game.
I think we behave very much like machines quite often
and we need something to push us out
of our mechanical ways of thinking.
And this is what happens certainly in the game of Go,
which is why I think that that moment
a couple of years ago marks a watershed
in what AI might be able to achieve.
And it is what we're looking at in this
exhibition here at the Barbican.
And it sparked me off on a kind of journey to see,
well if it can be creative in these confines zone of a game,
where else can it be creative?
Can it be creative in music, in the written word,
in mathematics or in the visual arts?
And so this afternoon,
I'm just gonna take a little look through the impact
that AI creativity has had on the visual world,
because that's actually one of the places
that AI has had great successes.
I made a program for the BBC horizon a few years ago
about AI and it was all very disappointing
what it was able to do.
And computer vision was one of the great hurdles
that they hadn't passed.
But machine learning allows an AI
to learn what's in an image
and then be able to perhaps produce
its own interesting images.
So here's your first exercise, a little AI art touring test.
One of these images is made by an AI
and one of these is made by a human.
This is an AI project that was done in Holland,
where if you're going to be creative,
perhaps looking at the creativity
of the past is the place to start.
What do we value?
What do we think is great art?
So they took all of Rembrandt's work,
learnt the particular things which marked out his style,
his use of light, the particular way he does a portrait
that kind of look in the eye.
He was looking into the soul of the person
that's being painted.
So they managed to achieve an understanding
of Rembrandt's work to produce I think something
which is pretty convincing.
So I wouldn't mind a little light up on the audience
'cause I'm gonna go ask you a question.
Which one of these do you think, can you identify the one
that was done by a human, by a soul?
Have a soul inside
and the one which somehow doesn't feel
like it's communicating.
So let's take the image on the left to start with,
put up your hand if you think that that's the one done
by artificial intelligence.
Now let's have a little survey.
Okay, and hands up if you think the one on the right
is the one done by artificial intelligence.
So actually a little bit, I would say a majority going
for the one on the right.
And who voted for the one on the right here?
Anybody.
What was sort of giving it away for you?
[Speaker] The detail on here, on the neck.
Something about that.
Something about the detail on the neck.
I mean I think detail is something interesting.
So you felt that meant
it was probably artificial intelligence.
Okay, so you are wrong actually.
So this is the one done by artificial intelligence.
And I think, you know,
the fact that you were pretty spirited,
I mean that was quite a Brexity vote actually, I think,
sort of quite committed either way.
But you know, I think it's a mark of how successful
it has been to be able to produce something
that split the audience so convincingly.
But what's the point of looking at the art of the past
making new things?
I mean, AI shouldn't be used for pastiche,
it should be looking for new things.
And certainly Jonathan Jones who hates anything to do
with AI and art really criticized this project.
He wrote what a horrible, tasteless, insensitive,
and soulless travesty.
Of all that is creative in human nature
when technology is used
for things it should never be used for.
But frankly, an art critic who wears a shirt like
that I don't trust terribly much on,
and he's very critical.
The review he wrote of the exhibition here
I think was very ill informed.
But you know,
but I think, you know, that's the point.
People don't understand that the AI is meant
to be pushing us into new realms.
It can help us to understand old art in a new way.
But I think the most exciting thing is how can we use this
as a tool, as we've seen with music just before,
to push us into doing new things.
And here are some paintings that have been created by AI,
which I think are starting to break the mold.
So again, I'm gonna ask you a test.
Four of these are done by a human,
four of these are done by AI.
Can you sniff out the AI?
So again, let's take a vote,
put your hand up if you think the paintings on the left are
done by the artificial intelligence.
Hands up who think those are done by AI.
Okay, very small number for that.
Who thinks the ones on the right then,
big votes for the ones on the right.
So sir, what was for you
the indicator that these may not be human?
[Speaker] Looks like a deep learning.
You can sniff out the deep learning.
I mean that's interesting.
I think there's a very sophisticated, you know,
so I can smell deep learning behind.
You are indeed right.
These are the ones created
by artificial intelligence.
What's interesting is this is created
by something called a creative adversarial network,
a sort of generalization into of something called
a general adversarial network.
A few of the projects
that have been shown in the exhibition here
at the Barbican can use this idea.
It uses the fact that you can use two algorithms
sort of working against each other.
So one algorithm learns about the art of the past,
it learns particular styles, pointless art, cubist art.
It almost becomes an art historian
by analyzing the art of the past.
Then it's asked to make something which breaks that mold,
can't be classified.
So it tries to move into the new, but it can't go too far.
It knows what we regard as art.
And so it's got these two ends of the parameter spectrum
that has gotta find the middle ground,
the discriminate algorithm then judges
whether, no, that I identify
as still something with a particular style.
You haven't broken the mold or else that's too much.
Not something that we regard as art.
And for me, I think this captures actually
how the human white mind works creatively.
Here's Paul Valery, French poet,
who said it takes two to invent anything.
The one makes up combinations and the other one chooses.
And there's a lot of evidence
that creativity is about an explosion of ideas,
but then being critical
and judging which one really is worth putting forward.
So it's interesting, I think, that the algorithm
behind pushing these into the new is using something
that we as creatives actually use in our ideas,
that sort of positive,
the good cop bad cop sort of approach.
But actually I think the most interesting thing about AI art
is to look at seeing whether we can use it
to understand the mind of the AI
or the emerging mind, potentially.
And so I think one
of the interesting projects I saw in writing this book about
creativity and code was actually,
produces rather kitsch sort of art, but it was a deep dream.
This idea that Google has very good
visual recognition software.
They were interested to know,
well, what actually is this vision seeing in these pictures?
It can classify them,
but could we actually see how it is classifying it?
The machine learning is producing code that is quite hard
to kind of look and see how it's making its decisions
and going forward that's gonna be really important.
If it's deciding on a job for you
or a medicine that you might take,
we'd like it to articulate why it's doing things.
We want to see how it sees the world.
So here's an image of a string quartet I play in.
The Google recognition software
identified musical instruments, people, chess.
But then deep dream says, okay,
show me what you are really seeing.
And so it says, just dial up
and accentuate any image that you can see inside there.
So when I put this through Google deep dream,
this is what emerged.
Lots of strange animals, faces, a car,
which began to emerge outta three lower strings.
You begin to understand
how the AI has learnt.
It's learned on lots of images, of faces, of animals.
It starts to give us a sense of of how it sees the world.
And in this way we can pick out important ideas about bias.
For example, these are images, just random pixels
that it started to see dumbbells beginning
to emerge in this, but all the dumbbells had
arms attached to them.
And we began to realize, well, of course,
it had only ever learned about dumbbells
which were being picked up by humans.
It thought it was part of our anatomy.
And so by using the art we can start
to sniff out perhaps potential biases that are beginning
to creep into this code.
And as Marshall McLuhan once said, you know,
Art is our distant early warning system
that can be always relied on to tell the old culture
what is beginning to happen to it.
Which is why I think it's so important
that this event is taking place here at the Barbican.
And we are looking at how the world of music ,visual art,
plays can help us to understand how this new AI is emerging.
Because I think it's gonna be
a real challenge going forward.
This AI is becoming more and more sophisticated.
It's quite possible that it might one day become conscious,
you know, suddenly my phone might suddenly say,
iPhone think, therefore iPhone am,
and I'm gonna have to, you know question,
is there a consciousness inside there?
I think art is perhaps our best FMRI scanner
for understanding what our internal worlds are like.
After all, why do we produce art?
Why do we paint?
Why do we write music?
Because we want to express our internal world
and share it with another human being.
The hard problem of consciousness is all about the fact
that how can I know what your pain is like
and whether it's anything like my pain.
And so I think going forward as AI begins to get
or already an internal world that we don't understand
that these tools of art
and creativity could well be our best way of kind
of sharing our worlds between us.
Because as Stein said, if a lion could speak,
well, we're not gonna be able to understand him.
And as AI emerges, it's gonna be a very different
sort of internal conscious world
if it ever does appear from ours.
And its creativity, probably,
and its art will be the best way to understand
what it might be like to be a piece of AI.
Thank you.
[people clapping]
