Professor Alan Winfield - Bristol Robotics Laboratory HARDtalk

of developing some cancers.

Now it's time for HardTalk.

Welcome to HARDtalk

with me, Stephen Sackur.

For now, the ABC employs human

beings like me to question the way

the world works. For how much

longer? -- the BBC employs human

beings.

As research into artificial

intelligence intensifies,

is there any sphere of human

activity that won't be

revolutionised by AI and robotics?

My guest today is Alan Winfield,

a world-renowned professor

of robot ethics.

From driving, to education,

to work and warfare,

are we unleashing machines

which could turn the dark visions

of science fiction

into science fact?

Alan Winfield, welcome to HARDtalk.

Delighted to be here, Stephen.

You do have a fascinating title,

director of robot ethics,

I'm tempted to ask you,

what's more important to you,

the engineering, robotics

or the ethics, being an ethicist?

Both are equally important.

I'm fundamentally an engineer

so I bring an engineering

perspective to robotics but more

than half of my work now

is thinking about...

I'm kind of a professional

warrior now.

-- worrier now.

Would you say the balance has

shifted over the course

of your career?

You started out very much

in computers and engineering

but increasingly as you have dug

deeply into the subject,

in a sense the more philosophical

side of it has been writ

large for you?

Absolutely right.

It was really getting involved

in public engagement,

robotics public engagement, 15 years

ago that if you like alerted me

and sensitised me to the ethical

questions around robotics and AI.

Let's take this phrase,

artificial intelligence.

It raises an immediate

question in my mind,

how we define intelligence.

I wonder if you could

do that for me?

It's really difficult,

one of the fundamental philosophical

problems with AI is we don't

have a satisfactory definition

for natural intelligence.

Here's a simple definition,

it's doing the right thing

at the right time, but that's not

very helpful from a scientific

point of view.

One thing we can say

about intelligence is it's not one

thing we all have more or less of.

What about thinking?

Are we really, as in the course

of this conversation,

talking about the degree

to which human beings can make

machines that think?

I think thinking

is a dangerous word.

It's an anthropomorphisation

and in gact more than that it's

a humanisation of

the term intelligence.

A lot of the intelligence

you and I have is nothing to do

with conscious reflective thought.

One of the curious things about AI

is that what we thought would be

very difficult 60 years ago,

like playing board games,

chess, Go as it happens,

has turned out to be not easy

but relatively easy,

whereas what we thought would be

very easy 60 years ago,

like making a cup of tea in somebody

else's kitchen, has turned out to be

enormously difficult.

It's interesting you light

upon board games so quickly

because in the news

in the last few days we've seen

something quite interesting,

Google's DeepMind department has

this machine, computer,

call it what you will,

the AlphaGo Zero I think they call

it, which has achieved I think

astounding results playing this

game, I'm not familiar with it,

a game called Go, mainly played

in China, extremely complex,

more moves in it, more complexity

than chess, and this machine is now

capable of beating it seems any

human Grandmaster and the real thing

about it is it's a machine that

appears to learn unsupervised.

That's right.

I must admit, I'm somewhat baffled

by this, you said don't think

about thinking but it seems this

is a machine that thinks.

It's a machine that does

an artificial analogue of thinking,

it doesn't do it in

the way you and I do.

The technology is based

on what are called artificial neural

networks and they are if you like

an abstract model of biological

networks, neural networks,

brains in other words,

which actually we don't understand

very well curiously but we can

still make very simple

abstract models, and that's

what the technology is.

The way to think about the way it

learns, and it is a remarkable

breakthrough, I don't

want to over-hype it because it only

plays Go, it can't make

a cup of tea for you,

but the very interesting thing

is the early generations effectively

had to be trained on data

that was gleaned from human experts

and many, many games of Go.

It had to be loaded

with external information?

Essentially, that's right.

That's what we call supervised

learning, whereas the new version,

and again, if I understand it

correctly, I only scanned the nature

paper this morning, is doing

unsupervised learning.

We actually technically call it

reinforcement learning.

The idea is that the machine

is given nothing else

than if you like the rules

of the game and its world

is the board, the Go

board and the pieces,

and then it just essentially plays

against itself millions

and millions of times.

It's a bit like, you know,

a human infant learning how to,

I don't know, play with building

blocks, Lego, entirely

on his or her own by just learning

over and over again.

Of course, humans don't

actually learn like that,

mostly we learn with supervision,

with parents, teachers,

brothers and sisters,

family and so on.

You're prepared to use

a word like learning,

thinking you don't like,

learning you're prepared to apply

to a machine?

Yes.

What I want to get to,

before we go into the specifics

of driverless cars and autonomous

fighting machines and all of that,

I still want to stay

with big-picture stuff.

The human brain, you've already

mentioned the human brain,

it's the most complex mechanism

we know on this planet.

In the universe in fact.

Is it possible, talking

about the way that Google Mind

and others are developing

artificial intelligence,

that we can all look to create

machines that are as complex

with the billions and trillions

of moving parts, if I can put it

that way, that the human

brain possesses?

I would say in principle, yes,

but not for a very long time.

I think the problem of making an AI

or robot if you like,

a robot is just AI in a physical

body, that is comparable

in intelligence to a human being,

an average human being if you like,

averagely intelligent human being,

is extraordinary difficult and part

of the problem, part of the reason

it's so difficult is we don't

actually have the design,

if you like, the architecture

of human minds.

But in principle you

think we can get it?

What I'm driving at really is this

principle philosophical question

of what the brain is.

To you, Professor, is the brain

in the end chemistry?

Is it material?

Is it a lump of matter?

Yes.

Does it have any spiritual

or any other tangible thing?

It is chemistry?

I'm a materialist, yes,

the brain is thinking meat.

That is a bit of a copout.

You said thinking meat,

it is meat and the way that meat

is arranged means it could think,

so you could create something

artificial where, if it was

as complex and well arranged

as human capacity could make it one

day, it could also think?

I believe in principle, yes.

But the key thing is architecture.

In a sense, the way to think

about the current work on artificial

intelligence, we have these

artificial neural networks

which are almost like

the building blocks.

It's a bit like having marble,

but just having a lot of wonderful

Italian marble doesn't mean

you can make a cathedral,

you need to have the design,

you need to have the architecture

and the know-how to build that

cathedral and we don't have

anything like that.

One more general point and then

I want to get down to the specifics.

Nick Bostrom at Oxford University,

you know him, I know you do

because he works in the same field

as you, you have to think of AI

as a fundamental game

changer for humanity.

It could be the last invention that

human intelligence ever needs

to make, he says, because it's

the beginning of a completely

new era, the machine intelligence

era and in a sense we are a bit

like children playing with something

we have picked up and it happens

to be an unexploded bomb

and we don't even know

the consequences that

could come with it.

Do you share that vision?

I partially share it.

Where I disagree with Nick is that

I don't think we are under threat

from a kind of runaway super

intelligence, which is the thesis

of his book of that subject,

Superintelligence, but I do think

we need to be ever so careful.

In a way, I alluded to this earlier,

we don't understand what natural

intelligence is, we don't have any

general scientific theory

of intelligence so trying to build

artificial general intelligence

is a bit like trying to do particle

physics at CERN without any theory,

without any underlying

scientific theory.

It seems to me that we need both

some serious theory,

which we don't have yet,

we have some but it isn't unified,

there isn't a single

theory if you like like

the standard model physics.

We also need to do responsible

research and innovation.

In other words we need to innovate

ethically to make sure any as it

were unintended consequences

are foreseen and we had them off.

Let's talk in a more practical

sense, unintended consequences may

well come up.

Let's start with something I think

most of us are aware of now,

and regard as one of the most both

challenging and perhaps exciting

specific AI achievements,

that is the driverless car.

Yes.

It seems to me all sorts of issues

are raised by a world

in which cars are driverless.

A lot of moral and ethical issues

as well as practical ones.

You work with people in this field,

are you excited by driverless cars?

I am, yes.

I think driverless cars have

tremendous potential

for two things...

Do you see them as robots?

I do, yes, a driverless

car is a robot.

Typically once a robot becomes part

of normal life we stop calling it

a robot, like a vacuum cleaner.

I think there are two tremendous

advances from driverless cars we can

look forward to, one is reducing

the number of people killed in road

traffic accidents significantly,

if we can achieve that,

so I'm going to be cautious

when I speak more on this.

The other is giving mobility

to people, elderly people,

disabled people who

currently don't have that.

Both of those are very practical

but Science Magazine last year

studied a group of almost

2,000 people, ask them

about what they wanted to see

in terms of the morale at the almost

of using driverless cars,

how the programming of the car

would be developed to ensure that,

for example, in a hypothetical,

if a car was on the road

and it was about to crash but if it

veered off the road to avoid a crash

it would hit a group

of schoolchildren being led

by a teacher down the road.

The public in this survey wanted

to know that the car

would in the end accept its own

destruction and that of its driver,

human passenger rather,

as opposed to saving itself

and ploughing into the children

on the side of the road.

How do you as a robot ethicist cope

with this sort of challenge?

The first thing I'd say is let's not

get it out of proportion.

You have to ask yourself as a human

driver, probably like me you've got

many years of experience

of driving, have you ever

encountered that situation?

Not in my case, but I want to know

if I ever step into a driverless car

someone has thought about this.

I think you're right.

The ethicists and the

lawyers are not clear.

The point is we need

to have a conversation.

I think it's really important that

if we have driverless cars that make

those kinds of ethical decisions,

you know, that essentially decide

whether to potentially

harm the occupants...

You're doing what you told me

off for doing, you're

anthropomorphising, it wouldn't be

making an ethical decision,

it would be reflecting

the values of the programmer.

Those rules need to be decided

by the whole of society.

The fact is, whatever those rules

are, there will be occasions

when the rules result

in consequences that we don't

like and therefore I think

the whole of society needs

to if you like own the

responsibility for those cases.

So you are making a call,

be it driverless cars or any other

examples we're thinking

about with AI, of the technological

developments in lockstep

with a new approach to monitoring,

regulation, universal

standardisation.

And a conversation, a big

conversation in society

so that we own the ethics

that we decide should be embedded.

But that will not help.

Much of the development here,

I mean, you work at Bristol

in a robotics lab but a lot

of cutting-edge work is being done

in the private sector.

Some of it is done by secretive

defence establishments.

There is no standardisation,

there is no cooperation.

It is a deeply competitive world.

It jolly well needs to be.

My view is simple.

The autopilot of a driverless car

should be subject to the same levels

of compliance with safety standards

as the autopilot of an aircraft.

We all accept...

You and I would not get

into an aircraft if we thought

that the autopilot had not met those

very high standards.

It is inconceivable that we could

allow driverless cars on our roads

that have not passed those kind

of safety certification processes.

Leaving driverless cars

and going to areas that are,

perhaps, more problematic

for human beings.

They developed the idea

of the machine, the future

intelligent machine,

taking jobs and roles that have

traditionally always been done

by human beings because they involve

things like empathy and care.

And compassion.

I'm thinking about roles

of social care and education.

Even, frankly, a sexual partner

because we all now read

about the sexbots that

are being developed.

In these roles, do you feel

comfortable with the notion that

machines will take over

from human beings?

No.

And I do not think they will.

But they already are...

Japan has carers.

A care robot may well be

able to care for you,

for example, for your physical

needs, but it cannot care about you.

Only humans can care about either

humans or any other animal.

Object, robots cannot care

about people or things,

for that matter.

And the same is true for teachers.

Teachers typically care

about their classes.

You think some people are getting

way overheated about this?

One of the most well-known teachers

here in Britain he now says that

in his vision of a future education

system, many children will be taught

one-on-one in a spectacular

new way by machines.

He says it is like giving

every child access to

the best private school.

Ultimately, there may well be...

And we are talking

about into the future,

some combination of machine

teaching and human teaching.

You cannot take the human out.

An important thing to remember

here is particularly human

characteristics of

empathy, sympathy...

Theory of mind, the ability

to anticipate, to read each other.

These are uniquely human

characteristics as is our creativity

and innovation and intuition.

These are things we have no idea how

to build artificially.

Jobs that involve

those things are safe.

Interesting.

People nowadays are looking

at doomsday scenarios

with the development of robotics

and AI where frankly,

most jobs one can think of -

and I was being flippant earlier

about being replaced by a robot -

but you suggest to me that so many

different jobs, not just blue-collar

but white-collar as well will be

done by machines...

Again, are we overstating it?

I think we are.

Yes.

I am not saying that will not happen

eventually but I think that

what we have is much more time

than people suppose to find

a harmonious, if you like,

accommodation between

human and machine.

That actually allows us to exploit

the qualities of humans

and the skills, the things

that humans want to do

which is, you know...

If you don't mind me saying,

you seem both extraordinarily

sanguine and comfortable

and optimistic about the way

in which AI is developing under

the control of human beings,

and your faith in humanity's ability

to co-operate on this and establish

standards seems to run

in the face of facts.

In one area is weaponisation.

The notion that AI and robotics

will revolutionise warfare

and war fighting.

You were one of 1000 senior

scientists who signed an appeal

for a ban on AI weaponry in 2015.

That that will not happen, will it?

The ban...

It may...

Did you see what

Vladimir Putin said?

He said that artificial intelligence

is the future for Russia

and all of humankind -

and this is the key bit,

"Whoever becomes the leader in this

sphere will become the ruler

of the world."

I am an optimist but I am also very

worried about exactly this thing.

We have already seen, if you like,

the political weaponisation of AI.

It is clear, isn't it,

that the evidence is mounting that

AI was used in recent elections?

You are talking about the hacking?

And some of that

believe is from Russia?

That is political weaponisation

and we do need to be worried

about these things.

We do need to have ethical

standards and we need

to have worldwide agreement.

I am optimistic about a ban

on lethal autonomous weapons

systems.

The campaign is gaining traction,

there have been all kinds

of discussions and I know some

of the people involved quite well

in the United Nations.

But we know the limitations

of the United Nations

and the limitations of politics

and we know that human nature

usually leads to the striving

to compete and to win,

whether it be in politics

or in the battlefield.

Can I leave you with this thought?

It seems that there is a debate

within science and you are on one

side being sanguinary

and optimistic.

Perhaps on the other side,

Stephen Hawking recently said

that the development of full

artificial intelligence could spell

the end of the human race.

Do you find that kind of thought

helpful or deeply unhelpful?

Deeply unhelpful.

The problem is that

it is not inevitable.

What he is talking about here

is a very small probability.

If you like it is a very long

series of if this happens,

then if that happens,

then this and so on and on.

I wrote about this in

the Observer back in 2014.

Before Stephen Hawking got

involved in this debate.

My view is that we are worrying

about an extraordinarily unlikely

event, that is the

intelligence explosion,...

Do you think we have actually been

too conditioned by science fiction

and by the Terminator concept?

We have.

There is no doubt.

The problem is that

we are fascinated.

It is a combination of fear

and fascination and that is why

we love science-fiction.

But in your view it is fiction?

That scenario is fiction that

you are quite right,

there are things we

should worry about now.

We need to worry about jobs,

we need to worry about weaponisation

of AI and we need to worry

about standards in driverless cars

and care robots, in

medical diagnosis AIs.

There are many things that

are here and now problems

in the sense that are kind of more

to do with the fact that AI is not

very intelligent so we need to worry

about artificial stupidity.

That is a neat way of ending.

Alan Winfield, thank you very much.