0:02

this is philosophy bites with me,

0:04

not to wolverton & me, david

0:06

edmonds if you enjoyed, please,

0:09

purchase with conium

0:11

funded through donations, would be gratefully

0:13

received for for details go to www dot philosophy bites dot-com humans

0:19

are creating artificial intelligence machines

0:22

and systems that use a i already

0:24

making decisions that our lives in

0:26

ways so,

0:27

how do we build ethics

0:29

into machines and

0:31

mike we learn something about ethics from

0:33

a ethics peter wales

0:35

and is a philosopher at the university of michigan

0:38

rosen welcome to foresee bites

0:41

thank you very much very glad to be the

0:43

topic we're going to focus on is a

0:45

i and ethics i'm

0:47

not actually sure about ethics bullets

0:50

get clear about ai first what are we talking about

0:52

a eyes a very general term

0:54

well for thing for me to say is i'm not an expert

0:57

on a i i'm a philosopher

0:59

who's worked in moral philosophy and through

1:01

believes that there are many ethical challenges

1:03

raised by a i and that they're

1:06

coming onto us much quicker than we expected

1:08

they would and so it's important

1:10

it's think for philosophers who art expert naias

1:13

to try to do what we can to scramble

1:15

the

1:16

the i literally artificial intelligence

1:18

what kind of intelligence is relevant

1:21

most people i think would say that the most

1:23

general idea of intelligence is this capacity

1:26

to solve problems that , mean

1:28

for example capacity to take what

1:30

has been learned in the past and solved new

1:32

problems problems now i think a lot

1:34

of people would include learning and intuitions

1:37

a data banks just sitting there is not by

1:39

itself into isn't because is not thinking

1:41

of thinking and it's not seeing

1:43

how they work out when learning from whether

1:45

they did workouts so that systems

1:48

that were mostly thinking about now are ai

1:50

systems that do learn in fact learning

1:52

is their fundamental principle their operation

1:55

and when i first started to see such

1:57

systems a couple of years ago i

1:59

was

1:59

the only excited because i think

2:02

that these actually have a chance of

2:04

being something like that human intuitive

2:07

knowledge system obviously there are

2:09

modeled on neurons in

2:11

neural networks in the mind but they aren't strictly

2:13

like those and no one would say

2:16

that were anything like level of reaching something

2:18

like mind but it does

2:20

seem to me that compared to the old

2:22

method the i which

2:24

was a method that was based on sort of logical

2:26

processing and essentially

2:29

required human experts to

2:31

program that knowledge into the machine

2:34

and then you'd rely on the machines power

2:36

and memory and so on to crank through

2:38

possibilities but the machine

2:40

itself was problem solving only and that

2:42

instrumental sense what's exciting

2:45

now is that this machines are

2:47

learning things things that

2:49

we didn't know

2:51

and are able to advance

2:53

the way that we think about a range of issues

2:56

could you give an example of the kind of thing yourself most

2:58

i guess most famous example is the distinction

3:01

between something like deep blue

3:03

the chess playing program that allegedly

3:05

be test for an alpha

3:07

go or alpha zero

3:09

now or muse euro for says next

3:12

phase deep blue was

3:14

programmed by experts just

3:16

experts just experts thus systems

3:18

and so on over many years

3:21

it was, in fact a machine that didn't

3:23

itself beat kasparov would

3:25

be casper, obviously human information and

3:28

knowledge that was put into the machine and

3:30

it turn the crank gave moves

3:32

it wouldn't really be fair to say that

3:35

mushy intelligence defeated

3:37

a human a system like alpha go

3:40

starts with no knowledge of go strategy

3:42

only the rules of the game

3:44

and users simulation playing

3:46

against itself

3:48

in developing it's own method of evaluating

3:51

simulations the only learning signal

3:53

it gets is whether it wins again or not

3:55

and through that developed not only

3:58

discovered some known strike it is

4:00

ago but invented new strategies and

4:03

was able to defeat of world champion mortuary

4:06

fast but he was able to do this not

4:08

with years and years of training and expertise

4:10

that through this process or reinforcement

4:13

learning this kind of learning machine

4:16

of as he has huge potential but

4:18

it also has also quality which seems

4:21

to be getting close to the human mind human

4:23

mind respects and obviously

4:25

we hold individual human

4:27

beings responsible for all kinds of things doing

4:29

this just a world in which will start

4:31

to hold machines responsible

4:33

for thing that's a good question because

4:36

responsibilities else very thorny issue

4:38

there are various kinds of responsibility but

4:40

one question is whether we will

4:43

allow them allow certain amount of autonomy

4:45

in decision making

4:46

that's in effect recognizing that weekend

4:48

interest certain decisions with them and

4:51

that's allowing them allowing certain responsibility

4:54

does that mean we think of them as moral agents

4:56

know because for moral agency

4:58

requires a lot more than just the ability

5:01

to make decisions we think of it as involving

5:03

for example self reflection consciousness

5:06

capacity for morley motion and

5:09

nothing like us to these machines but

5:12

they will be given responsibility

5:14

in the question is

5:15

are we entitled or can we be

5:17

justified in giving them these responsibilities

5:20

what would they have to be like for that me

5:22

result and ,

5:24

speaking the future but they're already been given

5:26

lots of responsibilities responsibilities

5:29

this means that be perfect sarah

5:31

seriously very quickly about

5:33

the question of okay well in what

5:35

sense might these machines be themselves

5:38

capable of responding to a morally relevant

5:40

features of situations i can

5:42

imagine prison for a machine

5:44

that runs an automated heart

5:47

operation that might actually be

5:50

more responsive to subtle changes in the

5:52

body as as impressions going on better

5:54

processing data that's coming three

5:56

more accurate and is cussing than a certain could

5:59

be the you have this machine

6:01

i doing a heart operation

6:03

on a very vulnerable person we've

6:06

, our trust in that if the machine

6:08

does the job better them certain it seems to

6:10

me what's the moral issue that

6:13

good these machines because they're learning

6:15

machines they don't have

6:17

first principles that were given to them their

6:19

first principle so to speak come from the data

6:21

income from experience and so

6:24

a machine that does doing that won't

6:26

necessarily have the knowledge a

6:28

surgeon would have of certain signs

6:31

as the patient is not doing well he

6:33

could make a mistake not because

6:36

it's a poor however or because

6:38

it's for monitoring these vital signals that

6:40

because there are other signals and patience

6:43

or stations body might give that something's

6:45

going wrong we don't know in advance

6:47

which those kinds of false are going to be we've

6:50

already found that there are cases

6:52

for example with self driving vehicles

6:55

where they will see a scene which

6:57

we would clearly identify as involving

6:59

a human being let's say and not see

7:01

the human because of the reflectance of the light

7:03

or something like that and so

7:05

they will make mistakes as

7:08

any learning agent will

7:10

the mistakes want me the same ones perhaps

7:12

that we would make and we won't be able to anticipate

7:14

all of them we can try begin read many

7:16

simulations but the question

7:18

is it in what sense are we

7:21

justified if the kinds of mistakes

7:23

they could make her not just a technical mistakes

7:25

but a ethical mistake the

7:28

subtle difference between a technical mistake

7:30

and unethical mistake consider

7:33

a program that is

7:35

learning to make good decisions

7:37

about the amount of parole

7:40

someone should have

7:41

that machine is going to take in all manner of variables

7:44

demographic variables included some

7:46

, those will concerned things like race and gender

7:48

and so on and we might

7:51

think that it's inadmissible to

7:53

take someone's race or gender into account

7:55

when making account when like this but

7:57

if it's statistically correlated and

8:00

maybe for of complex sociological reason

8:02

statistically correlated with recidivism

8:04

repeat offense then the

8:06

machine will use that information that's

8:09

not a technical mistake

8:11

because it is predicting accurately

8:13

using that information but we might

8:15

think it's immoral mistake it's an injustice so

8:18

the question is given that the machine is harvesting

8:21

all this information and that there will

8:23

be the statistical relevance relations are

8:25

there some which we think it would be impermissible

8:28

to use we have some idea that

8:30

in no law and ethics but

8:32

that isn't something that is automatically discovered

8:35

just having a challenge of making sure that

8:37

you make recommendations that are statistically

8:39

as likely as possible to

8:41

predict who's going to commit a new crime

8:44

or not or how soon they can be trusted so

8:46

does not mean that you have to then find

8:49

a way of building in some kind of ethical

8:51

imitation some kind of cut

8:54

out switch is it worth that doesn't allow

8:56

the machine to act in those sorts of cases

8:59

yes one model of how we might

9:01

respond is well maybe we could build

9:03

in some ethics and

9:06

certainly you could build into machines

9:09

the rhodesians when using certain date now

9:12

because these are highly and capable at

9:14

a probabilistic learning it's not

9:16

all that evident always have the limit

9:18

what information they have because they can follow

9:21

information and project information and

9:23

simulate information and draw inferences

9:26

from that and then see whether those edwards's were

9:28

accurate and then imagine that information

9:30

for example the weather's this hidden variables that

9:32

i wasn't given but that predicts to recidivism

9:35

and it turns out that hidden variable is something

9:37

very much like race so

9:39

did it take race into account not

9:41

clear because what's happening is that the society's

9:44

set up and a discriminatory way and

9:46

, discriminations going to get translated

9:49

into the data and then modeling

9:52

that data as accurate me and pack the

9:54

as possible is going to discovered that there's

9:56

discovered that a latency here we might think

9:58

it's a latency that he can't use of as a know

9:59

are you can't look at that

10:01

but of course if that layton variable is

10:03

highly predictions and machine wealth and

10:06

said have to use it then

10:08

you have a hierarchy of machines we have the moral

10:10

messina the tells the then shuts down

10:12

the ones which are not behaving well

10:14

there's a big it was an interesting literature

10:16

on this question of the off switch for ai

10:19

systems one thing that's often said

10:21

is that as ai systems become agents

10:24

not just capable of modeling data and

10:26

suggesting hypotheses and so on but

10:29

capable of taking actions they will

10:31

have the features of agents bostick

10:33

estimates about the state of the world they

10:36

will have a value function of some

10:38

kind it might be goals it might be

10:40

if it's an animal it's a reward function

10:42

and they will also have a capacity

10:44

to imagine alternative courses

10:46

of action assign expectations

10:48

to them and make decisions correspondingly

10:51

so suppose we give a machine a particular goal

10:54

well one condition of achieving most

10:56

goals is that the machine continue to exist

10:59

and so therefore the machine gets a sub goal

11:02

of protecting his own existence now

11:04

how my to do that if we're tried to shut it off

11:07

well it might anticipate when

11:09

we would shut it off or look for signs of that

11:12

and replicate itself that has access

11:14

to the web let's say and it puts replicas

11:17

of itself out there in there web somewhere and somewhere we flip

11:19

the switch thinking return the machine off

11:21

but we've actually duplicated the machine that's

11:23

man nefarious on the machines part

11:26

it's thinking i can't achieve this goal with i

11:28

don't exist there will be elements

11:30

of what the machine is trying to do that

11:32

we aren't immediately aware of be don't immediately

11:35

understand and anticipating

11:37

those and you know be said well let's build a machine

11:40

that but despite that they're me to

11:42

ah but what about turning that machine off will

11:44

have to build a machine the when anticipates it's

11:46

actions and then him what what's his the top

11:49

what's the ethical master and the

11:51

thing is that ethics has not produced for

11:53

us and axiomatic says

11:55

we have capital series

11:57

which give us general principles but we

11:59

don't have

11:59

the axiomatic system from which you can deduce whether

12:02

given behaviors right or wrong and that's

12:04

true even for what look like the most axiomatic

12:06

system for utilitarianism for example

12:09

many utilitarian now our rule utilitarian

12:12

so the question is what would be that set of rules

12:14

which would have the best acceptance utility bill

12:16

what the heck is answer that question and

12:19

with we don't know the answer to that question that we don't know

12:21

was right or wrong maybe with

12:23

it has a capacity to model

12:25

once he would be like of society follow

12:27

certain rules that we don't have now for

12:30

example machines are capable of modeling financial

12:32

markets and figuring out what's going on and those

12:35

if we had we had that were capable

12:37

of learning not just a

12:39

particular area but more broadly about

12:42

tendencies in human behavior is about social

12:44

interactions and so on we ,

12:47

be able to gain some knowledge about well what might

12:49

this set of rules be like and what might be like if

12:51

they were internalized and foliage a certain

12:53

degree and so even for

12:55

the most dedicated rule utilitarian

12:58

there's going to be a lot about ethics they don't know the

13:00

machine might be able to know that

13:03

to some degree better than an ordinary human

13:05

agent we rely on intuition a great deal

13:07

and that's fine because it's

13:10

more complicated systems more don't think there are axioms

13:12

of human behavior of human there'd be

13:14

axioms of ethics though

13:17

we can't think then that

13:19

we have a an algorithm to write down

13:22

which will say know this actions permissible seconds

13:24

impermissible having said that you

13:27

to terrorism is more plausible

13:29

in that respect many other ethical approaches

13:32

because it uses quantify pull data

13:34

it because it tries to be based on facts

13:37

and consequences and actions

13:40

whereas other types of theory

13:42

maybe theory maybe by theory or

13:44

right space theories it it gets much more complicated

13:47

to think how you could possibly program and

13:49

machine to implement and

13:51

go beyond human understanding

13:54

of essex it wouldn't want to see is

13:56

understanding that x these

13:58

are machines that the attack patterns

14:00

use patterns to form complex models

14:03

but does that mean they have an understanding of it well

14:05

noah alpha though doesn't understand

14:07

that is playing against a human psychology whereas

14:09

any go player would understand turns

14:12

out you can write a very successful go playing

14:14

program that doesn't depend upon knowledge of

14:16

human psychology ethics i don't know

14:18

this has kind of be possible because

14:21

, not a game with a kind of rules and bounded

14:23

as on the other hand there's some ethical series had

14:25

tried it acts as

14:27

acts set of rules and so yes you

14:29

could take an ethics it was a set of rules

14:31

and you could have the machine have

14:33

those rules in the same way that go machines have

14:35

the go rules the tunnels as you wouldn't be

14:38

very happy with it's behavior because

14:40

we don't have a set of rules you could write down we could

14:42

say in this circumstance you should do such and such

14:45

even rules like don't murdering innocent

14:47

person or something like that wealth there are

14:49

lots of cases where that may be the only

14:51

way to prevent others very

14:53

serious harms if the rule just as

14:55

don't do as and be get all these famous problems

14:58

do lie to the dictator who

15:00

comes to your door looking for your our friend

15:02

who's hiding the basement spell if you were told never

15:04

to lie you hidden no

15:07

we could give the machines rules we

15:09

probably wouldn't like the result of and

15:12

one thing that i've been trying to argue is

15:14

that even though we don't think of these agents

15:16

and shouldn't as a stand think of them as having

15:19

ethical understanding they certainly

15:21

don't have consciousness of or as we can tell

15:23

they don't have the kind of moral emotions

15:26

that we house they don't have the kind

15:28

of appreciation of what it is to be

15:30

a person nor respect a person or something that's

15:32

so given their capacities

15:35

the question really is can we make them sensitive

15:37

to ethically relevant features of situations

15:40

those ethically relevant features a situations

15:43

they arise in situations not

15:45

just because we're human and even

15:47

not just because we're conscious

15:49

so if you're a person to a game theory

15:52

you have agents agents make moves

15:54

they have strategies they have interests

15:57

they have estimates and so those

15:59

a that's will

16:01

in certain circumstances either get themselves

16:03

into a collision or a paradox

16:06

unable to solve the coordination problem for example

16:08

even though their goals would require that

16:10

the coordinates or they won't in

16:12

on that view you can say well what

16:14

are the principles that would allow them to do that

16:16

or whether the motivations are goals

16:19

that would allow them to do that and those might

16:21

be quite universal they might not depend

16:23

upon any particular kind of agent

16:25

but they will be true of asians in general the

16:27

you're thing that if you think of ethics in terms of

16:30

human corporation then

16:32

the pope perhaps that machines

16:34

could come up with solutions

16:36

to problems that we haven't solved well

16:39

that seems like one possibility in the immediate

16:41

term i think part of it is can

16:44

we develop communities of

16:46

cooperation with artificial agents

16:49

because they are already out in the world

16:51

in fact for doing something very similar with autonomous

16:54

vehicles all the time there are coordination

16:56

problems there are problems of communication

16:59

of signaling to one another there are problems

17:01

were if we can cooperate

17:04

our able to achieve a result like smooth

17:06

going traffic that are the not and

17:09

those agents out there are learning

17:11

about that from our behavior from

17:13

their behavior and so they're acquiring

17:16

some of this kind of knowledge or

17:18

information that's a sense in which

17:20

we didn't have a program into them the

17:23

way of solving certain sorts of problems because they

17:25

will be trying things perhaps at random events

17:27

unfortunately and so we

17:30

couldn't program ethics into them the question is

17:32

ah good could they have experiences such

17:35

that they learn and people are beginning

17:37

to do some work now with cooperative

17:39

games rather than competitive games like chess

17:42

go and with a multi

17:44

agents settings not just individual

17:46

asian in a good question

17:48

is what kinds of forms

17:51

of behavior can they evolve then

17:53

what would their values function look

17:55

like so that they would have that

17:58

kind of behavior is not just a but

17:59

cuellar solution to this problem

18:01

that at a generalized motivational

18:03

structures that will enable them to go on to new problems

18:07

and solve them in a way that would produce things

18:09

like outcome that were positive some

18:11

have better for them and better for us we've

18:14

been talking a lot about the potential

18:16

of idea skull issues that arise the complexities

18:20

of trying see program

18:22

ai to access actually

18:24

as it were which isn't the contribution

18:27

of philosophers should be or could

18:29

be in this area very good again

18:32

i'm not thinking we program essex into them because

18:34

we don't know how to write doing the

18:36

formula are forgotten how to write down the axioms

18:39

or we don't know how to say well these

18:41

are the following thirty two thousand and ethically

18:43

relevant teachers are situations on

18:45

the other hand if you've machines that

18:47

can evaluate situations and simulate

18:50

out and figure out what what are the things that

18:52

are being affected by this variable in the situation

18:55

they might have that capacity of situational

18:58

responsiveness weaker that moral

19:00

intuition that moral

19:02

intuition i would say is in humans

19:04

actually quite sophisticated and well developed

19:07

and , promise of these machines

19:09

is that they will be much more light

19:11

intuitive knowledge systems than like

19:13

logic machines and so they can

19:15

accomplish the kinds of things that perhaps

19:17

only intuitive intuitive

19:19

can accomplish the way in

19:21

which we actually do

19:23

learn about the world that are it's common

19:26

sense knowledge or even sophisticated physics

19:28

involves a tremendous amount of intuitiveness

19:31

we , have the prospect of machines

19:33

that can have intuitive knowledge in something

19:36

like this sense and that

19:38

therefore could be used know

19:40

just as so tools

19:43

that could function as as agents

19:45

i do think more of loss we can help us

19:47

understand the way that humans learn

19:49

ethics and maybe something about the

19:51

nature of ethical theories what might

19:53

ethical knowledge look like what kinds

19:56

of principles of so decision

19:58

making are learning by be involved not

20:00

necessarily ethical principles it

20:03

might be quite general principles and

20:05

i'm arguing that the same principles

20:07

that enable us to do something like sustain the

20:09

cooperation necessary for a language

20:12

large created episteme a community where

20:14

we share information which humans have managed

20:17

to do in a way that other species of

20:19

know the capacities we have to do

20:21

that are really into row with our capacities for

20:23

morality and so philosophy

20:25

can try to help give us insight into will what

20:27

is that set of capacities

20:31

and then because they can smell this

20:33

smell this of what we would have to see as

20:35

a prospect for these machines the

20:37

runaway looking at these machines might

20:39

actually give us a better understanding as the

20:41

row for instance of moral intuition yes

20:44

i think that's right one of the things i've discovered

20:46

in my own teaching i teach in a way

20:48

that i pose problems to students in

20:51

they're able to respond anonymously using

20:54

little remote and over time

20:56

i've been able to ask them a lot of questions

20:59

other than the usual questions not

21:01

just i should you turn the trolley or something like

21:03

that but how would you feel if

21:05

you during the trolley or how would you feel

21:07

as you learned your roommate had turned the

21:09

trolley or if you were

21:11

to become contrite after

21:14

having pushed someone off a footbridge and

21:16

you went to that person's family to try

21:18

to express your deep

21:21

regret the what you did or thought you had to go

21:23

through you feel how would you

21:25

expect them to feel what would you try

21:27

to come out of that interaction the

21:29

interesting thing is my students don't just as

21:32

are sick questions like said you push them and they have

21:34

answers to those questions and they have

21:36

been very quickly the mother tells

21:38

me about ethical intuition is that it actually embodies

21:40

of very complex set

21:42

of understanding of human relations surprising

21:45

is how robust some of these intuitions are and

21:48

you my she will why are they so robust

21:50

and the answer might be because this because this kind of thing that

21:52

we all learn from our

21:54

human experiences interestingly

21:57

enough there's a have you know about psychopathy

22:00

that is actually a kind of a learning disorder

22:02

though i think that important

22:05

feature ethical intuition it's not some

22:07

kind of a seeing of platonic

22:09

principles neither is it instinctive

22:13

we can have a situation where we

22:15

learn about philosophical

22:17

problems like

22:18

the status of intuitions

22:20

i'm trying to model what's going on

22:22

in intuition and so one thing these

22:25

machines can do is give it a whole body

22:27

of human intuitions try to extract

22:29

from that what kind of implicit theories which

22:32

philosophers have been trying to do but we do it piecemeal

22:35

and very hard to do we

22:37

might find that actually there are these implicit

22:40

series involved in our intuitions and then we can

22:42

look at those critically there

22:44

were no question the measuring

22:47

the world in which they i could

22:49

actually be somewhere where we outsource

22:52

some of ammo thinking what if it gets good at

22:54

modeling moral intuitions why

22:56

don't we need to use intuitions anymore

22:59

isn't there a risk that ah capacity

23:01

for making judgments

23:03

on the fly could atrophy

23:05

as atrophy result of that well yes

23:08

and an example of this would be something

23:10

like the way in which people so sense of direction

23:12

or ability to stay oriented and in physical

23:14

space as and atrophied to some extent thanks

23:17

to the fact that people are using their phones

23:19

new symbol , i expect

23:21

to that some of the social skills of drivers

23:24

may atrophy because they're just counting

23:26

on the car to do this and they wouldn't

23:28

be as good at it if they were put in a position

23:30

of driving one way to think about the idea

23:32

of learning here is that

23:34

learning as learning continuing process

23:36

and to the extent that we do that actually

23:39

there's going to be less effective learning by

23:41

these machines because part

23:44

of what they will learn from is the way that we interact with

23:46

them and we're going to have to interact with them as

23:49

agents that they are needed that we are you

23:51

know you can have a gentle picture of knowledge in

23:53

which okay knowledge is not some unified

23:56

axiomatic structure it benefits

23:58

from the diversity of input that

24:00

it gets added expenses add diversity collapses

24:03

then it's going to be biased heavily by the

24:05

pool no machine can know

24:08

everything so there will always be some

24:10

bias from the amount

24:12

or kind of data is going in in

24:14

humans would be irresponsible if

24:17

save resigned their role because

24:19

just think about the way in which in our lifetimes

24:22

there has been change in the way in which

24:25

gender categories have got

24:27

ethical implications or are connected

24:30

with ethical issues that's a historical thing

24:32

it's not a thing that would be something that them machines

24:35

would be tuned into if

24:37

we turn over there making all these decisions

24:40

then they will reflect twenty

24:42

years ago moral intuition bread ours

24:44

is constantly being updated were constantly

24:46

learning and , is our responsibility

24:49

not to allow that to happen and

24:52

to decide to we do will actually be getting

24:54

less intuitive rosen

24:57

thank you very much for thank you very much

25:00

appreciated wishes

25:06

tomorrow's supposed to be bites go to www

25:09

dot philosophy bites dot com you can

25:11

also find details there a philosophy bites

25:13

books and how to support us