0:01

Ted Audio Collective Hi,

0:08

I'm Adam Grant. I think about work a

0:10

lot. That's why I wanted to tell you

0:12

about Canva Docs, which will help you expertly

0:14

craft your work communications. They

0:16

have an AI text generator built in

0:19

called Magic Write, powered by OpenAI. You

0:22

can generate any text you want.

0:24

Job descriptions, marketing plans, sales proposals.

0:27

Just start with a prompt and you'll have a

0:30

draft in seconds. Tweak your draft

0:32

and you're done. Try Canva

0:34

Docs with an AI text generator

0:36

built in at canva.com. Designed

0:38

for work. This

0:42

episode is brought to you by Progressive. Most

0:45

of you aren't just listening right now. You're

0:47

driving, cleaning, and even exercising. But

0:49

what if you could be saving money by switching to

0:51

Progressive? Drivers who save by

0:53

switching save nearly $750 on average. And

0:57

auto customers qualify for an average of seven

0:59

discounts. Multitask right now.

1:01

Quote today at progressive.com. Progressive

1:03

Casualty Insurance Company and affiliates.

1:06

National average 12 month savings of $744

1:08

by new customers surveyed who saved with

1:10

Progressive between June 2022 and May 2023.

1:15

Potential savings will vary. Discounts not available

1:17

in all states and situations. You're

1:22

growing a business and you can't afford to slow

1:24

down. If anything, you could probably use a few

1:26

more hours in the day. That's

1:28

why the most successful growing businesses

1:31

are working together in Slack. Slack

1:33

is where work happens with all

1:35

your people, data, and information in

1:37

one AI powered place. Start

1:39

a call instantly in huddles and

1:41

ditch cumbersome calendar invites. Or build

1:44

an automation with Workflow Builder to take

1:46

routine tasks off your plate. No coding

1:49

required. Grow your business in Slack. Visit

1:51

slack.com to get started. updated

2:00

on Everything Business on TED Business,

2:02

a podcast hosted by Columbia Business

2:05

School professor Modupe Acanola. Every week

2:07

she'll introduce you to leaders with

2:09

unique insights on work, answering questions

2:11

like, how do four day

2:14

work weeks work? Will a machine ever

2:16

take my job? Get some surprising answers

2:18

on TED Business wherever you listen to

2:20

podcasts. Whatever

2:23

it is that is the solution to

2:25

humanity's problems, I'd argue it's probably not

2:27

in our imagination because if it was,

2:29

we'd be doing it. So what we're

2:31

looking for are things that are outside

2:33

the sphere of human

2:36

imagination. Hey everyone,

2:40

it's Adam Grant. Welcome back to Rethinking,

2:42

my podcast on the science of what

2:44

makes us tick with the TED Audio

2:46

Collective. I'm an organizational psychologist

2:48

and I'm taking you inside the minds

2:50

of fascinating people to explore new thoughts

2:53

and new ways of thinking. My

2:58

guest today is tech pioneer, Aza Raskin.

3:00

As co-founder of the Center for

3:02

Humane Technology, he's a leading advocate

3:05

for the responsible reimagining of the

3:07

digital world to prevent polarization and

3:09

promote well-being. Aza's work

3:11

focuses on solving some of the biggest collective

3:13

problems of our age, especially

3:15

as our tech rapidly evolves. AI

3:18

is like the invention of the

3:20

telescope and when we invented the telescope

3:22

we learned that Earth was not the

3:24

center. I've been thinking a

3:26

lot about the implications of

3:29

what happens when AI teaches us that

3:31

humanity is not the center. If

3:34

you don't know Aza by name, you know some

3:36

of his creations. He designed the

3:38

feature that makes students growing possible, which he

3:40

now regrets. And he coined

3:42

the phrase, freedom of speech is not freedom

3:45

of reach. Since then, he's

3:47

expanded his scope by co-founding the

3:49

Earth Species Project, where he's using

3:51

tech to decipher animal communication. Between

3:54

improving social media and talking to whales, we

3:57

had a lot to discuss. And

3:59

Aza challenged me to rethink my assumption that

4:01

these two missions aren't as different in any way.

4:11

Hey, Aza. Hey, Adam. I'm excited

4:13

for this. I feel like there is

4:15

so much ground we could cover, I

4:17

hardly know where to begin. Yeah, it's

4:20

only care for first all humans and

4:22

then after that, all beings. So

4:25

you grew up in tech and

4:28

I understand we have you to blame

4:30

for infinite scroll. Everyone's

4:32

just gonna start pelting me with tomatoes. I

4:35

did invent infinite scroll and I think it's

4:37

really important to understand my motivations and then

4:39

what went wrong because it was a

4:41

big lesson for me. When I invented infinite scroll, this

4:43

was before social media had really taken off. This was

4:45

way back just when MapQuest, you know, I don't know

4:48

if you remember that. Of course I do. Right? Like

4:50

we have to click and then the map would move

4:52

over and then you have to reload the page. And

4:55

the thought hit me like I'm a designer. Every

4:57

time I asked the user to make a decision

4:59

they don't care about, I failed. When you get

5:02

near the bottom of a page, that means you

5:04

haven't found what you're looking for. Just load some

5:06

more stuff. And I was designing it for blog

5:08

posts. I was thinking about search results. And it's

5:10

just honestly, it is a better interface. And then

5:13

I went around to like Google and Twitter and

5:15

say, Oh, we should adopt this interface. And I

5:17

was blind to the way

5:19

that my invention created with positive

5:21

intent was going to be picked

5:23

up by perverse incentives of what

5:26

would later become social media, where

5:28

it wasn't to help you, but

5:31

to essentially to hunt you, right? To extract something

5:33

from you using an asymmetric knowledge

5:35

about how the human mind works, which is

5:37

that your brain doesn't wake up to ask,

5:39

do I want to continue unless

5:41

it gets something like a stopping cue? What does that mean?

5:43

That means like generally you don't ask, do I want to

5:45

stop drinking wine until I get to the bottom of the

5:47

cup of wine? So my

5:50

invention got sort of sucked up by

5:52

a machine and waste on the

5:54

order of 100,000 human lifetimes per

5:57

day. Now it's, it's horrendous. And

6:00

this is what I think people miss all

6:03

the time in the invention of technology,

6:06

that it's not about the

6:08

intent, good or bad, of

6:10

the inventor. When you invent

6:12

a technology, you uncover a new class

6:15

of responsibilities. We didn't need the right

6:17

to be forgotten until the internet could

6:19

remember us forever. And then

6:21

two, if that technology

6:23

confers power, you're going to

6:26

start a race for that power. And if

6:28

there is some resource that we need, that

6:30

we depend on, that you can be exploited

6:32

for that power, in this case, like attention

6:35

and engagement with the attention economy, then that

6:37

race will end in tragedy unless you can

6:39

coordinate to protect it. And

6:41

so I was completely blind to that structure

6:43

when I was creating infinite scroll, and you

6:46

can see the results. That thing we call

6:48

doom scrolling would not exist without infinite scroll.

6:51

So I mean,

6:54

obviously, there's a tension between social

6:57

media business models, and what we think

6:59

is the humane option here. But a

7:03

lot of people hate doom scrolling, why have

7:05

we not seen a company yet experiment with

7:07

a limit on that? What would you do

7:09

at this point? How would you think about

7:11

solving this? It's a great question. So the

7:13

way we've often talked about the attention

7:15

economy is it's a

7:18

business model that is fundamentally about

7:21

getting reactions from the human nervous system.

7:23

You get people angry, you show them

7:26

things that they cannot help but look

7:28

at. So you would get them if

7:30

the incentive is to get reaction and

7:33

make reactive the human nervous system, it's

7:35

sort of obvious that we're going to

7:37

get polarization, narcissism, more outrage,

7:39

eventually democratic backsliding, like that's all

7:41

a predictable outcome of just make

7:43

the human nervous system more reactive

7:46

and get reactions from it. And

7:48

that's why we're able to call

7:50

it in 2013, building all the way up to

7:52

the social dilemma in 2020. And

7:55

so if we're going to think about solving

7:57

it, it's not a thing that an individual

7:59

company do. We get into that

8:01

paranoid logic. If we don't do it, we

8:03

lose the person who does. So you have

8:05

to do something to the entire space as

8:07

a whole so everyone can start competing for

8:10

the thing that is healthy and humane. I've

8:12

talked with multiple social media companies about this

8:14

over the last few years is just

8:17

run the A.B. test of let's

8:19

have people preset how many hours a day

8:21

or ideally minutes a day they actually want

8:23

to be scrolling and then it

8:25

just flags that their time is up. I think

8:27

to your point that would reduce

8:29

attention and engagement perhaps, but

8:32

it would also make people less angry at the platform

8:34

and I wonder if there's a net benefit there and

8:36

at least I would want to test that. I

8:38

think you probably have a better idea, but tell me what's wrong

8:40

with mine and then where you would go. There's

8:43

a different version of yourself before you started eating french

8:45

fries and a different version of yourself after you started

8:47

eating french fries. Before you've eaten one single french fry,

8:49

you're probably like, I don't know if I want to

8:51

have french fries. After you've eaten one, you're in a

8:54

hot state. You're just going to keep eating them until

8:56

they're gone and that's sort of the thing I think

8:58

you're pointing at. In psychology, we

9:00

would talk about this as a fundamental want-should

9:02

conflict. You know you should stop

9:04

scrolling, but in that moment you want to

9:06

keep doing it and it's hard to override

9:09

the temptation. The good

9:11

news about your should self is that although

9:13

it's weaker in the moment, it's smarter in

9:15

the long term. If we

9:17

can activate the should self in advance and pre-commit,

9:19

as you're saying to that target, the

9:22

probability should go up that you

9:24

would be willing to stick to that commitment

9:26

once you've made it. It's not a perfect

9:28

solution, but what I like about it as

9:30

an example is it's something one company could

9:33

try. They could be differentiating in a positive

9:35

way and doesn't require congressional intervention

9:37

or you know all of the companies to

9:39

form a coalition. Right. So tell me what

9:41

I'm missing there and what your more systemic

9:44

approach would look like. I

9:46

used to be addicted to both twitter and

9:48

to reddit and I'm like how

9:50

do I get myself off? So as

9:52

a technology maker, every

9:55

designer knows that retention

9:58

is directly correlated related with

10:01

speed. That is, the faster your website

10:03

loads or your app loads, the more

10:05

people continue to use it. Amazon really

10:07

famously found that for every 100 milliseconds

10:10

their page loads slower, they

10:13

lose 1% of revenue. Wow. So

10:16

using this insight, I actually wrote myself a

10:18

little tool that said the longer I scrolled

10:21

on Twitter, the more I used Reddit, the

10:23

longer I would get a random delay as

10:25

things would load. Sometimes it'd be sort of

10:27

fast, sometimes it'd get slower, the longer I

10:30

used it, the slower it would get. And

10:32

what I discovered is that this let my

10:35

brain catch up with my impulse.

10:38

And I would just get bored, do I really

10:40

want to be doing this? And it wasn't a

10:42

lot. It was like 250 milliseconds. It's human reaction

10:44

time. It gave me just enough time to overcome

10:46

the dopamine addiction loop. And within a couple of

10:48

days, honestly, my addiction was broken because I'm just

10:51

like, oh, no, I actually don't want to be

10:53

doing this. Wow. Okay, so

10:55

this is an ingenious invention. How

10:58

do I download it? And are you going to make

11:00

it widely available? For everyone listening to

11:02

this podcast, this is not a super hard thing

11:04

to make. I did it with my own personal

11:06

little VPN and proxy. If anyone wants

11:08

to come help build this thing, please,

11:11

it's not hard. And I think there's a big

11:13

opportunity. I just personally don't have the time. I

11:15

just made it for myself. So let's

11:17

put up that plea. Let's then

11:19

jump up to the global solution.

11:22

We need to hit these companies essentially in their

11:24

business model. We have to hit them where sort

11:27

of a scorecard for the

11:29

effects of social media on teen

11:31

mental health, depression, and suicide, on

11:34

the health of public discourse,

11:36

on the backsliding of democracies.

11:39

You make the big list of all the different kinds of

11:41

harms. And then you have some

11:44

way of evaluating how well

11:46

each company is doing. And

11:48

then you institute

11:51

some kind of latency

11:53

sanction. Like, all right, looks

11:55

like Facebook's doing really badly on teen

11:57

mental health. We're going to in

12:00

a democratic way, slow Facebook down after the

12:02

first n number of minutes. It'll never get

12:05

more than, I don't know, 500 milliseconds of

12:07

delay. It's not like you're stopping it. There's

12:09

no censorship. You're just saying, hey, you're having

12:12

negative action allies, you're affecting the whole, and

12:14

you would imagine how very quickly these

12:16

companies are going to innovate their way

12:19

to solving those metrics, like quarter after

12:21

quarter. The first thing,

12:23

of course, I'm sure in people's minds is

12:25

that seems really scary. Like who'd want to

12:27

give the government the ability to slow down

12:29

websites? This speaks

12:31

to the next thing, which is

12:33

especially in the era of exponentially

12:35

powerful technology as we move into

12:37

AI, we're going to need forms

12:39

of trustworthy governance that are hard

12:41

to capture. We're going to need

12:43

to have the equivalent of citizens'

12:45

juries or other things, like other

12:47

forms of institutions which are hard

12:50

to capture and are capture and

12:52

corruption resilient. This, I think, would

12:54

be an excellent place to start prototyping

12:56

what that future vision of what resilient

12:58

institutions would look like. I

13:00

actually think that would be really compelling. Note

13:03

that this is a solution that

13:05

never touches content. It never touches

13:07

content moderation. It never touches censorship.

13:10

It's a solution born out of seeing

13:12

the world as incentives leading to outcomes

13:14

and trying to shift things at

13:17

the incentive level so that you

13:19

can unleash the amazing amount of

13:21

creativity and ingenuity inside of

13:24

these companies that are just doing the

13:26

thing that the incentives tell them to do.

13:29

For uninitiated listeners, can you explain what

13:32

a bridging algorithm does? You

13:34

have used Twitter and

13:37

have seen community notes. This is

13:39

actually a bridging algorithm in

13:42

practice. Essentially what

13:44

bridging algorithms do is they look

13:47

for consensus across groups that

13:49

normally disagree. Once it

13:51

finds statements that

13:53

people across multiple different

13:56

divides agree on, then

13:58

it raises that up. It sort of

14:00

promotes those. One of my favorite

14:03

examples of thinking about this is something

14:06

called the perception gap. And

14:08

the perception gap is

14:11

how differently I

14:13

perceive you and your

14:15

beliefs than your actual beliefs.

14:18

When we are fighting, we are

14:20

often not actually fighting with the other side,

14:22

we are fighting with a mirage

14:24

of the other side, a sort of

14:26

a caricature. And now we

14:28

end up in a really interesting place because

14:31

we could start to measure what

14:33

kind of content at scale increases

14:36

perception gap, that is, sort of fills

14:38

you with false beliefs about the other

14:40

side beliefs. And which kinds of content

14:43

decreases, helps you see more accurately. And

14:45

you could imagine then an algorithm

14:48

which helps go viral, the content

14:50

that lets us see each other

14:52

correctly. It doesn't make all

14:54

disagreements go away, but it says at

14:56

the very least, we should be able

14:58

to accurately see what all

15:00

the other sides are saying.

15:03

And because we are actually closer than we believe,

15:05

it's sort of like bringing the two sides of

15:08

a wound closer together so it can start to

15:10

heal. I think you alluded

15:12

to your skepticism that a social

15:14

media platform would like this idea

15:16

because outrage is more

15:19

activating in some ways than connection.

15:21

That's right. We're not entirely convinced.

15:24

I just wonder if we haven't tried the

15:26

right approaches to bridging yet. So for example,

15:29

there was some evidence that was published a few

15:32

years ago showing that people would rather have a

15:34

conversation with a stranger who shared their political views

15:37

than a friend who didn't. And

15:40

I think people recognize that as a massive problem

15:42

in their lives. If we think about family members

15:44

and friends and close colleagues who are

15:46

not speaking to each other or having a hard time

15:49

getting on the same page, that's

15:51

an audience for a bridging algorithm. Facebook,

15:53

they discovered a very simple thing that

15:55

they could do for fighting

15:58

hate speech, disinformation, misinformation, and other

16:00

things. information, all the worst

16:02

stuff, what was that one

16:04

simple thing that they could do? It

16:07

was they could remove the reshare button

16:10

after two share hops. That is, I

16:12

could share something, somebody else could click

16:14

the reshare button, somebody

16:16

else could click the reshare button, but after

16:18

that the reshare button would disappear. Now, if

16:20

you're really motivated, you could copy the text

16:22

and paste it again. So again, no censorship,

16:24

it's just introducing a

16:26

little more friction. But

16:28

it comes at the cost of engagement. That's

16:31

why I think we're not going to

16:33

see much traction with bridging algorithms until

16:35

those fundamental incentives are fixed. Did

16:38

you co-coin the phrase, freedom of speech is

16:40

not freedom of reach? I

16:43

coined it and then Renee DiResta,

16:45

a brilliant researcher at the Internet

16:47

Observatory now, she wrote an

16:49

article in Wired that popularized it. So

16:52

I love that phrase. There's a

16:54

certain level of reach that concerns

16:56

me when the content is consequential. So

17:00

thinking about during COVID, health

17:02

information and misinformation or disinformation,

17:05

thinking about posts that

17:07

are safety relevant in

17:09

an area where there might be danger or

17:11

a threat of violence. I've

17:14

often wondered when health and

17:16

safety information reaches a certain

17:18

level of virality, why

17:20

isn't it flagged to not be re-shared

17:23

unless it's fact checked? And

17:25

why isn't there a process for that? Is something like

17:27

that viable? Other countries do

17:29

this. So it is viable. I'm

17:32

thinking of Sinan Aral's work showing that

17:34

lies spread faster and farther than the

17:36

truth. Falsehoods go six times

17:38

faster than truths. This is really important because

17:41

one of the things we want to sidestep

17:43

is, is this piece of

17:45

content true or false? Fact

17:48

checking is hard. And then the

17:50

thing that's next to fact checking is frame

17:52

checking. And now it

17:54

gets very hard to adjudicate. So we

17:56

should be looking not at specific

17:59

pieces of content. content, but at the

18:01

context surrounding them, how fast they're spreading, what

18:04

is the way that they're spreading, what are

18:06

the incentives for it spreading, and so

18:08

that we can move out of the morass

18:10

of free speech. Because as soon as

18:12

we head down that pathway and solutions that

18:15

require a debate about free speech, we lose.

18:18

I'm realizing it's not so simple as

18:20

just fact checking. Frame

18:22

checking is almost impossible. Yes, that is

18:24

exactly right. And that is why that

18:26

whole program of let's get more fact

18:29

checkers in is just going down the

18:31

wrong solution branch. And so we need

18:33

to be thinking about it at a

18:35

more systems level,

18:37

incentive level, context level

18:40

than content. Congress

18:43

realizes they're a bunch of luddites. They

18:45

put you in charge of a committee to

18:47

make a series of recommendations for what ought

18:49

to be done, societally.

18:54

What are you proposing? We are

18:56

at the cusp of the next era

18:59

of technology, of AI. Which

19:01

way is it going to go? Are we going to

19:03

get the incredible promise of AI, or are we

19:05

going to get the terrifying peril of AI? And

19:09

our point was the same as it's always been, which

19:11

is if you

19:13

want to understand where it's going to

19:15

go, look to the incentives. That's how

19:17

we're able to predict social media. So

19:19

what are the incentives for AI? To

19:23

grow your company as fast as possible,

19:25

to grow your capabilities, get

19:27

them into the market as quickly as

19:29

possible for market dominance so you can

19:31

sort of like wash, rinse, repeat. And

19:33

the shortcuts you're going to take are

19:35

always going to be shortcuts around safety.

19:37

And we are going to recapitulate all

19:40

of the problems of social media, just

19:42

orders of magnitude bigger. And

19:45

the way we like to say it is that social

19:47

media was actually humanity's first

19:50

contact with AI. And

19:52

whereas AI in social media, it's the

19:55

thing that sits behind the screen choosing

19:58

which posts and which videos. hit

20:00

your eyeballs and your eardrums. It's

20:03

the algorithm. It's the algorithm. And

20:05

it was a very simple, unsophisticated

20:07

version of AI, and

20:09

its small misalignment, optimizing for the

20:12

wrong thing, sort of broke

20:14

our world. So tell

20:16

us, what might you do? If

20:19

I could wave a magic wand and say,

20:21

all right, every one of these major AI

20:23

companies, there needs to be some way for

20:25

them to give, I don't know, 25%, 40% of their

20:28

compute to

20:34

forecasting all of the foreseeable

20:36

harms that they

20:38

can possibly foresee using the

20:41

new sort of cognitive labor

20:43

that AI affords, so

20:46

that there's now some kind of appropriate

20:48

liability for not doing

20:50

enough to constrain

20:53

those foreseeable harms. And

20:55

then you could imagine, we're gonna need

20:58

some kind of graduated sanctions, and the

21:00

sanction comes in the form of, like,

21:03

I don't know, like a compute tax or

21:05

something like that. This is very, like, much

21:08

of a sketch, because this is hard, but I'm just

21:10

trying to give a flavor of how we might start

21:12

to think about it in a way that's at the

21:14

incentive layer, not at what is a specific thing that

21:16

one company can or cannot do later. I

21:19

think what's tricky about it then is,

21:21

okay, you're gonna pour all those

21:24

compute resources into anticipating risks, and

21:27

then you're not just gonna rely on

21:29

the AI to decide which

21:31

of those risks are high versus

21:33

low probability, or high versus low severity.

21:35

We need to bring in human judgment.

21:38

And then what do we do when we have

21:40

a low probability, high

21:42

severity threat? This work gets

21:45

very, very challenging, because we as

21:47

human beings are not very good

21:49

at, like, emotionally

21:51

attuning to tail risks, especially when

21:53

on the other side of the

21:55

equation, right, because the AI could

21:57

enable, like, terrible bio.

22:00

weapons and race-based viruses, a whole

22:02

bunch of terrible things. And you

22:04

can imagine AI just increasing all

22:06

of those tail risks. But on

22:08

the other side, we get incredible

22:10

benefits and the benefits are concrete,

22:12

like cancer drugs. They happen for

22:14

you immediately versus these risks, which

22:16

are diffuse, probabilistic,

22:19

amorphous, and our brains just

22:22

can't deal with that trade

22:24

very well at all. I'm actually very curious what you

22:26

would say about it. How do you make those kinds

22:28

of risks? It's a really good question.

22:30

It's a hard one. Frankly, I don't think

22:32

we've cracked it yet. I

22:34

want to just delegate the problem to

22:37

Phil Tetlock and his team of duper

22:39

forecasters and say, okay, we have individuals

22:41

who have demonstrated a consistent ability to

22:43

do this. So let's treat them

22:45

as one of your juries. They

22:48

know what they know and they know what they don't know. We

22:50

know a lot about how to train people to be

22:52

better forecasters. A second is, I

22:54

think we can make some of

22:56

this probabilistic information easier to

22:58

digest. I think of the

23:00

work of Gerd Gigerändzer, for example, and colleagues where

23:04

they've shown that natural frequencies are

23:06

easier to process than statistics. And

23:09

so instead of saying that something is 0.1% odds, say this

23:11

is one in a thousand

23:15

and all of a sudden people are more

23:17

likely to take it seriously. Like, wow, that

23:19

could happen. You have to

23:21

make them visceral. You have to feel it. What

23:23

we need is a process. And

23:27

this is where I think it

23:29

gets really exciting because the United

23:32

States was founded on the principle

23:34

that we could build a

23:36

more trustworthy form of governance. No

23:39

one really, I think, deeply trusts

23:41

like the institutions that we have now.

23:44

And if we just handed power to

23:46

regulate AI to the government,

23:48

it would probably mess it up in some

23:50

way. There's probably some kind of deep centralization

23:52

of power that would happen. And that's super

23:54

scary in the era of essentially forever

23:57

dystopias. There is no such thing as

23:59

privacy. in the AI

24:02

world, everything that can be decoded

24:04

will be decoded. Governance needs to

24:07

scale with AI because otherwise, as

24:09

AI increases its intelligence, you're driving a car

24:12

whose engine is going faster and faster, but

24:14

your steering wheel is going faster and faster,

24:16

that thing is going to break. And we're

24:18

going to need a way of having human

24:21

collective intelligence scale with

24:23

AI. Otherwise, AI will overpower human

24:25

collective intelligence, which is another way

24:28

of saying we lose control. And

24:30

obviously, this is a complex, hard

24:32

topic, and they're like mini-publics.

24:34

And Audrey Tang's work from Taiwan,

24:36

I think, is the best living

24:38

example of how you can put

24:40

these values in to practice.

24:43

That it isn't just sort of like

24:46

a theoretical framework. She,

24:48

with a whole community, has built the

24:50

tools that do these kinds of bridging

24:53

algorithms we're talking about, so

24:55

that citizens can

24:58

set the agenda for

25:00

government to have to listen

25:02

to. They did this for,

25:04

say, how should Uber and

25:06

other ride-sharing apps, how should

25:08

they integrate into society? And

25:11

they asked everyone in society

25:13

to give them the ability to

25:15

contribute what their values are, what they care about.

25:17

And it's a lot of these incredible

25:20

little design philosophies that I

25:22

find super fascinating. Like, in

25:25

her system, there isn't a

25:27

reply button. You can only say your

25:29

value. And if you disagree, you don't

25:31

disagree. You just have to state in

25:33

the positive your value. And

25:36

now you have this big map of everyone's

25:38

values, so that when you then can thumb

25:40

up something, be like, oh, yes, I agree

25:42

with this idea, they can use

25:44

a bridging algorithm to say, well, we know

25:46

what everyone's positive value statements are. So let's

25:48

find the policy

25:51

or the agenda that we really care about

25:53

that sits across those divisions

25:56

in our society. So we're finding the things

25:58

that knit and heal versus the... things that

26:00

divide. Adri Tang was a

26:03

software programmer. And then they

26:05

created a new position for her, and now

26:07

she's Taiwan's first ever Minister of Digital Affairs.

26:10

Yes. Why are we not doing that? I

26:13

think honestly it's an imagination gap.

26:16

We cannot imagine a

26:18

system different than we

26:21

have now. It turns out there are a

26:24

huge number of really brilliant people working on these

26:26

kinds of things. So if

26:28

the US government were to put, I don't know,

26:31

let's just say $10 billion per year into

26:35

upgrading democracy

26:38

itself, not just digital democracy

26:40

adding more forms online, but

26:42

let's do the most American

26:44

thing, which is innovate our

26:46

way to a new form of what

26:49

democracy looks like. I want to

26:51

vote not left, not right, but upgrade. So

26:53

if I were to draw a Venn diagram of a

26:55

techie and a hippie, do you live in the middle?

26:58

I spent a lot of time in nature and

27:00

there is something profound about feeling

27:02

the smallness of your breath against the

27:05

largeness of the universe.

27:08

But I don't know if I'd say I'm a hippie and I don't know if I'd say I'm

27:10

a techie. Well that actually is

27:13

a great segue to where I wanted to

27:15

go. I was stunned when you said

27:18

last summer that you thought it might be possible

27:20

for us to understand whales

27:22

and maybe even talk to them one

27:24

day. Why do you want to

27:26

communicate with whales? We're

27:28

trying to talk to whales and already that's

27:31

not why we're trying to do it. We do

27:33

not change when we speak. We change

27:35

when we listen. The goal for

27:38

Earth Species Project is to

27:40

learn how to listen to

27:42

whales and orangutans and parrots,

27:45

the other non-human cultures

27:47

of Earth, sometimes which have been

27:50

communicating for 34

27:52

million years, passing down languages

27:54

and dialects and cultures because

27:57

whatever it is that is the solution to the

27:59

Earth. humanities problems, I'd argue it's

28:01

probably not in our imagination, because if

28:03

it was, we'd be doing it. So

28:05

what we're looking for are things that

28:08

are outside the sphere of

28:10

human imagination. And just to preempt, I

28:12

think, your listeners' questions, like we're talking

28:14

about animal languages, does such a thing

28:16

even exist? And I just want to

28:18

give a couple quick examples that

28:21

I think will, like, help illustrate this.

28:23

Many animals have names that

28:25

they will call each other by,

28:27

sometimes even in the third person.

28:29

Parrot parents will spend the

28:32

first couple of weeks of their

28:34

chicks' lives leaning over and whispering

28:36

into each of their individual children's

28:38

ears a unique name, and

28:41

the children will sort of, like, babble back

28:43

until they can get it, and they will

28:45

use that unique name for the rest of

28:47

their lives. Mind blown. And

28:51

then, just to give another

28:53

example, a 1994 University Hawaii

28:55

study where they were teaching

28:57

dolphins two gestures. And the

29:00

first gesture was, do something you've never

29:02

done before. And it takes a lot

29:04

of patience and a lot of fish to, like, communicate that

29:06

idea to a dolphin, but they will get it. Were

29:09

you a kid who was obsessed with Aquaman? What's

29:13

the origin story of this? I was a kid

29:15

that was obsessed with everything. It

29:17

must have been a very annoying

29:19

kid. This idea really came, actually,

29:22

from hearing a story on NPR

29:25

about this incredible animal, the

29:27

gelada monkey. The researchers said

29:29

they had one of the

29:31

largest vocabularies of any primates,

29:33

humans accepted. And when you

29:35

listen to them, they sound like women and children

29:37

babbling, and they sort of do turn taking, and

29:39

it's this complex vocal thing. And she's like,

29:41

we don't know what they're saying, but I

29:43

swear they're talking about me behind my back.

29:45

They were out there with like a hand

29:47

recorder, hand transcribing, trying to figure out what

29:50

they were saying. And the thought

29:52

sort of struck me like, why are we

29:54

using machine learning to translate? And

29:56

that changed in 2017, when suddenly AI developed

30:01

the ability to translate between

30:03

human languages without

30:05

the need for any Rosetta stone

30:07

or any examples and that's the

30:09

moment that it was time to

30:11

start Earth species project, start actually

30:14

going out to the field and learning

30:16

from biologists and the why really grew

30:18

with it. When I look out at

30:21

the structure of humanity's largest

30:23

problems, like I think

30:25

there's a connective thread between all of

30:28

them, whether it's the opioid epidemic

30:30

or the loneliness epidemic or climate

30:33

change or inequality, it

30:36

always takes the form of a narrow

30:38

optimization at the expense of a whole. Some

30:41

part of the system like

30:43

optimizing whether it's for GDP at

30:45

the expense of climate or

30:48

whether it's trying to grab people's

30:50

attention at the expense of mental

30:52

health and backsliding democracies. It's

30:54

always a narrow optimization that breaks

30:57

the whole and Earth species is

30:59

fundamentally about reconnection and a narrow

31:01

optimization at the expense of the

31:03

whole is fundamentally a different

31:05

way of saying that is disconnection from

31:08

ourselves, from each other, the natural world,

31:10

a disconnection of our systems to their

31:12

large-scale effects. When you think about

31:16

many of the just so stories, indigenous

31:19

myths, they almost always start out with

31:21

human beings talking with nature,

31:23

talking with animals, and that

31:25

moment of disconnection is symbolized by

31:28

the moment we can no longer communicate

31:30

with nature. This isn't just a question

31:32

of what we

31:34

must do. Fundamentally, this is a question

31:36

of who we must be, like

31:39

to change our identity, to

31:42

change the stories we tell ourselves in

31:44

order to live, to change our myths,

31:46

to reconnect us. At the deepest

31:49

level, that's the hope of what

31:51

Earth species can help bring

31:54

about and just to name and self-awareness that

31:56

no one thing can do this. There is

31:58

no silver bullet, but maybe there is

32:00

silver buckshot. Hey,

32:04

rethinking listeners, we're supported by our

32:06

friends at Working Smarter, a new

32:09

podcast from Dropbox exploring the exciting

32:11

potential of AI in the workplace.

32:14

Working Smarter talks with founders, researchers, and

32:16

engineers about the things they're building and

32:18

the problems they're solving with the help

32:20

of the latest AI tools. Tools

32:23

that can save them time, improve collaboration,

32:25

and create more space for the work

32:27

that matters most. On Working

32:30

Smarter, hear practical discussions about what AI

32:32

can do so that you can work

32:34

smarter, too. Listen to Working

32:36

Smarter on Apple Podcasts, Spotify, or

32:38

wherever you get your podcasts. Or

32:41

visit WorkingSmarter.ai. Hi,

32:45

I'm Ben. I suffer

32:47

from a condition called writer's block.

32:50

It strikes when I'm at work. That's

32:52

why I choose Canva MagicRide. It

32:55

works bad. Generating text in

32:57

seconds. It works to AI. Common

33:00

side effects include increased productivity, compliments

33:02

from coworkers, feelings of satisfaction. Now

33:05

I can say bye-bye to writer's

33:07

block. Ask your boss if Canva

33:09

MagicRide is right for you at

33:11

canva.com, designed for work. Let

33:16

me suggest now we go to lightning round. What

33:18

is the worst advice you've ever gotten? That

33:21

feeling in your body is telling you be careful

33:24

or something's up. Don't listen to that.

33:27

Push through. Oof. If

33:30

you could talk to any animal species, which

33:33

one would you choose? If

33:35

you were to talk to everyone on the Earth species

33:37

team, each person would have a different animal they're most

33:39

excited about. But for me, it's

33:41

beluga. Because beluga,

33:44

if you actually listen to them, they

33:46

sound like nothing you're expecting. They sound like

33:49

an alien modem. The

33:51

cultures of belugas and dolphins and

33:53

whales, they go back 34 million

33:56

years for something to have survived 34

33:58

million years of cultural history. evolution,

34:00

there has to be some deep wisdom in

34:02

there. I am so curious

34:04

to get the very first glimpses of

34:06

what that might be. You're

34:09

in conversation with a beluga whale. If

34:12

you could ask one question, what would it be?

34:14

I'd want to know, what does

34:16

it feel like to be them? What

34:18

is the question you have for me? You prove,

34:21

okay, animals think

34:23

they have language, there's an interiority.

34:26

What for you changes? What do you think the implications are?

34:30

I guess my hope is that we start

34:32

to realize that we need to do a

34:34

much better job, both avoiding

34:37

harm to and taking care of

34:39

species that aren't human. And

34:42

that this is a watershed moment. The

34:46

skeptical side of me says, we've

34:48

tried this with a lot of human cultures

34:50

and failed pretty much every time. It's

34:53

so easy to dehumanize people that

34:56

we already know are sentient and

34:58

entire groups that we already know

35:00

feel extreme pain. Why

35:03

would it be any different with animals? Whenever

35:05

I think about that, I'm like, well, it is

35:07

true that even though we know other

35:09

humans speak, we still do terrible things to

35:12

them. And imagine how much

35:14

worse it would be if they couldn't speak at all.

35:17

You mentioned earlier whales,

35:19

orangutans, parrots. How

35:22

did you go about deciding which animals? A

35:24

lot of which animals we decide to

35:27

work with are driven by the deep

35:29

insights of the biologists that have been

35:31

out there in the field. So

35:33

for instance, why start thinking about orangutans?

35:36

It's because one of our partners, Adriano

35:39

Lamera, was able to

35:41

show in the last couple of years that

35:43

orangutans have a kind of past tense. They

35:46

can refer to events that happened

35:48

at least up to 20 minutes ago. It's probably longer,

35:50

but that's as far as he's been able to show

35:52

so far. And when you think about language,

35:54

two of the big hallmarks of language are being

35:57

able to talk about things that are not here

35:59

and not now. Parrots, as we're talking

36:01

about, they have names, they call each other by. And

36:03

I honestly think even

36:05

just a campaign that let the world

36:07

know that animals have names, like that

36:10

would already start to shift human culture

36:12

and how we relate. I

36:14

think probably for a long time

36:16

I assumed that cognitive

36:19

capabilities tracked with vocal range.

36:22

But we all know that's not true. Parrots,

36:24

they can say incredible things. I

36:27

don't think their thinking capacity

36:29

is anywhere near what a dolphin is,

36:31

for example. How do

36:33

you weigh those two sets of factors? I'll

36:36

push back a little bit. There was a

36:38

nature publication maybe

36:40

three, four years ago now where they're

36:43

looking at ravens and crows and

36:46

their cognitive capabilities compared to, say,

36:48

the great apes. And they're

36:50

on par. This is the general thing

36:52

we find, which is as human beings, our

36:54

ability to understand is limited by our ability

36:56

to perceive. And generally speaking,

36:59

we just haven't been perceiving

37:02

enough. It seems like

37:04

this is long overdue because I've looked for

37:06

years at these supposed

37:09

intelligence rankings of animals and

37:11

said, well, this is just a function of the tasks that

37:13

we've given. Yes, exactly. And the way that we know how

37:16

to score them. And it's really

37:18

easy to discover that a

37:20

pigeon is dumb if you don't give

37:22

it a navigation task. Yes. And

37:24

then all of a sudden you do and you realize, wow, it's

37:26

a lot smarter than us when it comes to finding its way

37:29

around the world. And I wonder how

37:31

many species we've underestimated that way. One

37:33

of my favorite examples of this comes from

37:35

the mirror test is when you take

37:37

an animal and you paint a dot on them where they

37:40

can't see and they're unaware of it, they look in the

37:42

mirror and then they start trying

37:44

to get that dot off of them or investigate

37:46

it. It's a test

37:48

that tests self-awareness. They

37:50

have to look into a mirror and say,

37:53

oh, that image

37:56

in the mirror, that is me. step

38:00

to take. It means there's an interiority and a sense

38:02

of self. It was thought for the longest time that

38:04

elephants couldn't pass the mirror test, but

38:07

then it was turned out

38:09

that it's just because scientists were using small mirrors.

38:12

No. Right? It's

38:14

just like if you measure the thing wrong, all

38:16

it needed was a bigger mirror, then

38:19

suddenly what looked unintelligent becomes very

38:21

intelligent. You were really careful

38:23

to stress that we should just listen or

38:26

that listening is the primary goal. It's at

38:28

the center. It's the primary goal exactly. As

38:30

soon as we're capable of deciphering and understanding,

38:32

someone is going to want to communicate. What's

38:36

your answer to the question of should we

38:38

open Pandora's box? Because I feel like the

38:40

standard Silicon Valley response to this is not

38:42

satisfying. It's, well, somebody else is going to

38:44

do it if we don't and we're more

38:46

ethical than they are. So we need to

38:49

do it first. Which to me is just

38:51

dripping with narcissism and arrogance. Yeah. It's the,

38:53

like, well, I want to do it. So

38:55

I'm going to find the belief that

38:57

lets me do the thing I want to do. Exactly.

39:01

So why do you want to open the box

39:03

despite that risk? We're going to uncover a whole

39:06

bunch of new responsibilities about what does it mean

39:08

to be able

39:10

to communicate with the other

39:12

non-human cultures of Earth. And

39:15

of course, if it confers any kind

39:17

of power, it's going to start a race and that race will

39:19

end in tragedy. So I think

39:21

to be a sort of

39:23

humane technologist or responsible technology, really

39:25

just be to be a technologist

39:27

means to pre-think through all the

39:29

ways you're going to start some

39:31

kind of race. What are the

39:33

ways that your technology is going

39:35

to be abused or cause harm?

39:38

We might create like a whale QAnon or something.

39:40

We don't know. So

39:42

we need to be really careful about

39:45

going out and starting to just speak

39:47

in the same way. You could imagine

39:49

factory farms using it. You could imagine

39:51

poachers using it to attract animals.

39:53

You could imagine ecotourism

39:55

using it to attract animals. So there is

39:58

no such thing as a technology. doesn't

40:00

have externalities and doesn't have

40:03

bad actor abuses. So

40:06

what do we do? So that means we

40:08

need to race ahead and start thinking about

40:10

what are the international norms

40:13

and treaties and laws and other

40:15

things that can bind those races.

40:17

I think we're going to need

40:20

whatever the equivalent of a Geneva

40:22

Convention for cross-species communication is. And

40:25

to give another example, when we started our

40:28

species, we were doing everything open source.

40:31

We're like, it's good to get these

40:33

models out to as many of the

40:35

scientists as possible because as we

40:37

build the tools to decode animal communication

40:40

and translate animal language, we're also building

40:42

the tools that it turns out all

40:44

biologists need just to do their work and

40:46

their conservation work. And we've realized, actually,

40:48

that was a naive

40:50

value, that we can't just

40:53

open source everything. We're going to have to go through

40:55

a gated release. So as we build these models, we're

40:57

just not going to ship them to everyone. There's going

40:59

to have to be some kind of application process. And

41:01

then we're going to have to start thinking through, and

41:03

this is not just for us, but the wider space,

41:06

what is the right way so

41:08

that we as one entity

41:11

can't sort of abuse our centralized power? How

41:13

do we find these processes that we've been

41:15

talking about that make it a trustworthy process

41:18

for who gets access to the models? How

41:21

do you think about the problem of privacy

41:23

violations? In

41:26

general or for animals? For

41:28

animals. I'm thinking that you're trying

41:30

not to disrupt or disturb whales by

41:32

listening in. Yeah. But they

41:34

also didn't give you permission to listen in. In the

41:36

process, if we learn how to ask

41:39

whether we're violating consent, then we

41:41

can actually just ask and find out. If

41:43

we think about whales, for example, what year do

41:45

you think it'll be when we

41:47

can understand everything that they're saying, or

41:49

maybe not everything, but where we

41:52

can decipher a significant

41:54

chunk of their communication? And of

41:56

course, we're talking about science here, so I just want to

41:58

call out any prediction. of where

42:00

we're going to be, but we are this

42:03

year heading into our first non-wild

42:06

two-way AI to animal communication experiment.

42:09

And we're seeing, can we essentially

42:11

pass the Turing test for a

42:14

specific kind of sogginbird, a zebra

42:16

finch? Can you swap one zebra

42:18

finch out for the AI zebra

42:20

finch and see if the actual animal can tell

42:23

the difference? What sound does an elephant

42:25

make or two elephants make when they come together? And

42:27

that means something about greeting, affiliation, but maybe it means

42:29

they miss you, or maybe it means I'm glad to

42:31

see you. Maybe it means their

42:33

name, but you can see, okay, well, it happens with one

42:35

elephant is running really quick and running

42:37

its ears, and we know that that has

42:40

emotional conjugation to it. And

42:42

so you can see that as we start to pass

42:44

the Turing test, we get towards decoding pretty quickly. This

42:47

is a real frame shift. We started out, and

42:49

I was under the impression that the goal

42:51

is to learn something that can benefit humanity.

42:55

And also, it would be really nice if it got

42:58

us to be kinder to animals.

43:01

It did not occur to me that you're in a position

43:03

to actually help the very animals

43:05

that you're communicating with. So the idea, for

43:07

example, that you could put a warning device

43:09

on every major ship that

43:11

would signal to any underwater creature, get out

43:13

of the way. That

43:16

could be very meaningful. You could potentially

43:18

do the same thing in a rainforest with

43:20

birds, right? Is one

43:22

of your aspirations to be able

43:24

to use some of what you learned to actually

43:27

save species from extinction? Yes,

43:30

absolutely. And I just want to paint a

43:32

picture in everyone's head for what might a

43:34

translation look like? Because are we

43:36

just talking about a Google Translate, and you say whatever you

43:38

want, it comes out. And it probably won't look

43:41

like that. I think there are parts of

43:44

the experience that we share with animals.

43:46

We know that whales will carry their

43:49

dead children for up to three weeks.

43:51

Pilot whales do this. And

43:53

it looks like grief is a shared

43:55

experience. But then there are huge portions

43:58

of their experience that we might not. never be

44:00

able to directly translate. Like

44:03

Spermwell spent 80% of their

44:05

life a kilometer deep in

44:07

complete darkness and seeing in

44:09

3D sound. That's

44:12

not anything in the human experience.

44:15

What might those translations be? I think those

44:17

translations are likely to be much more poetic.

44:19

It might be a snatch of music with

44:22

a specific kind of color. It's

44:24

some kind of multimodal translation.

44:27

We won't know what it means exactly, but

44:29

we will get a sense over time

44:31

and maybe it'll be our children who grow

44:34

up immersed in these odd translations from other

44:36

beings and other cultures that they're like, oh,

44:38

I get it. I have a sense for

44:40

what that thing is. Things to look

44:43

forward to and brace ourselves

44:45

for. Yeah, exactly. Awesome. Thanks,

44:47

Zaeza. Thank you so much, Adam. This

44:49

was super fun. To be continued. Agreed.

44:54

As we're waiting for digital platforms to

44:56

evolve, I have one thought about a

44:58

small step we can each take to

45:00

reduce polarization and misinformation. People

45:03

often say, I'm entitled to my opinion.

45:06

I want to rethink that. Yes, you're

45:08

entitled to your opinion in your head. But

45:10

if you decide to share that opinion, it's

45:13

your responsibility to change your mind when

45:15

you come across better logic or better

45:17

evidence. Rethinking

45:22

is hosted by me, Adam Grant. This

45:24

show is part of the TED Audio

45:27

Collective. And this episode was produced and

45:29

mixed by Cosmic Standard. Our

45:31

producers are Hannah Kingsley Ma and Asia Simpson.

45:33

Our editor is Alejandro Salazar. Our

45:36

fact-checker is Paul Durbin, original music by

45:38

Hans-Dale Stu and Alison Leighton Brown. Our

45:41

team includes Elijah Smith, Jacob

45:43

Winnick, Samaya Adams, Michelle Quint,

45:45

Banban Cheng, Julia Dickerson, and

45:47

Whitney Pennington-Roggers. Humpback

45:56

Whales, their song goes viral. And

45:59

for whatever reason, they're not. and Australian humpbacks are

46:01

like the K-pop singers and their songs

46:04

will spread, and we don't know why, over

46:06

the entire world within a couple

46:08

of seasons sometimes and then everyone

46:11

is singing like the Australian pop

46:13

songs. Support

46:15

for the show comes from Brooks Running.

46:17

I'm so excited because I have been

46:19

a runner, gosh, my entire adult life

46:22

and for as long as I can remember, I

46:24

have run with Brooks Running

46:27

shoes. Now I'm running with a

46:29

pair of Ghost 16s from Brooks,

46:32

incredibly lightweight shoes that have

46:34

really soft cushioning, it feels

46:37

just right when I'm hitting my running

46:39

trail that's just out behind my house.

46:41

You now can take your daily run

46:43

in the better than ever Ghost 16,

46:46

you can visit brooksrunning.com to learn more.

46:52

PRX.