0:01

I'm Dr. Brian Goldman, host of the CBC

0:03

podcast, The Dose. Each week

0:05

we answer vital health questions that will help

0:07

you thrive, like, what does my mental health

0:10

have to do with my gut? How

0:12

can I prevent melanoma? How much sleep

0:14

do I really need? And how can

0:16

I manage my health without a family doctor?

0:19

I chat with the top experts to bring you

0:21

the latest evidence in plain language, all in about

0:23

20 minutes. Join The Dose on

0:25

the CBC Listen app or wherever you get your

0:28

podcasts. This

0:31

is a CBC Podcast. Hi,

0:35

I'm Nora Young. This is Spark. If

0:37

spoken language is a sort of spontaneous

0:40

creative act between speaker and listener, will

0:43

machines ever truly match our

0:45

playful linguistic games? And

0:47

in the cat and mouse game between

0:49

humans and content-moderating algorithms, we creative humans

0:51

seem to have the edge for now.

0:54

Today in an episode that first aired

0:57

in May of 2022, from the quill

0:59

to the telegraph to morphing memes on

1:01

social media, how the tech we use

1:03

changes the way we communicate. Imagine

1:10

you run into a friend. Hey, how

1:12

was that party last night? And

1:14

you start telling them the story. Well,

1:17

first of all, the cake was

1:19

so fancy. But

1:21

in that story, there's a certain word.

1:24

And if you say that word,

1:26

cake, everything you just said disappears.

1:33

That's how it works online, except with words

1:36

a lot less wholesome than cake. On

1:39

social media platforms like TikTok and

1:41

YouTube, if your content contains certain

1:43

flagged phrases or banned words, your

1:45

post could be removed altogether. The

1:48

word police aren't people, they're

1:51

content-aware algorithms. But humans

1:53

are crafty and have developed a

1:55

workaround called AlgoSpeak. AlgoSpeak

1:58

is a direct derivative of English

2:00

or whatever language so that the

2:02

machine can't censor the user. This

2:05

is Jamie Cohen. He's an assistant professor of

2:08

media studies at CUNY Queens College in New

2:10

York City. The practice

2:12

of word swapping has helped content

2:14

creators fool algorithms for some time

2:16

now. For example, people would

2:18

often write panini instead of pandemic in

2:21

their posts. Now, the

2:23

point of this algorithmic monitoring is

2:25

to catch harmful terms, things

2:27

that might spread disinformation or

2:29

racist language or violent extremism. But

2:32

because it's automated, nuance and context

2:34

is often missing and speech that

2:37

shouldn't be flagged is. I'll

2:39

go speak comes directly from the TikTok app

2:41

because TikTok's app is a much more content

2:43

aware system than content moderation systems

2:46

like Twitter or Facebook or so forth,

2:48

which are usually algorithmic, but don't have

2:50

a content aware system. It's more of

2:52

a line that somebody has to report

2:54

you. Okay. On TikTok, the system, if

2:56

it hears you, literally hears you or

2:58

sees you say with the captioning words

3:00

like dead or sex, the

3:03

video clip either gets demonetized or unpublishable. A

3:05

term is shadow banned. Your clip stays online,

3:07

but it doesn't end up in the for

3:09

you page. Okay. Or it simply just gets

3:11

removed from the feed itself. And

3:14

that's for very minor infractions of language. So

3:16

what users have done is figured out ways

3:18

around speech or words. So instead

3:20

of the word sex, it would say segs, S

3:23

E G S or instead of saying dead, it

3:25

would be made on a made on alive. And

3:27

so these words when together make sense to the

3:29

reader, but the machine itself can't actually read them

3:31

in the way that the content moderation system

3:33

would. Okay. But is there any chance

3:35

that the content moderation sort

3:38

of learns that people are using made

3:40

on alive instead of dead? Yeah.

3:43

So when machines read media,

3:45

they learn that human behavior is learned by

3:47

machine learning. And so every time you use

3:49

a computer, whether you're using your Facebook feed,

3:52

whether you're using TikTok, whether you're using anything

3:54

that has machine learning engaged with it, it

3:56

gets better. This is like the same thing

3:58

as auto correct. typing autocorrecting, you have

4:01

a slang term that you use with a

4:03

friend, that eventually autocorrect stops correcting it because

4:05

it learns that that's the word that you're

4:07

going to be using. Now the downside here

4:09

is that algorithms like TikTok are going to

4:11

learn the workarounds very simply and then what's

4:13

going to happen is we're going to have

4:15

to evade them further and eventually

4:17

language will be disguised in a way that

4:19

will be somewhat unrecognizable to a reader. You

4:23

talked a little bit about TikTok

4:25

and its 4U recommendations. Can you expand

4:27

on that a little bit? How do the different ways that algorithms

4:29

work on different social platforms

4:32

shape the needs for users to disguise their

4:34

language? When I teach about algorithms in

4:36

class, we have to keep in mind that they're black boxes.

4:38

We have no real insight as to how the machine learning

4:40

was written, how the code is derived, who's

4:43

writing it, what the outcomes are. We do

4:45

know for a fact though that the algorithms

4:47

are written to sell us data, you know,

4:49

sell our data to advertisers and it works

4:51

back to sell us material. So we know

4:53

that the algorithm is a loop to kind

4:55

of move content back into our eye view

4:57

so that advertisers can get it. But unlike

4:59

Facebook and Twitter where the algorithm is designed

5:01

for attention or the attention economy, it's designed

5:03

to give the most grab for the first

5:05

few clicks that are supposed to cause engagement,

5:07

whether that's a like, a comment

5:09

or a share. TikTok's content aware

5:11

system is what's known as a time spent

5:13

algorithm. So it puts content in front of

5:16

your face that makes you not want to

5:18

leave the app. It makes you want to

5:20

keep watching. So the goal of TikTok is

5:22

an infinite, is infinity that eventually

5:24

you're just not going to leave your screen. You're

5:26

just going to keep thumbing up the next clip.

5:28

Okay. But users have become very aware as to

5:30

how the content aware system works. And they know

5:33

that if they spend too much time on any

5:35

given clip, the algorithm shifts to show them more

5:37

of that type of content. Whereas they know if

5:39

they skip a clip very quickly, the algorithm learns

5:41

they don't like that type of content. So we

5:44

train the machine in real time on

5:46

all of our social media. We're constantly training it.

5:48

This is what causes echo chambers, our inability to

5:50

get outside of our own feedback loop. Okay. But

5:52

when it comes to TikTok, the for you

5:55

pages, the primary way people interact with that

5:57

app and don't go to user profiles that

5:59

go specific. specifically to the feed. Taylor

6:02

Lorenz explicates this in her Washington Post piece,

6:04

but the 4uPage is the dream of a

6:06

TikToker. That's where everybody gets their views. So

6:08

you have to make the content for the

6:10

4uPage rather than for another user. The

6:13

design of AlgoSpeak seems in a way

6:15

to hinge on a kind of insider versus

6:18

outsider binary, right? You need to

6:20

get the language to understand what's going

6:22

on. Does this have implications

6:24

for who can engage on platforms like

6:26

TikTok? Yes, this causes

6:28

a nuance issue. So as we know,

6:31

as from a journalist or

6:33

an educational perspective, nuance is a hot

6:35

commodity. You know, it's not easy

6:37

for a reader to get meta

6:39

text or understand reading between the lines. And

6:42

so when we move to a place where

6:44

we're using AlgoSpeak, it does require a savvy

6:46

use of the internet. It's no longer like,

6:48

oh, just your basic access point. It actually

6:50

moves the access point further away from the

6:52

basic user and it creates it inside outside.

6:54

This is where like digital divides

6:57

actually really emerge. Like digital divides

6:59

of the early 2000s

7:01

were like a big thing. Everybody was talking about

7:03

the digital natives and digital immigrants, people who couldn't

7:06

speak to one another because of generational gaps. Now

7:08

everyone has access, but the esoteric nature

7:11

of the web, the specific use is

7:13

going to become inaccessible to some users.

7:15

And we're going to see a divide,

7:18

whether it's generationally or technically, from how we

7:21

use those social media apps. And

7:23

I mean, more broadly, what does it

7:26

mean for overall accessibility on social media

7:28

platforms if words in

7:30

a way have more than one meaning? Yeah,

7:32

this is where memes are really the... My

7:35

primary subject of study is memes. And the

7:37

reason I study memes is because they are

7:39

reductionist media. They take big concepts and they

7:41

make them very small and shareable. The downside

7:44

is by reducing them, you eliminate a lot

7:46

of the context. You remove a lot of

7:48

the understanding of it because it makes you...

7:50

A reader of a meme is required to

7:52

understand the reference. They have to know what

7:54

it's referring to in order to even understand

7:56

it. All those people do the same thing.

7:58

It will literally... cause us to speak

8:01

a mimetic language where culture is being shared.

8:03

But if you're outside the culture, it's almost

8:06

another language. It would almost be something foreign

8:08

to a user. In that

8:10

case, we're actually going to create these very

8:13

interesting and niche communities that will have

8:15

their own sub-languages on social media platforms.

8:18

Which, I mean, is interesting, but it also

8:20

kind of makes you question the whole point

8:22

of these places is they're supposed to be...

8:25

You know, we can argue whether this is actually

8:27

the case, but they're supposed to be these virtual

8:29

town squares where everyone can meet and have these

8:31

conversations together. Right. Yeah, and

8:33

it causes enforced tribalism. Another

8:35

problem with enforced tribalism in this case

8:37

isn't so much the organic nature of

8:40

that. I think part of using social

8:42

media or even a public square is that you have communities. You

8:44

have sub-communities and you have affinity groups,

8:46

people who like what you like. But

8:48

in this case, in the specific

8:50

algo-speak case, it's machines making those

8:52

communities happen, rather than people organically.

8:54

So instead of a bottom-up grouping,

8:56

like a subreddit, for example, a

8:58

subreddit is a bottom-up community that

9:00

creates their own affinity group. In

9:03

algo-speak, the machine itself is causing people to

9:05

hide from a machine splintering the

9:07

community into different factions that may or may

9:10

not be able to retrieve the original community

9:13

back. Does algo-speak actually work? To

9:15

get the sense that creators substituting

9:17

phrases or using codified words

9:19

can and do actually avoid detection by

9:21

the algorithm? To an extent, it does

9:24

work. Yeah. And we notice

9:26

this because heavy users, heavy producers on

9:28

TikTok have figured out their view count

9:30

does change based on how much they

9:32

engage with algo-speak. Now

9:35

what that means to the platform is questionable.

9:37

Being that it's a black box, we're not

9:39

sure how the algorithm works, or not to

9:41

be literally conspiratorial, we don't know if it's

9:44

by design. In other words, maybe algo-speak is

9:46

part of the process of the way that

9:48

social media operates. Maybe TikTok

9:50

likes this. Maybe TikTok appreciates this

9:53

change in behavior. We're not really sure what the

9:56

outcome is of that. What we do know is

9:58

that it does work. You can hide. from

10:00

language or detection like the example

10:02

that Taylor Lorenz brings up is LaDollarBean

10:04

which is instead of typing the

10:06

word lesbian into the caption it's L-E

10:09

dollar sign BN and the machine the

10:11

machine reader the over-bright reader says LaDollarBean

10:14

because it can't actually read the dollar

10:16

sign so people have adapted now to

10:18

say that word out loud so it's

10:20

first typed up be hidden from the

10:22

algorithm then said to be hidden from

10:24

the algorithm and then adopted and codified

10:26

into modern language. Which is

10:28

remarkably creative and clever right like there's

10:30

something really quite

10:32

charming about it I know you mean about the

10:35

sinister aspect of the machine but there is an

10:37

aspect of it that's really quite charming. I

10:40

do think so I this is where

10:42

I think we're destined to think of social

10:44

media in a cynical way we're really in

10:46

this space where we really understand that it's

10:48

not in our benefit we don't have really

10:50

control over the systems we don't have

10:52

a way of as we just saw

10:54

with like Elon Musk we don't really have a say

10:56

in how these deals get done we have no idea

10:58

how these up systems operate on the other hand the

11:01

systems are made up of people all of

11:03

this is user generated content it is people

11:06

connecting with one another and so to

11:08

go against the cynicism that overt is to

11:11

say you know this is a way of

11:13

making community in a very interesting way and

11:15

it's endearing and earnest to me to

11:17

realize that people want to talk about things and

11:19

it doesn't matter what the machine says they're going

11:21

to talk about them anyway. We

11:35

speak with memes archives. My

11:37

name is Kenyatta cheese I am co-creator

11:39

of the internet meme database know your

11:42

meme. There's this mistake that

11:44

a lot of folks who come from a more

11:46

traditional media space make that media and social media

11:48

are same thing like egg and eggplant are right

11:50

they have the same exact word but completely different

11:52

things. One is content the other

11:55

one's conversation and the reason

11:57

why we want to be able to make.

12:00

those things malleable. The reason why we want to

12:02

put our own mark on that meme and share

12:04

our own versus and experiences because we're using it

12:06

for conversation. My

12:12

name is Inma Sichman and the name

12:14

of the book is Memes in Digital Culture.

12:16

The main concept is

12:18

a digital concept. It was coined

12:21

in 1976 by Richard Duncan and

12:23

it describes small units of culture

12:26

that spread from person to

12:28

person by copying or imitation.

12:30

Academics argued about this concept

12:32

for ages but then the

12:34

internet came. Memes are

12:37

the building blocks of contemporary

12:39

digital culture and they

12:41

express deep fears,

12:45

motivations and social sisters

12:48

and you cannot ignore them just because

12:50

seemingly they see. CBC

13:06

Radio. I'm Nora

13:08

Young and today we're talking about how digital

13:11

technology is changing the way we talk online.

13:14

Right now my guest is Jamie Cohen,

13:16

a digital culture expert and assistant professor

13:18

at CUNY Queens College. So

13:20

far we've heard how ALGO speak

13:22

helps users evade censorship on platforms

13:24

like TikTok by swapping out banned

13:26

words for coded words that imply

13:29

the meaning provided you're in the

13:31

know and we've talked about

13:33

how that can lead to accessibility barriers. But

13:36

beyond ALGO speak are there other

13:38

examples of how technology has shaped

13:40

and changed our communication patterns. Yeah,

13:43

memes. So visual culture communicating graphically is

13:45

probably our biggest step towards a change

13:48

in language. And so around the early

13:50

2010s people moved out of the space

13:52

of low cats where the cat was

13:54

saying the word and then the doge

13:57

meme where the dog was saying the

13:59

words. to referential media that

14:01

changed our language. So we've

14:03

been speaking memetically, at least

14:05

culturally, memetically, digitally. Memes

14:08

go all the way back to ancient times,

14:10

but digital memes or digital internet language is

14:13

really from about 2014 to

14:15

present, where we've been encoding

14:17

our language in ways that

14:19

hide from both people, censors,

14:22

and outside communities. Can

14:24

you expand on that a little bit? What

14:26

is it about the nature of memes? The

14:31

nature of memes is actually similar to the

14:33

nature of emojis. When we want to share

14:36

an emotion that's unusable in text, like sarcasm,

14:38

it's impossible to say sarcasm in a text space

14:40

because people might read it wrong. Again, back to

14:42

nuance. When you use

14:45

an emoji, you could fill in

14:47

the gaps of your emotions. It's a graphical

14:49

way of communicating, and the user usually understands

14:51

it. Emojis are fairly easy because they're representative

14:53

of faces. There's a meme

14:55

is a replacement for the

14:58

emotion. Many examples of

15:00

memes will be sort of like algo speak.

15:02

You take a word that replaces another word

15:04

and you expect the people to get it.

15:06

I'll give you a very interesting coded example.

15:09

For a while, there was this meme that was

15:11

about misspelling how to talk about food. Somebody

15:14

would say, I'm going to make gourmet

15:16

food, but it was spelled G-A-R-E-M-A-Y. And

15:19

then it said, bon app the teeth, instead of bon

15:21

app the teeth. And

15:23

then eventually people said, bon atrophy.

15:26

It sounded like bon app the teeth. And then

15:28

eventually it said, I'm going to make gourmet food

15:30

osteoporosis. So

15:33

osteoporosis became the stand in for bon atrophy, which was

15:36

a stand in for bon app the teeth. And

15:38

so it was this leveling referential memetic language

15:40

that could only be understood if and

15:43

only if you've seen the origin memes from

15:46

this. In other words, that's a four layer

15:48

referential removal from its original meaning. And

15:50

if you think about algo speak in Twitter, we're on

15:52

our way to that. The other thing that I've

15:54

heard about in this context is leet speak. Can you tell me

15:57

a bit about leet speak? Yeah. LeetSpeak

16:00

was the original version of hiding, and

16:02

that was more word replacement. It was

16:04

early internet users that wanted to hide

16:06

from content moderators, who were humans. And

16:09

the way that humans would do it

16:11

at scale is they would highlight bulk

16:13

code and delete curse words and so

16:16

forth. But if you replace letters with

16:18

numbers, like Leet, EliteSpeak, so EliteSpeak would

16:20

be LeetSpeak, L33T, SP43K, so

16:22

it's numbers that look like letters. And

16:25

now when you highlight that code, it

16:27

just doesn't show up. It's just simply

16:29

invisible. But left to right reading

16:32

still allowed you to read the language. You

16:34

could put it together. It's kind of like that thing on the

16:36

internet when people do that joke where you just put the first

16:38

letter and the last letter of any word, you could still read

16:40

the word in contact. LeetSpeak did that.

16:43

That wasn't the way that memetic language

16:45

or algo speak works, which is complete

16:47

word replacement, like words that mean something

16:49

completely different, like being made on alive.

16:52

Like those are two words for one

16:54

word, but that's a word replacement. It's

16:56

completely removed from its original syntactical use.

16:59

You know, this kind of automated moderation

17:02

may be intended to block harmful language

17:04

or harmful content, but it

17:06

can also suppress, you know, really

17:08

important conversations around things like sexism or

17:11

racism or assault. Do you

17:13

think that these algorithms could eventually have the

17:15

nuance of a human moderator for

17:18

images or for text where it's able to discern,

17:20

okay, this is a valid conversation, this is not

17:22

a valid conversation? Wow, that is one of the

17:25

best questions ever. This

17:27

is the discussion I consistently have in classrooms about

17:29

content moderation with Gen Z. And

17:31

what we talk about is bad faith users and

17:33

good faith users. And can a machine understand the

17:36

difference between good faith and bad faith? And

17:38

in good faith, it means you're literally willing to have

17:40

a discussion. And in bad faith, it means you're trying

17:42

to troll somebody, you're making somebody feel bad. Can

17:45

a machine detect that type of nuance? That

17:47

is something that I don't know yet because

17:51

these machines aren't designed for good conversations

17:53

they're designed for advertising. They're designed to

17:55

sell us product. So I think

17:57

it would require in many ways us to go back in

17:59

time. and rewrite the original code. I

18:01

mean, to be honest, the biggest problem

18:04

we have with many social networks, specifically

18:06

Facebook, is the original algorithm was designed

18:08

to rank women. So that's still in

18:10

Facebook, that ranking software, that sexist ranking

18:12

software is still running behind the scenes.

18:15

So it does curb good conversations. And

18:17

also just not to get too dark

18:20

here, but it does enable bad actors

18:22

to use coded language to enable trolling,

18:24

which we call dog whistle, dog whistle

18:26

politics, which is encoding bad words or

18:29

coding bad behavior. So that's

18:31

accidentally shared by mainstream media or shared

18:33

by public users. So many

18:35

people are using algo speak to do

18:37

this, but the machines themselves can't detect

18:39

good faith versus bad faith. And I

18:42

hope someday we figure out

18:44

that part, if we're going to keep going with

18:47

machine learning, we should try to make sure that's

18:49

part of it. Yeah. Do

18:51

you think this kind of coded language is always

18:53

going to have a place? Like, for

18:55

example, in the context of a repressive

18:57

government that that's censoring political dissent? Yeah.

19:00

Unfortunately, in a right

19:02

word or authoritarian leaning future, we're

19:04

moving towards a place where we're going to have to

19:06

be a bit more subversive or covert in our language.

19:08

And we actually see the meme, the reason I bring

19:10

up memes is in China,

19:13

Tiananmen Square memes, or tank man

19:15

memes are invisible, the machine will actually detect those

19:17

and delete them. And people have figured out how

19:19

to how to get those types of information back

19:21

into the public by creating memes to go around

19:23

it. And this is why in the end, I

19:26

think algo speak is actually somewhat

19:28

short term in terms of language structure, I actually

19:30

think we're going to be speaking much more graphically.

19:33

Just generally. Just generally, we're going

19:35

to be sharing images far more than

19:38

language as a subversive or covert technique,

19:40

or even as to put it earnestly,

19:42

as a really cute way of organizing

19:44

communities. Of course, you

19:47

know, on the darker side, there are online

19:49

communities that are using euphemism to promote things

19:51

like hate speech or incite violence. Do

19:53

you think that algo speak makes

19:56

it harder to track radicalized

19:58

online content? 100%.

20:01

I'm actually, I've been doing talks and conversations with

20:03

the Trust and Safety Collective quite often. And

20:05

our big conversation is exactly what you

20:07

asked, which is how do you detect

20:09

radicalization or radicalization and process or even

20:11

grooming in terms of how

20:14

language is operated because young people speak

20:16

at a different language comprehension than adults.

20:18

And the ability to manipulate young people and

20:21

their trust and safety of the platform is

20:24

also used covertly by bad actors in

20:26

that sense. So trust and safety and

20:29

content moderators are when we use the internet, they

20:31

are the internet. And I think we have to put a lot of

20:33

understanding that a lot of social media is just

20:36

about how we moderate that content. And I know

20:38

a lot of people think, oh, they have a

20:40

too much of a heavy hand and this and

20:42

that it's humans get we can hurt we can

20:45

get in pain and we could be end up

20:47

in places that cause trauma by using these social

20:49

media platforms. We learned this in the Facebook papers.

20:52

Content moderators are designed to keep us safe,

20:54

not to censor us. This

20:56

is sort of a philosophical question. But do

20:58

you think there are downsides to letting

21:01

the technology influence how we as

21:03

humans modify language? Yeah,

21:06

there's there's a definite downside. And this is it's

21:08

philosophical. So the answer is going to be philosophical,

21:10

which is if we're trained to allow machines to

21:12

teach us to make new language, then

21:15

we're also being trained to allow other systems to

21:17

do that as well. So by

21:19

accident, we're being trained in forms or

21:21

functions that are outside of social media

21:23

that we're not aware of maybe at

21:26

this moment, it may only take one

21:28

really charismatic authoritarian leader to run a

21:30

similar system in real life to change

21:32

our language. It's interesting, you know, we're

21:34

kind of trapped in this sort

21:36

of cat and mouse game where you know,

21:38

the words

21:40

that we put out there are used to train the

21:42

systems, but then the systems are also training us at

21:45

the same time. Yes, correct. Yeah. So do you think

21:47

overall tech is helping or hurting our capacity to create

21:49

that sense of shared meaning? I'd have to because we

21:51

can't go back in time and turn it off, I

21:53

have to say it helps. I think one of the

21:55

things we have to keep in mind is there's a

21:57

responsible way of using tech or using tech as a

21:59

way of using technology. But tech

22:01

is not going away, so I believe it's

22:03

an overall benefit to humans. I think the

22:05

opposite at this point would be

22:07

far worse than anything else because we've already

22:09

engaged with that. We couldn't just turn it

22:11

off or unfaze it. So I do believe

22:14

it's a net benefit to human expression, to

22:16

human connectivity. I just think that

22:18

we have to – I think the machines

22:20

are designing a space where we're thinking about

22:22

tiny things. We're thinking about very small moments

22:24

where we're taking small scandals and making the

22:26

daily issue, when we should really take a

22:28

step back and think about the larger aspects

22:30

of it, like anything from climate change

22:32

to social justice to the idea of

22:35

just inequality in general, just to have

22:37

that in our mind. We don't have to engage with

22:39

it, but rather than thinking about the daily – what

22:41

they call the main character of the day, we

22:44

could start thinking about the issues at hand that actually

22:46

make these systems work. Jamie, thanks so much for your

22:48

insights on this. Thank you so much for having me.

22:51

Jamie Cohen is an assistant professor at

22:53

CUNY Queens College in the Department of

22:55

Media and

23:26

Archives. I'm Tim Wu,

23:28

professor at Columbia University. You

23:31

famously coined the term net neutrality. How does

23:33

net neutrality fit into these questions about the

23:35

Internet? The most important thing about net neutrality

23:38

is it is a tool for challenging monopolies.

23:40

It suggests that a startup can always go

23:42

to the Internet, try to get started, and

23:44

challenge whoever is in power. You

23:47

know, if you think about somebody like William

23:49

Gibson, Hello, I'm William Gibson. He doesn't predict

23:51

the web, but he coins the term cyberspace.

23:54

Myself, I think of cyberspace

23:56

as very much a

23:58

heritage term. But

24:01

I also, you know, by the same

24:03

token, I think of the real world

24:06

as very much a heritage term. What

24:09

we thought of as cyberspace colonized

24:12

and then effectively became,

24:14

has become what we

24:16

think of as the

24:18

real world. My

24:21

name is Baratunde Thurston and all around

24:23

strange but awesome guy. Basically,

24:25

I got really frustrated with all these different ways

24:27

like, get at me over here and be over

24:30

there and do this and all these different buttons.

24:32

I thought, why don't I just make up my

24:34

own term that encapsulates it all? Friend,

24:37

fan, subscribe, and follow. You

24:39

add all that up and

24:41

you get friend-scralo. I

24:55

own you. Though it was

24:57

originally an internet typo, home is

24:59

now common video game components. Game

25:03

D'ar. I'm Celeste

25:05

McQuarver and that's my word. I coined the

25:07

term Game D'ar. Well, Game

25:10

D'ar is kind of like gaydar. You know,

25:12

I take the gamer to know a gamer.

25:14

It's really cool in fact because I can

25:16

just be in a room with a crowd of

25:18

people and I can just instinctively tell, you know,

25:20

who plays Tetris and Pac-Man. Can only gamers have

25:22

Game D'ar? I mean, I'm just thinking that even

25:24

some stray people have gaydar. If

25:27

they're exposed to it, I suppose they

25:29

can pick up Game D'ar. Hello,

25:43

I'm Jess Milton. For 15 years,

25:45

I produced the Vinyl Cafe with the late,

25:48

great Stuart McLean. Every week, more than

25:50

2 million people tuned in to hear funny,

25:52

fictional, feel-good stories about Dave and his family.

25:54

We're excited to welcome you back to the

25:56

warm and welcoming world of the Vinyl Cafe

25:58

with our new poll. podcast backstage at

26:01

the final cafe. Each week

26:03

we'll share two hilarious stories by Stewart and for

26:05

the first time ever I'll tell you what it

26:07

was like behind the scenes. Subscribe

26:09

for free wherever you get your podcasts.

26:13

I'm Nora Young and this is an episode of Spark that

26:15

first aired in May 2022 all about how

26:18

the technologies we use are reshaping

26:20

language. Think

26:22

about how much we communicate with others

26:24

on an average day. We gestured a

26:26

merging car, we nod to a stranger,

26:28

we talked to a co-worker, a message

26:31

a friend, replied to a comment online.

26:33

We're constantly connecting. But

26:36

how did we get here? How did humans

26:38

develop shared meaning and language to begin

26:40

with? Without a time machine

26:43

we don't really know scientifically, however, we

26:46

do have some ideas about how it might

26:48

have started. Hi, I'm

26:50

Morten Christiansen, I'm a psychologist at

26:52

Konell University working on the cognitive

26:54

science and language. I'm also at

26:57

Orwisch University in Denmark and

26:59

a senior scientist at the Haskins Labs

27:01

in Connecticut. Morten is also

27:03

the co-author of the book The Language

27:06

Game, how improvisation created language and

27:08

changed the world. We

27:12

start out a book with a

27:14

little historical minet about a meeting

27:16

between Captain Cook and his

27:18

crew of the HMS Endeavor and

27:21

a band of Hausch indigenous

27:23

people in the Bay of

27:25

Good Success on Sierra Del

27:27

Fuego in January 1769. Soon

27:29

after dropping anchor, Cook and his

27:33

men, they went ashore and they were

27:35

soon met by the Hausch. At

27:38

first the Hausch sort of

27:40

retreated, but then two

27:42

of Cook's men went forward on their

27:44

own and then the same happened

27:46

with two of the Hausch people. Very interestingly

27:49

what they did is that they helped

27:51

out in front of them sticks and

27:54

then showed them and then threw them aside

27:57

and Cook and his men took that as an indication.

28:00

that they were friendly. And indeed,

28:02

that was true. But of course,

28:04

they had no common language and

28:06

inhabited utterly different worlds. But

28:08

nonetheless, and this is crucial, they were

28:11

able to communicate through what we think

28:13

of as a high-stakes game of cross-cultural

28:15

charades. And

28:19

so what we are suggesting in our

28:21

book is that this historical meaning illustrates

28:23

how language might have emerged through charades

28:25

like interactions between early humans. And

28:28

so from that perspective, language is like a game

28:30

of charades, where what we are trying to do is

28:33

to improvise, to provide clues to each

28:35

other to get our ideas across what

28:37

we want to say. And this means

28:39

that language is first and foremost a

28:41

product of cultural evolution, rather

28:44

than being built in sort of a language instinct

28:46

or something like that. Because

28:48

I think we tend to have this idea that, well,

28:51

you know, language is based on a

28:53

set of rules. We know dictionary

28:55

definitions, we know grammatical rules, and

28:57

we just implement them. But what

28:59

is your view suggesting about how

29:01

we use language alternatively? Well,

29:03

the way we're looking at language

29:06

is sort of fundamentally improvised way

29:08

of communicating. And with contrast

29:10

with this sort of notion that it's relying on

29:12

rules, or like a fixed code that allow us

29:14

to sort of bottle up our thoughts into a

29:16

stream of words, and then these

29:18

are decoded and uncalled by

29:21

the listener. But instead, what we

29:23

are suggesting is that language is

29:25

just like when we place charades, we

29:27

communicate as best as we can using

29:29

hints and clues, which are created and

29:32

presented through the powers of

29:35

human ingenuity. And so

29:37

when it comes to rules, we are suggesting

29:39

that rules are really something that emerge gradually

29:41

as various kinds of patterns of how we

29:43

use language get overlaid on top of each

29:45

other, rather than being there sort of a

29:47

priori, as it were. Your

29:50

book really stresses the important role that spontaneity plays

29:52

in language creation. Can you tell me a bit

29:54

more about what makes that so essential? Well,

29:58

so what we're trying to do... we are

30:00

communicating is really just trying to be

30:02

in the moment and trying to sort of

30:04

indicate what it is we are trying to

30:06

get across in the same way when we

30:08

are playing charades. So in a game of

30:10

charades, what you're trying to do is that

30:13

you're looking at the people you're playing with

30:15

and you're trying to get say the title

30:17

of a movie or something else across and

30:19

so what you're doing is you're paying attention

30:21

to what your audience is understanding what they

30:23

are perceiving in what you're trying to do

30:25

and you sort of adapt what you're trying

30:27

to sort of gesture to

30:29

your audience and this is what

30:31

we are suggesting is exactly what we're going

30:33

on in language. We are improvising in all

30:35

sort of clever ways in order to try

30:37

to get our point across and in doing

30:40

that we are relying not only on the

30:42

gestures themselves but also what we know about

30:44

our audience, what we know about the world

30:46

and sort of what we can take for

30:48

granted and what we sort of think maybe

30:50

our audience doesn't know and then we try

30:52

to put all that together spontaneously to collaborate

30:54

in order to understand one another. Yeah

30:57

and building off this idea of charades, it really is

30:59

this collaborative process where the role of

31:01

the listener is much more important than

31:03

just being a sort of passive recipient

31:05

of information, right? Exactly.

31:08

There's a tendency to sort of treat

31:10

language understanding as if we are sort

31:12

of kind of like a computer where

31:15

we're sort of waiting until somebody has

31:17

said something and then we sort of

31:19

spring interaction, decoding it and so on.

31:21

But our suggestion is that just like

31:24

in charades, we're sort of actually engaging

31:26

with the person talking in order trying

31:28

to figure out and collaborate to generate

31:30

a sort of common understanding of whatever

31:32

the topic is. But

31:35

it can't be all just sort of linguistic chaos, right?

31:37

There must be some sort of order

31:39

as well. Well, so

31:42

here we are referring to what's

31:44

in the science of complex systems

31:46

is called self-organization. So

31:49

what we are suggesting is that over time,

31:51

as we are playing the same game of

31:53

charades over and over again, we might reuse

31:55

certain gestures that we used before

31:57

and they might become stylized over time.

32:00

time and what we are suggesting is

32:02

that something similar is happening in

32:04

language as well and there's a

32:06

whole sort of subfield of linguistics

32:08

that's called grammaticalization that's all about

32:10

how that actually happens in language

32:12

also. So what

32:14

happens is that patterns are not there beforehand

32:16

but they build up over time as we

32:18

sort of communicate with each other over and

32:21

over again. And the kinds of

32:23

patterns vary from linguistic group to

32:25

linguistic group. Exactly. So

32:28

when we look across the world it's about 7,000 languages. We

32:31

see an amazing diversity in ways of

32:33

expressing ourselves. So there are some languages

32:35

like first the First

32:37

Nation language in Canada called straight

32:40

Salish where it seems

32:42

that they don't have a distinction between nouns and

32:44

verbs. They have a different

32:46

way of dividing up how they communicate. Of course

32:48

we also have sign languages that don't even use

32:51

spoken words at all yet. They're able to

32:53

communicate just as well as we are in

32:55

spoken language. So the

32:57

variety of human languages is

33:00

just amazing and it actually

33:02

there's a major contrast between

33:04

human languages and human communication

33:06

systems and animal communication systems.

33:08

So when you look across

33:10

the sort of the animal kingdom

33:12

there's an amazing variety

33:14

in ways that animal

33:17

communicate. So there are bacteria that uses

33:19

chemical sensing to communicate. There are the

33:21

cuttlefish that uses amazing visual displays and

33:23

you have a beast doing the whackling

33:26

that's behind to indicate where a nectar

33:28

might be found and the quality of

33:30

it and you have monkeys that use

33:32

different kinds of calls to indicate whether

33:35

there's this kind of predator or that

33:37

kind of predator. But

33:39

when you look within a

33:41

particular species of animal you

33:43

find that there's very little

33:46

variation in how these animals communicate

33:48

across individuals. But when

33:50

you look across human languages we just

33:52

see this amazing and astonishing variety of

33:54

ways in which we use different

33:57

ways of putting words together or

33:59

different ways of using signs in order

34:01

to communicate. And this is

34:03

really a major distinction between human language

34:05

and other animal communication systems. And this

34:07

is what gives us the flexibility to

34:09

express ourselves no matter what kind of

34:11

culture we live in and what kind

34:13

of environment we live in. So

34:16

when it comes to this kind of spontaneity and

34:18

importance of context in understanding language,

34:20

could you give me an example

34:22

of that in English, of how

34:24

it's contextual and how it varies? So

34:27

consider the word break. A breakdown

34:29

is terrible. So if you're breaking down with

34:32

your car, that's obviously terrible. A breakup is

34:34

also not fun for anyone. But

34:36

a breakthrough is excellent. So if we're having

34:38

a break, we are very excited. And of

34:40

course, more generally, breaking things is considered to

34:42

be bad. But if we're in a sort

34:44

of a long meeting, a long day meeting,

34:47

maybe a day-long meeting, having a break is

34:49

excellent, especially if we get coffee, of course.

34:51

And so here you have different contexts, sort

34:53

of changes what we mean by the same

34:55

word. In this case, you're a break. I

35:18

mean, there has been speculation that the

35:20

advent of the internet has meant that

35:22

English is a much more dominant force and

35:24

that more niche languages are going to fade away.

35:26

Do you have thoughts on that? Well,

35:29

there certainly have been some pressure towards

35:32

English becomes sort of a lingua franca

35:36

in some cases. But on the other hand,

35:38

also, the internet has allowed a small group

35:40

of individuals to sort of band together and

35:43

sort of maintain their languages that may not

35:45

be a sort of majority language. So I

35:47

think it goes in both directions. And of

35:49

course, we as a society

35:51

can be supporting, say, First Nation

35:53

languages or other languages that might

35:55

otherwise be in danger of

35:58

dying out. And I think there are some movements in

36:00

that direction. But certainly it is

36:02

the case that across the world there

36:04

are many languages that are near dying

36:07

out and thus we

36:09

need to both document them but also

36:12

support the people who are speaking those

36:14

languages. Because once those languages disappeared, it's

36:16

also a part of that culture that

36:18

disappears as well. So the language is

36:21

not just a language, but it's also part of a

36:23

culture and it carries with it much

36:26

information about that culture and how people

36:28

interact with one another. And that's of

36:30

course an incredible and terrible loss

36:33

once each language that has

36:36

existed dies out. From

36:47

this park, Language Preservation Archives.

36:50

Who am I without my language and

36:54

where do I come from without my language?

36:56

It is who we are as a Houssaint'n'ch people.

36:59

It ties into

37:01

our laws and our beliefs and teachings. Asa

37:04

Renee Sampson, just a lattice, an'akkou saint'n'ch.

37:08

Hul nuk sant. Ies en

37:10

kunas asla. Che'is en, dana

37:12

squeal sin chasen. Huluk sin

37:14

kunas nah asla hila. I

37:17

tat anuk sant dana kunay ex

37:19

sasqueal. And I'm First

37:21

Nations and I'm really happy to

37:23

be here. And I'm starting to

37:25

learn my sin chasen talk

37:28

and that I'm starting to learn my

37:30

teachings and my language. That's what I

37:32

said. There's a

37:34

lot of push for

37:37

language revitalization in our community.

37:40

There's word lists on there that you can go that has

37:43

different multimedia, has a picture and it

37:45

has the word

37:47

and sin chasen. Many

37:49

different other First Nations people are at

37:51

a different level. Some of them have

37:53

a lot of audio clips and

37:56

they have songs. So you go

37:58

on there and you can basically have

38:00

access anywhere in the world

38:03

to your own language. Learning

38:06

my language is learning who I am and

38:09

it's important for me to learn my

38:11

language for my children's sake because they

38:13

will grow up and they will

38:16

know who they are and who they're connected to. My

38:23

name is Sqachaltin or

38:25

Khalsilim. Those are my ancestral

38:27

names that I carry from my ancestors.

38:31

My friends that speak the language, we text to each other

38:33

in the language. And then

38:35

the funny thing like English does is that

38:37

people started abbreviating things and finding shortcuts to

38:39

say Squamish words. So just like LOL

38:42

and things like that how English would do those

38:44

things. People started doing that with Squamish

38:47

but they were like completely Squamish concepts

38:49

and Squamish words that they were shortening

38:51

and communicating instead of actually writing out the whole word.

38:54

My goal isn't to get to a place in

38:56

the future where young people can go to downtown

38:59

Vancouver and order a hot dog in

39:01

Squamish. You can do that

39:04

in English. We're always going to have English. We're never going to

39:06

really escape from English. The text messaging

39:08

stuff though and being able to communicate to

39:10

these new ways, it's a way that I

39:12

think for myself and for the young people

39:14

is that we're trying to find avenues to

39:16

reclaim our identity, reclaim our language, reclaim our

39:18

little place in this world that gives us

39:20

a sense of pride and strength and

39:23

encouragement. You're

39:41

listening to Spark from CBC Radio.

39:44

I'm Nora Young and today we're talking about

39:46

how technology is driving changes in the way

39:49

we communicate. Right now, my

39:51

guest is Morten Christensen, a cognitive scientist

39:53

and co-author of the book The Language

39:55

Game. So far we've

39:57

heard about how spontaneity and collaboration are

39:59

foundational. pillars of human language. But

40:02

what did the introduction of digital communication mean

40:04

for the evolution of language?

40:07

In general, across human history, of course,

40:10

we have come up with many different

40:12

kinds of technologies, and that has influenced

40:14

our way of communicating with one another

40:16

in a variety of ways.

40:19

Just to first consider a different

40:21

technology, namely writing systems. So writing

40:23

systems allows us to more easily transmit

40:25

knowledge from one generation to another. And

40:27

that, of course, is a major advantage.

40:29

But it also introduces a certain kind

40:31

of conservatism in how we interact with

40:33

one another, if we sort of adhere

40:35

too slavishly to what was written before,

40:38

rather than going with the flow of how we talk

40:40

now. And oftentimes, when new

40:42

technologies have been introduced, they

40:44

have been decried, or the impact on them

40:47

have been decried by especially previous generations. So

40:49

for example, going back to

40:51

writing, when the Gutenberg printing press was invented,

40:54

there was actually quite a lot of concern

40:56

about this spread of literacy, because they were

40:58

concerned about what it might do to the

41:00

mind of the general population. But of course,

41:02

today, we're sort of continuously emphasizing

41:05

kids should be reading, we should all be reading more,

41:07

and so on. And of course, when we

41:09

look at sort of modern technologies, when texting

41:11

became popular, especially amongst young people, there's a

41:13

lot of consternation about what they would do

41:16

to their language skills. And some

41:18

of these language maiments, they were sort

41:20

of very concerned that they were just

41:22

completely, that language skill would degenerate, and

41:24

so on, because they used new abbreviations

41:26

like LOL or JK, and

41:28

so on. But of course, the actual

41:30

advent of texting is actually a nice

41:33

example of what we think of as

41:35

cultural evolution of language in action. Cultural

41:37

evolution of language is how we think

41:39

language has evolved. And cultural evolution is

41:42

constrained by whatever medium we're using for

41:44

communication. So for example, spoken language is

41:46

limited by how we can move our

41:49

mouth around, our memory for

41:51

sound sequences, and so on and so forth.

41:53

But the same is true when it comes

41:55

to texting. So initially, when we started texting,

41:57

we'd want to know that we had these sort of tiny cell phones.

42:00

I had nine buttons with typically

42:02

each button. There was three letters

42:04

going to that. So it's really

42:06

cumbersome to type out long messages.

42:08

So what happened is that these abbreviations

42:11

and other ways of doing

42:13

shorthand, these were really

42:15

adaptations by the texting language, as it

42:17

were, to those constraints in

42:19

order to make it easier to

42:21

communicate. But of course, once we

42:23

got our modern day smartphones with

42:25

their virtual keyboards and predicted completion,

42:28

we sort of gone back to

42:30

spelling out words fully again. And

42:32

so the texting language sort of kind of reverted

42:34

to a more standard written format. Now, of course,

42:37

with the kind of added flourishes of things like

42:39

emojis and all sort of stuff. So language sort

42:41

of adapts to the technology. And I think that's

42:43

sort of a nice example of at least a

42:46

positive one of how it has happened in this

42:48

case here. Yeah, and certainly indicates a lot

42:50

of the kind of playfulness that you talk

42:52

about in your book as well. You

42:54

say in your book that successful communication relies on

42:56

this kind of shared knowledge and a shared context.

42:59

But I'm wondering if online

43:01

culture makes that more difficult if we

43:03

just think about how quickly cultural touchstones

43:05

move in and out of fashion. So

43:08

I think on the one hand, having

43:10

the internet has sort of provided

43:12

us with more input from all places

43:14

in the world. So that's a good

43:16

thing. So another thing can

43:18

create a sort of more common understanding

43:21

across different cultures to some degree. But of course,

43:23

it's also possible that you create these bubbles that

43:25

we sort of tend to stick within and that

43:27

could become sort of quite narrow minded within those

43:30

bottles. So I think there's a lot of different

43:32

variations and exactly how it's gonna play out in

43:34

the long term I think might be a little

43:36

hard to. There's a lot of different variations and

43:38

exactly how it's gonna play out in the long

43:40

term I think might be a little hard to

43:43

know at this point. But at least the

43:45

way we are looking at language as

43:47

a fundamentally collaborative endeavors does

43:49

suggest that in the long term, language is

43:51

about a collaboration. And if we take that

43:53

too hard, then maybe that can also help

43:56

us deal with things like find ways around,

43:59

no fake. and all these other problems

44:01

that are sort of plaguing us at

44:03

the moment. And I think that's sort

44:05

of an illustrative example of regard

44:07

to technology from a few years ago. So

44:10

you probably remember that at

44:12

some point, there was all these spam

44:14

email. And at that time,

44:16

those talk about at some point in the

44:19

future, email would be pretty much impossible because

44:21

we would just be inundated with spam. But

44:24

spam filters have become incredibly good. So now

44:26

we still get some spam, but not as

44:28

much as we used to. So hopefully, there

44:30

might be some solution in the future to dealing with

44:32

things like fake news and so on going

44:35

forward as well. At

44:37

one point in the book, you write that, quote,

44:39

words do not have stable meanings. They're tools

44:42

used in the moment, as we've been discussing. Overall,

44:45

in this episode, we've been looking at how users

44:47

on social media platforms have started

44:49

using codified words or algo speak

44:51

to keep algorithms from detecting and

44:54

censoring their content. For example, using

44:56

a phrase like leg booty instead

44:58

of LGBTQ. Is that

45:00

the kind of instability of meaning that you're talking about,

45:02

or does that fit into how you see the instability

45:04

of meanings? Well,

45:06

I think it's been said that the sort

45:08

of meaning of words are in constant flux.

45:10

And as we as a

45:13

society or as a culture evolve,

45:16

the words meaning will evolve with us in

45:18

that way. So I think that's a nice

45:21

actually example of how meaning is

45:23

never stable, but continues to change across time.

45:25

Again, sort of trying to fit into the

45:27

constraints of the moment. In

45:30

this case here, getting around being censored

45:32

by some of these bugs and so on.

45:35

Mm-hmm. Further

45:42

start. Speech police. Archives.

45:45

Mignon Fogarty. Grammar Girl. There's

45:48

just not an excuse for not using a capital

45:50

letter at the beginning of a sentence or not

45:52

capitalizing someone's name when you're writing a

45:54

professional email message and you're typing on

45:57

a regular keyboard. It

45:59

is kind of real. And when you're doing

46:01

the electronic equivalent of passing a note

46:03

to a friend, it's different. You know,

46:05

I think that when you don't capitalize

46:08

someone's name in a work email, you're essentially

46:11

saying, I don't care about you enough to

46:13

hit the shift key. Here

46:18

at Spark, we've had a number of conversations

46:20

and arguments about the best way to end

46:23

an email. Yes, arguments.

46:25

So today, my tech PSA is how

46:28

to sign off without sounding weird, or

46:30

forced, or British. Michelle! What,

46:33

Nora? I'm just calling

46:35

it as I see it. And as I

46:37

see it, when you sign off with cheers,

46:39

or best, you sound like you're from across

46:42

the pond, or clinking wine glasses at a

46:44

dinner party. Come on. Nora also uses cheers.

46:46

In fact, so many people do it that

46:49

I'm starting to wonder if it's a colonial

46:51

tick we can't shake. So

46:53

I think your safe is to end an

46:55

email with six. Yeah, just

46:57

thanks. It's simple, to the

47:00

point, polite and friendly. You

47:08

spent 30 years learning the difference between it's

47:10

and it's, and then you end up fighting

47:12

with the phone over it. Iver

47:15

Tossen. I'm a senior product manager at Buzzfeed,

47:17

and I hate typing on my phone. I

47:20

hate fighting with auto correct. I hate the

47:22

typos. I hate the whole experience. And it

47:24

surprises me that this is not

47:26

an area where there's not more active

47:28

innovation. The extent to which technology and

47:31

predictive software is surging ahead in some

47:33

levels seems out of

47:35

sync with this extremely pedestrian

47:38

fight that everybody is fighting

47:40

on a daily basis. Especially,

47:42

low but high to anyone who tries to

47:44

use Canadian spelling. Yes. Where

47:46

not only do you have the cultural classes of

47:48

the United States, dragging room in one direction, but

47:50

you also have other software. We'll be

47:52

tied you if you try and use we'll be tied you and see

47:55

what they've got a lot of corrects to. Well,

47:57

very much so. It's almost the same. It's almost

47:59

the same. like the phone has an idea

48:01

of how you should be speaking. And if you

48:03

deviate from that norm, it tries to drag you

48:05

back to it. Mm-hmm. I'm

48:15

Nora Young, and today on Spark, we're exploring

48:17

the way digital tech has changed the

48:19

way we communicate, for good and bad.

48:22

Right now, we're talking about the evolution

48:24

of language with my guest, Morton Christensen,

48:26

a cognitive scientist and co-author of the

48:28

book, The Language Game. You

48:32

spent some time in the book talking about AI,

48:35

and you write that computers can't match

48:37

the complexity of human linguistics. But

48:39

with artificial intelligence, mastering chess,

48:41

writing poems in some cases, responding

48:44

to our natural language queries from us on

48:47

Alexa or Siri, why

48:49

couldn't it also learn to

48:51

truly replicate human language patterns?

48:55

Well, in part because

48:57

they don't really understand anything.

49:00

I mean, computers today in AI is

49:02

amazing. I mean, they can steer spacecrafts,

49:05

and they can play chess or go

49:07

or any kind of computer games with

49:10

amazing skills. However,

49:12

when it comes to these language models, and

49:14

they are also incredible. So the ability to

49:16

do things like Google Translate or other kinds

49:18

of language systems, they can clearly create all

49:21

sort of complex language. But they

49:23

don't really understand what they're doing. What they're

49:25

relying on is taking little bits and pieces,

49:27

putting them together in a way that

49:30

makes it seem like it's true human

49:32

language. But they're not really interacting with

49:34

one another. So it's more like they're essentially

49:37

engaging in monologue. And one of the

49:39

things that we argue is that it's

49:41

very important for us to view language

49:43

as dialogue rather than monologue. And

49:45

in a sense, these bots or

49:47

AI language system, they're really just

49:50

playing monologue. And that's a major limitations,

49:52

which means that they can't really go

49:55

beyond that. And so clearly today, at

49:57

least, we haven't told computers to place

49:59

your rates. And I think we don't

50:01

really have to worry about sort of computer-seeking

50:03

or a language as such until

50:06

we see them playing charades. Now once they

50:08

– if and when they do that, then

50:10

we might want to be worried.

50:14

Indeed. So

50:17

I mean, I suppose from a more philosophical level,

50:19

we could say that it's because AI systems

50:22

don't have subjectivity. They're not bringing that

50:24

subjectivity to the moment of communication, I

50:26

guess. No.

50:28

I mean, there are so many not doing

50:31

that. I mean, they're able to digest billions

50:33

of words in a way that none of

50:35

us could ever do. So

50:37

I mean, they're much more well-read, so to speak, than

50:39

any of us. But yet, they

50:41

don't really understand what

50:43

the text, how the text is

50:46

sort of relating to the word. Nonetheless, they

50:48

can do amazing things. So here at

50:50

Cornell University, I'm in the middle of

50:53

a project where we are using one

50:55

of these big language models to look

50:57

at poetry. And we are having them

50:59

generate poetry. And they can produce

51:01

poetry. That's actually quite interesting. So we can

51:03

ask it to do a poem

51:06

in the style of Emily Dickinson or

51:09

Shakespeare or Walt Whitman or some other

51:11

poet. And they can do that quite

51:13

well. So one of the things that we are currently

51:15

trying to figure out is that how

51:18

good are they? So we're going

51:21

to have both undergraduates trying to produce poetry

51:23

based on a prompt, say two lines of

51:25

a poem, and they have to continue it.

51:27

And we ask sort of

51:29

one of these language models to do that, too.

51:31

And then we're going to do sort of a

51:33

poetry-touring test where we ask other people to judge,

51:36

was this generated by a person or machine? And

51:40

we'll see. We don't have the answer yet. So maybe we'll

51:42

talk at some later point, and then I can tell you

51:45

what the answer is. But they can

51:47

clearly do this. But of course, they don't

51:49

have any kind of emotion or anything like

51:51

that. So they're doing that by essentially having

51:53

read loads of poetry. So they've been

51:55

trained on poetry, books of poetry, and so on.

51:57

And they can use that to generate new poems.

52:00

But yeah, they don't really understand the

52:03

underlying meaning, the emotions, the culture that

52:05

goes into writing these poems. So they're

52:07

going to be hopefully missing that. So

52:10

we are interested in how they might

52:12

be able to generate poetry. So

52:15

does that sort

52:17

of limitation on artificial intelligence mean

52:20

that things like content moderation

52:22

are always going to be a problem because

52:24

AI systems are going to

52:26

have trouble dealing with the flexibility of language. They're going

52:28

to have trouble understanding when something is sarcasm

52:30

and when it's not, for example? Probably

52:33

yes. And also, humans are quite ingenious.

52:35

So they have, you know, as soon

52:37

as we come up with one sort of algorithm

52:40

to sort of try to deal with content moderation,

52:42

humans are going to come up with a clever

52:44

way of trying to get around it. But

52:46

of course, the question is whether in the

52:48

long term the algorithm sort of can keep

52:51

ahead of human ingenuity. We don't know.

52:53

But because language is so flexible and

52:55

humans keep on coming up with new

52:57

ways of expressing themselves, it's going to

52:59

be hard to prevent anything

53:01

to everything from new ways

53:04

of doing things from occurring. Yeah. And

53:07

just finally, Martin, looking into the

53:09

future, how do you envision our styles of communication

53:11

will change? Well, we've certainly

53:14

seen throughout human history that every time sort

53:16

of new technologies have come in, it's going

53:18

to change how we communicate. So like with

53:20

reading or texting, as we talked about

53:23

earlier. So likely there will be changes in

53:25

how we communicate. And as, you know,

53:28

smartphones, computers, etc, etc, becomes more

53:30

and more integrated in our lives,

53:32

it's likely to affect how we

53:34

communicate. Exactly how that change

53:36

will come sort of through is

53:39

uncertain, at least in my mind. I'm not sure

53:41

what happened. Of course, there's all sort of science

53:43

fiction scenarios and so on. But I don't really

53:45

know. But what I think everyone

53:48

can be confident about is that it will

53:50

change in some way, that we will adapt

53:52

to the new technologies just as they also

53:54

will to some degree adapt to us. Mm

53:57

hmm. Thanks so much for your

53:59

insights on this. So thank you. Thanks for

54:01

having me. Morten Christensen is a

54:03

cognitive scientist and co-author of the book

54:06

The Language Game. You've

54:11

been listening to Spark. The show is

54:14

made by Michelle Parisi, Adam Killick, McKenna

54:16

Hadley Burke, and me Nora Young. And

54:18

by Jamie Cohen and Morten Christensen. And

54:21

from the Spark Archives, Kenyatta Cheese,

54:23

Lamour Shiffman, Kim Wu, William Gibson,

54:26

Baratunde Thurston, Celeste McCorder,

54:28

Renee Sampson, K