0:00

the right question can greatly impact your

0:02

future, especially when it comes to

0:04

your finances. So when

0:06

you're seeking advice that fits your

0:08

financial goals, remember to ask. Are

0:11

you a CFP professional? Because

0:13

certified financial planners are ethical,

0:15

educated, and experienced professionals committed

0:18

to acting in your best

0:20

interest. That's why it's gotta

0:22

be a CFP. Find

0:24

your CFP professional

0:26

at let'smakeaplan .org. Artificially

0:30

intelligent? Yes. Socially?

0:33

Awkward. From American Public

0:35

Media, this is Marketplace Tech. I'm Stephanie

0:37

Hughes. for you

0:40

next video. Bye.

0:47

I met up last week with

0:49

Layla Ishak. She's a professor of cognitive

0:51

science at Johns Hopkins University. I

0:53

went to visit her lab where she's got

0:55

some watercolors of brains on the wall. Good

0:58

brain art is important. It

1:01

takes the senior scientist on a new

1:03

study looking at how good AI is

1:05

at reading social cues. She and

1:07

her research team took short videos. Sometimes

1:09

it's like two people chatting and dancing, two

1:12

guys doing this like

1:14

synchronized skate routine. They

1:16

showed the videos to human participants

1:18

and asked them questions like, are these

1:20

people communicating with each other? They

1:22

also gave the same videos to over

1:24

350 open source AI models. That's

1:26

a lot, though it didn't include all

1:28

of the latest and greatest. And

1:31

Ishik found that the AI models were

1:33

a lot worse than humans at understanding

1:35

what was going on. One

1:37

thing we found was that actually none

1:39

of the models could do a good

1:41

job of matching behavior or brain responses

1:44

to these different social attributes, like our

1:46

people communicating. Surprisingly, none of them

1:48

could even do a great job at telling us

1:50

things like are these people facing each other? So

1:52

I think we had a

1:55

feeling that there would be elements

1:57

of this that the AI could

1:59

not capture, but we were pretty surprised

2:01

by, in general, the poor performance.

2:03

Exactly how bad it was. And

2:06

so basically the AI across the board couldn't

2:08

tell if people were communicating, if they were facing

2:10

each other. There was some variety.

2:12

So like I said, we tested 350

2:14

models. Some models were better at it

2:16

than others, which yielded some interesting insights,

2:19

but no single model could provide a

2:21

match to all the human behaviors we

2:23

tested. Why does this matter? Like,

2:25

why would it be helpful for AI

2:27

to be good at this? Yeah, well, I

2:29

think anytime you want to have AI

2:31

interacting with humans, you want to know what

2:33

those humans are doing, what they're

2:35

doing with each other, what they're about to do next.

2:37

And I think this just really highlights how far a

2:39

lot of these systems are from being able to do

2:41

that. What do your findings

2:43

mean for possible business applications for

2:45

artificial intelligence? Yeah, I

2:47

think that businesses where this is probably

2:49

most close to being applied or currently

2:52

being applied are things like self -driving

2:54

cars. People, the drivers, have this intentionality

2:56

and the pedestrians, and you have to

2:58

be able to understand that. For

3:00

example, that, like, I think it's

3:02

very hard for self -driving cars to

3:04

make an unprotected left turn. It's

3:07

hard for humans, too. It's hard for humans,

3:09

too, sometimes. And when you do that, you

3:11

have to, like, really look around and think

3:14

about who's doing what next and those sorts

3:16

of things. And I think this... highlights how

3:18

much more work needs to be done both

3:20

in the development of these systems to improve

3:22

them. But I also think it highlights some

3:24

new ways to be stress testing these systems

3:26

against humans. We'll be

3:28

right back. This

3:31

marketplace podcast is supported by Commvault.

3:33

Businesses can recover data as well

3:35

as cloud applications and platforms after

3:38

a cyber attack in just minutes

3:40

with Commvault so they can become

3:42

a continuous business. You're

3:44

listening to Marketplace Tech. I'm Stephanie

3:46

Hughes. We're back with Leila

3:48

Ishik, Professor of Cognitive Science at

3:50

Johns Hopkins University. I

3:53

think some people envision this future

3:55

where we all work alongside our

3:57

AI colleagues or buddies, you know.

4:00

And I wonder, what do your findings

4:02

mean for the short term at

4:04

least about AI's ability to do that?

4:06

Like, will it be the Michael

4:08

Scott in the office? Perhaps,

4:11

but I think there are some

4:13

even more baseline findings than baseline

4:15

problems than that. Like I said,

4:18

you wanted to be able to tell what the

4:20

person is doing, what it's close to, who's

4:22

close to who, and even those more basic things

4:24

than just reading intentions that seems to be

4:26

lacking in as well. I feel

4:28

like I'm a grown up and I'm

4:31

still learning how to pick up on social.

4:33

like a lifelong process. Do

4:35

you think the AI will get there? Yeah,

4:37

you mentioned your grown up. I mean, I

4:40

think there are, it's really striking

4:42

how much of this even little babies can do

4:44

though. And so not to

4:46

the full sophisticated level that we keep

4:48

developing and refining as throughout development

4:50

and throughout our lifetime, but there are

4:52

some base abilities that seem to

4:54

be present from at least very early

4:56

in childhood. you

4:58

know, I think AI should be able to get

5:00

there. And I think the progress AI has been

5:02

making over the last decade or so has been

5:04

really amazing. But I think that

5:06

some of these problems might require a

5:09

fundamentally different approach than sort of the

5:11

brute force, just get more data and

5:13

bigger networks solutions that have been taken

5:15

us pretty far. But I think there

5:17

might be limits to that. Another

5:19

place that AI is being used

5:22

is in customer service. I wonder

5:24

how your findings what

5:26

your findings can mean for customer service and

5:28

then AI is used in that. Yeah, I

5:30

mean so I think right now a lot

5:32

of those applications are all text -based like chatbot

5:34

type things but if you really wanted to

5:36

scale that up you would want or any

5:38

sort of assistive robots AI you would want

5:41

them to be able to interact with people

5:43

based on visual cues. We use the visual

5:45

cues all the time to interact with each

5:47

other so I think that this has important

5:49

implications. anytime you want an AI to be

5:51

interacting with humans. Do you

5:53

have any advice for AI makers?

5:56

Yeah, I mean, I think historically, and

5:59

still loosely to some extent, AI

6:01

has drawn a lot of inspiration

6:03

from humans, from cognitive science, from

6:05

neuroscience. And I think

6:07

in the latest AI boom that those

6:09

two fields have sort of diverged. But

6:11

I think this is an important point

6:13

to start coming back together and where

6:17

The things we know that humans care about

6:19

and the sort of structure we imbue on

6:21

the world can help improve these AI models. That's

6:24

Leila Ishik, professor of cognitive

6:26

science at Johns Hopkins University. We've

6:28

got a link to her study on our

6:31

website, marketplace .org, and we've got a picture of

6:33

her brain arc, too. Daniel Shin

6:35

produced this episode. I'm Stephanie Hughes, and

6:37

that's Marketplace Tech. This

6:43

is APM. If

6:46

there's one thing we know about

6:48

social media, it's that misinformation is

6:50

everywhere, especially when it comes to

6:52

personal finance. Financially inclined

6:54

for Marketplace is a podcast you

6:56

can trust to help you get serious

6:58

about your money so you can

7:00

build a life you've always dreamed of.

7:03

I'm the host, Janelia Spinal,

7:05

and each week I ask

7:07

experts important money questions, like

7:10

how to negotiate job offers,

7:12

how to choose a college that

7:14

you can afford, and how to

7:16

talk about money with friends and

7:18

family. Listen to financially inclined wherever

7:21

you get your podcasts.