0:02

Jonah Maddox: Too Long, Didn't Read. Brought to you by the Alan Turing Institute, the National

0:06

Institute for Data Science and AI.

0:10

Welcome to Too Long, Didn't Read. AI news, developments and research delivered directly to your ear

0:16

holes from the experts and me.

0:19

I'm Jonah, a content producer here at the Turing, and

0:22

Smera Jayadeva: I'm Smera, a researcher in data justice and global ethical futures.

0:26

Jonah Maddox: Smera, in season one, you were the resident expert, brilliantly

0:30

answering questions on everything AI, from the ethics of AI labor, to history of the

0:34

chip wars, and even briefly stopping to advise Santa on a potential AI workflow.

0:40

But you have been given, I'm gonna say, a promotion.

0:44

You're my co presenter now! Smera Jayadeva: Yeah yes, that's right.

0:49

We have a slightly new format this season.

0:51

This time you and I will be discussing the AI news, but we'll also be seeking various

0:56

expert voices from a wide range of AI and data science disciplines, ultimately to

1:02

create an even more comprehensive podcast.

1:04

Jonah Maddox: Exciting stuff. On this episode of TLDR, we will be talking about misinformation, not a rising

1:11

star in an academic beauty pageant, but a very serious threat for democracy.

1:15

Smera Jayadeva: We'll also see if the age of dynamic, real time robotics is actually

1:20

upon us, and what it means for your job.

1:22

Jonah Maddox: And we look beyond the headlines when talking

1:24

about generative AI and sets.

1:30

That's funky music on the guitar. 2024 is the year of elections.

1:39

Over 80 countries and half the global population will be voting this year.

1:43

Since there's no collective noun for a group of elections, let's go with a group.

1:47

Electiontastic. Smera Jayadeva: Yeah, taking on elections is a huge endeavor.

1:53

Take India, it's seen as the world's largest democracy.

1:56

Elections began in late April and will go on till June.

2:00

And we're talking a massive population of nearly a billion people.

2:04

Jonah Maddox: Yeah, and as if sorting through the candidates campaign promises

2:07

wasn't tricky enough before, we now have good old AI to consider and how

2:12

it can help people spread falsehoods.

2:15

Smera Jayadeva: Yeah, and we discussed some of this in the first episode of our first season.

2:18

Remember Jonah, way back when? But essentially we face three strands of information manipulation.

2:25

First we have disinformation, which is the big bad boy, wherein

2:29

information is falsely construed with the intent of manipulating audiences.

2:35

Then we have misinformation, but without the intention of actually causing harm.

2:40

Finally, we have the complexities of malinformation, where information

2:45

is exaggerated or conflated to obscure the truth or the narrative.

2:49

This is also where secret or classified information is often shared at a

2:54

strategic time just to influence voters.

2:56

Jonah Maddox: Check out series one, episode one for the fuller explanation

3:00

on this we'll link it in the show notes. Smera Jayadeva: And there's also a few points to keep in mind when it comes

3:05

to the misuse of data and information.

3:08

For one, any group or individual manipulating information doesn't

3:12

necessarily have to follow a single route.

3:15

Even if they're intentionally planning on manipulating information, one can

3:19

begin by exaggerating historical events and follow it up with intentionally

3:24

false and misleading information only to galvanize voters towards their cause.

3:29

For instance, you know, I mentioned India early on and actually was

3:32

in India during the elections and there's a good chance that by the

3:35

time the This recording is out.

3:38

India is probably still going to be counting the votes.

3:41

But if one were to track the misinformation or disinformation

3:44

campaigns in the country, it's, it's rarely a day without reports of false

3:49

information making rounds on social media platforms or communication channels,

3:54

be it Twitter or X or even WhatsApp.

3:57

In fact, the world economic forum said India has the highest risk of

4:01

miss and disinformation in 2024.

4:05

Jonah Maddox: So, it's rife, and it's relevant, and we're going to deal with it.

4:08

It's probably time we should, uh, bring on our special guest to navigate this.

4:14

DLDR Expert Guest! This month, we are joined by an expert who has worked as an analyst

4:19

within the Defence and Security Research Group at RAND Europe.

4:22

She's led projects which assess the impact of emerging technologies

4:26

on the information environment and worked to identify the

4:28

implifications of disinformation and conspiracy theories in Europe.

4:32

That sounds cool. Her research has informed strategy and policy at the UK Home Office, UK Ministry

4:37

of Defence, the European Commission and the United Nations Development Programme.

4:41

From the Turing Centre for Emerging Technology in Security, CETAS,

4:45

we are very happy to welcome research associate Megan Hughes.

4:48

Woo, sounded a bit like the sort of beginning to blind date there where

4:53

Graham kind of brings him on like, from the Turing Center of Emerging Technology.

4:58

Hello Megan, I'll let you speak now. Megan Hughes: Hi Megan.

5:02

Thank you so much for having me. Looking forward to hopefully an interesting discussion.

5:06

Definitely. Jonah Maddox: So can you give us a very brief explanation of what CTAS

5:10

is and what a research associate does? Megan Hughes: Sure.

5:12

Yeah. So CTAS is the Center for Emerging Technology and Security

5:16

at the Alan Turing Institute. And I'm a research associate within the team and we work on policy

5:22

research relating to emerging technology and national security.

5:26

So we look at kind of the implications of emerging tech technologies and we

5:30

try to advise actors like the government on what they should do in response.

5:34

Jonah Maddox: Okay. So, we've learned sort of quickly in the intro from Smyrna about

5:38

misdisc and malinformation. But could you tell us a bit more about how it plays

5:42

Megan Hughes: out during election time? Sure. Yeah.

5:44

So I'll kick us off with a kind of traditional influence operation.

5:48

If we look back at the 2016 U.

5:50

S. presidential election, Presidential elections. We can see quite a typical example of a state sponsored influence operation.

5:58

So this was when Russian actors looked to influence US

6:02

voters ahead of the elections. And they did a number of things.

6:06

So it wasn't just a kind of misinformation, disinformation campaign.

6:09

It was much broader than that. So we had things like hack and leak techniques.

6:13

where hackers got into the Clinton campaign emails and then shared

6:18

these emails over a period of a few weeks to kind of distract

6:21

from the main campaign messages.

6:24

But specific to misinformation, Russian actors created a network

6:28

of fake accounts of bots. We're looking at about 50, 000 of them, and they were all sharing emails.

6:35

divisive content, fake news stories, reposting hashtags to make them

6:39

go viral, like Hillary for prison.

6:42

That was one of them. And they were also publishing political advertisements, criticizing Clinton.

6:47

So that's looking a few years ago, and that's looking at something that,

6:51

like I said, I'd kind of term that a traditional influence operation.

6:54

When we look to the past. past few years, and we look at AI examples from elections that have taken

7:00

place since we've looked at the start of 2023 and research I've been, I've

7:04

been doing, and I can talk to you about three examples of AI misinformation.

7:09

So you've got AI generated voice clones.

7:12

I don't know if you've, if you saw coverage on the Biden Robocalls.

7:16

So this is where we had a deepfake audio.

7:19

clips of Joe Biden urging voters not to turn out and vote in

7:24

the New Hampshire primaries. We've also got an example, if we look to Poland in their recent election, the

7:31

opposition party actually published a deepfake audio clip of the prime minister

7:36

reading a set of real leaked clips.

7:38

So you can see how kind of generated voice content we're seeing come up.

7:42

General AI generated content as well.

7:45

Over in the U S there are reports of whole news sites that have been generated by

7:49

AI sharing completely fake news stories.

7:52

So that's more text based content that's quite easily shareable.

7:55

And lastly, coming closer to home, looking at the London mayoral elections,

8:00

we saw AI powered bots, that were again, sort of similar to the tactics in the

8:05

Russian operation, circulating hashtags.

8:07

So the hashtag London voter fraud was circulated quite a

8:11

lot ahead of the elections. So those are some examples of techniques and tactics that

8:15

we've seen being employed. Right. And is there any evidence of them having the desired effect?

8:22

So that's really interesting. So when we look at Misinformation generally.

8:28

So if we kind of take the AI out of the context, a lot of studies have shown that

8:33

only a small minority of people actually see the majority of misinformation.

8:39

So I think there was a study in 2016 on X, formerly Twitter and it showed

8:44

that only 1 percent of X users actually were exposed to 80 percent of the

8:50

fake news content on the platform. And if you're exposed to misinformation, it doesn't necessarily

8:57

mean you'll be persuaded by it. So fake news, we know, is more likely to enhance existing views.

9:03

It's not as likely to radically change your behavior, so not as likely to

9:08

kind of influence voting intentions. And studies have quite consistently found that in relation to elections,

9:14

misinformation hasn't meaningfully The outcomes of elections.

9:18

And that's because there are loads of factors that contribute

9:21

to someone's voting choices. What we can look at is what's new with AI.

9:26

So looking forward to kind of upcoming elections.

9:29

Well, AI might make a difference in.

9:32

The amount of disinformation and misinformation that might be disseminated.

9:37

So it might help people, help actors reach more people.

9:40

It also might help to personalize misinformation.

9:44

So this is called micro targeting and it's a concept where personalized campaigns are

9:48

aimed towards individuals or groups, and it has been shown to be quite effective.

9:54

I think something that's quite relevant is the platforms on which

9:58

people are finding their news.

10:01

So we know that young people between 16 to 24, the majority

10:04

of them find their news online. So I think it's 80 percent find their news online.

10:10

And most of that is through social media. Not to kind of, you know, scare anyone because it's perfectly easy to.

10:16

Look at to see BBC news on social media.

10:18

It doesn't mean that people are just getting all of that news from fake sites.

10:22

But what's important is traditional social media sites use graph models where they

10:29

show you content based on the content that your network, your social network

10:34

is sharing and liking and engaging with.

10:36

When we look at TikTok, which is obviously going to be a big

10:39

player when it comes to sharing information before elections, TikTok.

10:44

doesn't use that model so much. So TikTok actually shows you information that comes from

10:50

outside your social network. It actually uses algorithmic recommendations to bring in new content.

10:57

So if we look at kind of what's new with AI in terms of being able

11:00

to personalise Disinformation or misinformation in, in able to,

11:05

being able to reach more audiences. Could we see more effective use of misinformation on platforms like TikTok?

11:12

Maybe, but that's not to kind of sow worries.

11:16

Smera Jayadeva: So would you say TikTok's the answer to echo chambers then?

11:19

We're breaking, we're breaking what, what things were before.

11:22

Megan Hughes: I wouldn't recommend spending all your time looking

11:26

for information on TikTok. I think it yeah, maybe, maybe The answer to echo chambers, but I know that groups

11:31

like, like Meta are exploring, changing.

11:34

They're kind of using the graph models, the social graph models.

11:37

So who knows, but I think you're definitely right that echo chambers

11:41

exist on traditional social media sites.

11:43

And we know that people are likely to kind of share things that they agree

11:47

with, with people that agree with them. Right. Smera Jayadeva: So surely voters are used to being sold something when it comes

11:53

to electoral promises and manifestos.

11:56

That's all the basis of electoral campaigning.

11:59

But doesn't that mean we've always been vigilant towards such, you

12:03

know, trends in communication?

12:05

Megan Hughes: Sure. In the interest of talking about a really, you know, timely, hot topic, can

12:12

I suggest we go back to ancient Rome?

12:15

Yeah, you know, just down the road, yeah.

12:19

Just down the road, you know, finding the really relevant facts here.

12:23

But there is an anecdote. There's a point.

12:25

So, so if we go back to the Roman Republic.

12:28

It's facing civil war. Octavian, who is Caesar's adopted son, wants to really get the public on side

12:35

so he can win against Mark Antony, one of Caesar's most trusted advisors.

12:41

So what does he do? He spreads a bunch of rumors that Mark Antony is a drunk.

12:46

And because he's having an affair with Cleopatra, he doesn't have

12:50

any of the traditional Roman values that would make a good leader.

12:53

I hope you see the point now. I'm trying to point out that this is a very, very early

12:59

example of misinformation, disinformation even, campaign.

13:03

So we can trace this back thousands of years.

13:05

Misinformation is definitely not a new problem.

13:09

It's something that, as you say, we've been dealing with for a while.

13:12

When we look at the impact of new technologies like AI,

13:16

there are some differences. So, you know, I mentioned being able to disseminate misinformation

13:21

more easily and to more, you know, more people, but there's also a

13:25

concept called the liar's dividend. I'm not sure if you've come across this.

13:30

No. So this concept was coined by a couple of US law professors and the concept

13:35

is that people can now claim that true information is false and you can avoid

13:42

accountability by relying on public scepticism and the belief that the

13:46

information environment is completely inundated with false information.

13:50

So that's something that, you know. We might expect to see we've we've seen an example of it actually in relation

13:57

to elections in Tamil Nadu in India, a clip came out of a minister accusing his

14:03

own party members of illegally finding finance or fraud, basically, and he

14:09

came out and said, No, I dismiss that. That's not true.

14:11

I never said that. But a later analysis of the clip by technical experts found that it It's

14:17

quite likely that the clip was authentic.

14:19

So that's one example we've seen.

14:22

We've not seen lots of examples of this, but it's definitely

14:25

something that, you know, there's potential there for it to happen.

14:29

Jonah Maddox: Yeah. So as that begins to happen, people's trust in truth will sort of disappear.

14:33

Bit by bit break down. It's funny, isn't it, that you think of kind of this as a sort of

14:38

highbrow topic, but it's basically just playground tactics, isn't it?

14:42

Completely. Smera Jayadeva: With all of this, how do we authenticate real information?

14:46

I know you said there are a couple of experts, but if there's so much of this

14:49

going around, are there any ways we can.

14:52

you know, trying to ascertain the truth, at least for an audience

14:55

that might not have that much time. So is there maybe someone out there doing this work for them?

15:00

I think Megan Hughes: there are a few things. So there's things that platforms can do and there's things that we can do.

15:05

And I think the first piece of advice I'd give is to maintain

15:09

a healthy level of skepticism.

15:12

It's important not to. believe all the hype and not to worry too much, because just as you mentioned

15:18

Jonah, if we get kind of really confused about the state of the information

15:22

environment and we think, you know, the waters are completely muddy, we

15:25

can't find true information anywhere. That's not going to help anyone.

15:28

And it creates a kind of sense of, of public anxiety that

15:31

might actually undermine things like real election results.

15:35

In terms of kind of practical things that, that people can do and platforms can do

15:40

as well, we've seen that, uh, pre bunking is a method that can be quite effective.

15:46

So this is a prevention rather than the cure method where you anticipate

15:50

the use of disinformation and you warn people about it before it spreads and

15:55

you provide factual information on a topic so people are kind of aware that.

16:00

Disinformation might be coming their way. I Jonah Maddox: read about pre bunking, that's not a word I'd encountered before.

16:04

Yeah, Megan Hughes: and it's actually been effective. They've used they've done some early studies on climate disinformation.

16:10

Um, and I think that platforms like Meta have actually started using

16:15

pre bunking techniques online.

16:17

So it's, it's proven kind of effective and platforms are deploying techniques.

16:21

Jonah Maddox: Looking at the stat you gave us about how few people are

16:24

actually exposed to misinformation means that The majority of information

16:28

we're getting is information and we should be told yeah, you can

16:31

believe a lot of what you're getting. Is that happening?

16:34

Megan Hughes: I think you're completely right. And it's really important that, you know, we do need to be encouraging

16:40

trust in the information environment.

16:42

So I think when you log on to Facebook, I think in the campaign period if you

16:46

share a post that's to do with a political party, for example, I, if I remember

16:51

rightly, a little comment comes up saying, you know, have you, have you checked this

16:54

source or have you checked the content? And I think that's a great example of something that could be done to just kind

17:00

of make people pause and think, Oh, okay.

17:03

Smera Jayadeva: So on the different methods that we're probably, we can

17:06

use either as an individual or that platforms are taking on, I've also

17:09

heard about the, The Coalition for Content Provenance and Authenticity.

17:14

So essentially content watermarking, is that going to have any real impact?

17:18

What can we see in the future when it comes to C2PA?

17:21

Megan Hughes: I think it's a great question. I think C2PA is.

17:25

a step in the right direction.

17:28

So it's a group of organizations that have come together and they have committed

17:32

to developing technical specifications to be able to trace the origin of media.

17:38

And there's lots and lots of ongoing research on watermarking, but there

17:43

are a lot of problems with it. So there's the adoption problem.

17:47

So if one platform adopts a form of watermarking and they're putting

17:52

notices out saying, you know, Oh, this content is AI generated, there might

17:57

be an assumption by users that any content that then isn't watermarked.

18:02

is legit. And that might not be strictly true.

18:05

So there's, there's, there's a, there's an adoption problem there.

18:08

And even if watermarking kind of becomes very good, I think we can assume that

18:14

sufficiently capable and sufficiently motivated actors, they'll get around it.

18:19

So it's a step in the right direction, but it won't be a kind of A great

18:23

solution that solves all of our problems.

18:26

So Megan, could you tell us about this CTAS report?

18:29

Sure. Yeah. So, so this has been a great project to work on and it's, it's ongoing.

18:34

So we've got a publication coming out soon.

18:37

That's, that's a briefing paper and then a longer form report due out later

18:41

this year, and what we've been looking at is the impact of AI enabled threats

18:47

to the to the security of elections.

18:50

And. We've been looking at examples of AI misuse from 2023 to date, and the kind

18:57

of takeaway that I'd like listeners to, to think of is that examples are quite

19:02

scarce and where they do exist they're really hyped up by mainstream media.

19:07

The, the risk isn't really in AI use during elections.

19:11

You know, there's, there's a small risk, but the major risk is the heightening

19:16

of public anxiety and the undermining of the general information environment.

19:21

And what we don't want is for people to lose trust in genuine,

19:25

authentic sources and information.

19:27

So that's the kind of key top line I'd want people to take away from our report.

19:31

Jonah Maddox: Yeah. Yeah. That's a really good point. Let's make sure that we're not Contributing to the hype about

19:36

misinformation with this podcast. So I suppose that that kind of leads us to any final thoughts from you.

19:44

A concluding statement, if you will. Megan Hughes: Sure.

19:47

I think the, the key message is, you know, misinformation has been

19:51

around for thousands of years. AI is relatively new to us all, but it is just a tool, so people will use it

19:58

for good and for bad, but please don't worry that it's going to hugely impact

20:04

all of the upcoming elections in this very important year for, for democracy.

20:09

There's a lot of hype, but we're yet to see any real evidence that AI has

20:13

actually impacted any election results.

20:15

So just think critically, check your sources.

20:18

Think about the content of news and that's it.

20:21

Smera Jayadeva: All right. So just before we leave, there's one final question.

20:26

Is you may hypothetically in a world where you are facing off for prime

20:30

ministerial elections in the UK, Megan, what legislation should you be elected,

20:35

what legislation would you spearhead?

20:38

Megan Hughes: Oh, That's a really good question.

20:41

I have to really think carefully because there's a lot of public

20:45

accountability with a public podcast.

20:49

I think that the Online Safety Act has made some good steps, but I think I

20:56

would like to see stronger legislation surrounding pornographic deepfakes,

21:01

because, you know, we've spoken about AI in the context of election security,

21:06

but 95 percent of online deepfakes are pornographic material, often of women.

21:13

So, you know, that's a huge problem that I think it got discussed a lot.

21:17

with what happened to Taylor Swift but the kind of conversation spiked

21:21

and then as dropped down a bit. So I think that that's a really important topic that we need to, we need to have

21:26

really strict laws in place to deal with.

21:29

Smera Jayadeva: I mean, that's a great point. I'd, I'd vote for you just on that.

21:34

Jonah Maddox: There's my campaign. We'll be coming to a bit of the the deep fake stuff later in the episode.

21:38

Thank you very much, Megan. You've been a wonderful guest.

21:41

We'll let you get back to saving the world.

21:44

Megan Hughes: Thank you very much for having me. This has been lots of fun.

21:51

Smera Jayadeva: Okay, Jonah. So for our second story, I really wanted to talk about robotics.

21:55

Robotics. So I saw a really interesting video the other day from Figure AI about

22:00

their new robot named Figure01 and OpenAI software has been integral

22:05

to the development of this robot. And the reason why.

22:08

I think I was so surprised by it is because of the way in which the

22:13

robot responds to some of the tasks that the person's asking them to do,

22:17

not only in terms of its movement, but also the way the robots spoke.

22:21

I think that was the first time I actually confronted the fact that, you know, this

22:25

isn't Something that's, you know, a few decades away, but this is something that

22:29

we're actively working on right now. Jonah Maddox: Yes.

22:31

It's a pretty amazing video. We will link it of course, in the, in the show notes.

22:36

For our listeners that haven't seen it, the launch video for figure

22:39

zero one has someone asking this shiny Chrome robot for some food.

22:44

It gives him an apple. And then proceeds to clean up a mess while explaining why it chose the apple because

22:49

it was the only edible thing on the table. I know the task of giving someone an apple doesn't sound hugely impressive,

22:55

but you do need to watch it to see how different it is at least of how I thought

23:00

of humanoid robots were progressing.

23:03

It's mad. Smera Jayadeva: Yeah. And you know, this, this startup figure AI is backed by some

23:07

of the biggest names in tech. Jeff Bezos, Microsoft a lot of companies have invested, I think over

23:12

a billion into the development of this technology and what's key to this new

23:17

shift that's happening is that open AI's recent generative AI software has

23:21

been a key part of the entire puzzle.

23:24

It's making the robot more dynamic and it's making that text to that

23:28

natural language speech a lot more impressive for the general audience.

23:32

And I think it, it really shows how quickly tech has been evolving.

23:36

I mean, if we see, you know, the industrial revolution times of the

23:39

early 1800s, early and mid 1800s to, you know, the quick jumps, the

23:45

rapid jumps that we saw from the year 2000 to now where, you know, we had

23:49

some basic computing and now we have Really, really, really smart phones.

23:54

And I just wonder if we're seeing this right now, what can we can expect

23:58

in like the next two or three years? Jonah Maddox: Yeah.

24:01

So are we going to see a massive increase in robots around us now?

24:04

Is, are we prepped for this? Smera Jayadeva: As I said before, you know, generative AI has been

24:08

instrumental to giving that boost to the robotics industry to make it more

24:12

dynamic and respond in real time.

24:15

But if you watch the product videos, it's far from our imagined idea of a

24:20

perfectly mobile and you know, a robot that's able to respond that quickly.

24:25

If you see some of these videos, especially of the ones

24:27

that look like little dogs. Yeah, it's a bit creepy to say the least, but But that's just talking about, you

24:33

know, more performance related aspects. I think there's also the general challenges of generative AI, some

24:39

of which we've already covered. Yes, Jonah Maddox: the impact on vulnerable communities.

24:44

Or is it the biases? Or the safety concerns?

24:46

Or the explainability? Smera Jayadeva: Or all?

24:49

Yes. It's pretty much all of that.

24:52

I mean, this isn't to say there aren't great users for robotics though.

24:56

We can use them to navigate difficult terrains.

24:59

For instance, NASA is working on a robot to navigate celestial bodies.

25:03

So you don't need to put a human at risk on, on the moon instead, a robot

25:08

may be able to walk around and, you know, pick up some space material

25:12

to bring back for research purposes. Yep.

25:15

But it is a giant leap for machines, jokes aside, there are studies showing that

25:22

there is success with AI and robotics in healthcare for mobility access and so on.

25:29

Interestingly, we can also integrate them into the larger internet of

25:33

things network infrastructure. And this might bring us one step closer to what we envision smart

25:39

homes and smart cities where all our devices are interconnected.

25:44

And they're perpetually consuming our data about our every

25:47

movement, our every decision. You know what I'm going with

25:51

Jonah Maddox: this. You say it like it's a bad thing, but I feel like I'm still so naive to how

25:56

this data collection really impacts me.

25:59

It's too easy to accept the TNCs we're bombarded with.

26:03

So what can we expect in the next few months? Smera Jayadeva: So for the next few months for manufacturers and this

26:08

ranges from Amazon to Boston Dynamics to Hyundai to Nvidia to Tesla, you

26:13

know, everyone's getting in on it. It's a rather even playing field as of now.

26:17

So if we're Jonah Maddox: imagining a sort of Jetsons esque future, then presumably

26:21

the production costs need to come down.

26:23

Smera Jayadeva: Well, if we continue on an unregulated path where robots

26:27

are affordable, it would actually come at the simple cost of your

26:31

data, your Agency, or even your job.

26:34

Who needs them? Do you think, do you think it is that dire?

26:39

It is interesting, especially from a market analysis point of view.

26:43

And, you know, if you take the language of these websites, these

26:47

robotics websites, it might lead you to believe we need these machines

26:50

to fill up these jobs and so forth.

26:52

And that we, in fact, are the more lazier humans, but that's

26:55

me reading between the lines. But fundamentally, many of.

26:59

The repetitive manufacturing jobs, which robots could replace not only very,

27:04

very low paying, but incredibly taxing.

27:07

So if one wanted to upscale and move out of, say, working in a warehouse where they

27:12

have rather repetitive tasks, they might not have the time because they're stuck

27:17

in endless shifts just to make ends meet.

27:19

Thus creating the working poor Jonah Maddox: side note.

27:22

Right. Or side thought. If you were to lose work, like, production lines, you could lose

27:27

the creativity that's born from it. Right. Is an interesting nugget.

27:31

Gordy Berry, who founded Motown Yeah. Was inspired by the production line.

27:35

He worked. on it and building cars in Detroit, he thought you could do the same

27:39

with a musician, like bring them in, send them up the production

27:42

line and come out with a hit.

27:44

He even had a quality control system like the car factory did where they would

27:48

make sure each song was like the best it could be before it left the hit factory.

27:52

Even Rerecording them with different singers and things like that.

27:55

So yeah, remove all repetitive jobs and we might not get another Motown.

28:00

Smera Jayadeva: Oh, wow. But I mean, are you saying we should continue keeping workers in very

28:05

repetitive factory jobs, Jonah? In case we get another

28:08

Jonah Maddox: Motown. Easy for me to say.

28:11

Yeah. Although I must say I used to be a very unskilled builders, builders laborer.

28:17

And that is easily the time that I've been most prolific in making music

28:22

and art and feeling really creative, not quite to Motown standard, but,

28:26

Smera Jayadeva: in all seriousness, there needs to be a lot more analysis

28:29

and review of what's going to happen to the state of our markets and,

28:33

you know, what economic models will look like with greater automation.

28:37

You know, we have a lot of fundamental assumptions about labor costs, about

28:41

knowledge, about information and so forth, but it really needs to, you

28:45

know, Get a proper deep dive as we see greater and greater automation.

28:54

Jonah Maddox: Click bait. I know what you are up to with your tantalizingly open-ended question and

29:00

error of seductive mystery . I thought I was kind of impervious to it until this

29:04

month when I found myself paragraph deep into an article titled OpenAI is exploring

29:10

how to responsibly generate AI porn.

29:13

Smera Jayadeva: Let me guess, they're not actually exploring

29:16

how to generate Porn at all? Basically, you're right.

29:19

Yes. Jonah Maddox: So what happened was this month OpenAI released draft guidelines for

29:24

how the tech inside ChatGPT should behave.

29:26

And with regards to not safe for work content, it says,

29:30

Basically, we don't do that. However, the article that I read that was in Wired and also Guardian focuses

29:36

on this note lifted from the document, and I quote, we're exploring whether we

29:41

can responsibly provide the ability to generate NSFW content in age appropriate

29:46

contexts through the API and chat GPT.

29:49

We look forward to better understanding user and societal expectations

29:52

of model behavior in this area. Smera Jayadeva: See you.

29:55

Can kind of see where the article got excited.

29:57

Jonah Maddox: Yes, can see where they got it from. But, but they were also told by an OpenAI spokesperson that we

30:03

do not have any intention for our models to generate AI porn.

30:06

So this segment is kind of at risk of becoming clickbaity itself.

30:10

Clickbait of a clickbait. Smera Jayadeva: But it does raise some important questions, I think,

30:15

about the future of generative AI and where we need to be more careful.

30:19

The platforms want users to have maximum control, but also don't

30:24

want them to be able to violate laws or other people's rights.

30:27

I think we touched upon it in our series where we looked at deep fakes

30:31

being used for generative porn, and since then there have been the very,

30:35

very public questions about it. Deepfakes of Taylor Swift.

30:37

Jonah Maddox: Yes. Yeah. As Megan touched on earlier in the episode.

30:40

And we'll obviously link the episode where Smera and Jesse talk about that from

30:45

last series as well in the show notes. So a month or so ago, the UK government created a new offense that

30:51

makes it illegal to make deepfakes. Sexually explicit deepfakes of over 18s without consent and OpenAI are

30:58

very clear that they do not want to enable users to create deepfakes, but

31:01

it is happening on some platforms.

31:04

I read an unpleasant article about the rapid rise in the number of schools

31:07

reporting of children using AI to create indecent images of other children

31:11

in their school, which is very sad.

31:13

Smera Jayadeva: I know, but I mean that we're talking about

31:15

something within schools in April, we saw the first of what will.

31:20

hopefully be a larger crackdown of sex offenders using AI.

31:23

A 48 year old man from the UK was prosecuted and banned from using

31:29

AI tools after creating more than a thousand indecent images of

31:33

Jonah Maddox: children. Yeah. So we, we need better tech, better regs and a better education

31:38

towards sex and respect in general.

31:40

Aside from the illegal and abusive uses of AI when we're talking about sex,

31:45

I can't see a future where some form of pornography isn't created by AI.

31:50

I imagine it's often the fringe communities that the tech isn't

31:52

specifically made for who improvise to make what they want and end up discovering

31:57

some new use case that no one thought of.

31:59

Surely it's going to play a part somewhere in the future of AI.

32:03

Smera Jayadeva: I would actually be more worried about AI driven porn.

32:06

There is no transparency on the data used to train some of the generative AI models.

32:10

And we also have the problem of poor explainability.

32:13

If we can even say there's any form of explainability.

32:16

In this case, there may be a chance that someone's photographic data

32:21

that may have been used to train a model and maybe somewhere down the

32:24

line, there's some gen AI porn, which looks very oddly familiar to you.

32:28

And I personally do not want to wake up to a future 20 years down the

32:31

line where a photo I uploaded on Facebook, completely non harmful ends

32:36

up being part of a training data set. That has very non welcome users.

32:41

Jonah Maddox: Yeah. And I wonder if there's something in the idea that if AI companies do explore

32:46

the more questionable avenues the resulting new architecture developed

32:49

could enable people with ulterior motives to jailbreak the system and use it

32:53

for their own even more dubious means. Smera Jayadeva: Oh yeah, definitely.

32:56

I mean, better tech doesn't mean we eradicate crime as much as

33:00

criminal justice, AI systems might make you want to believe.

33:04

The more interconnected our networks, I think there are more risks of cyber

33:08

operations, be it data theft, data leaks, or even model replication, where

33:12

they can reproduce some of these models and the outcomes and at the risk of

33:17

the person whose data is being used. Jonah Maddox: Yeah.

33:20

Okay, let's wrap it up there I suppose just to bring it full circle, back to

33:25

clickbait and having learned from Megan about being aware of what we read and

33:28

where we get our information, I suppose the message here is to be vigilant

33:32

although this clickbaity headline led us down a valid rabbit hole, sometimes you

33:35

could find yourself in a more spurious place, ew, think before you click.

33:41

Smera Jayadeva: At least we're on the right track when it comes to the law. It's good to see that, you know, there are active steps being taken to

33:47

make sure that people are protected and that there are court rulings now

33:51

that can be upheld in future cases.

33:53

Hopefully it's not the case, but you know, knowing how the world

33:56

tends to use tech, it wouldn't be surprising if we hear more about this.

34:00

as this technology improves. Jonah Maddox: Yeah, keep you posted.

34:07

Well, that's about it for this month. But before we go, Smera, I want to continue a tradition from the last series,

34:13

and that is our positive news segment.

34:15

So what made you feel optimistic about AI this month?

34:18

Smera Jayadeva: There's a lot been happening, but there's one story I want to focus on.

34:21

It's this big breakthrough with DeepMind's AlphaFold 3, essentially.

34:26

I've Jonah Maddox: heard of it. Smera Jayadeva: So the big breakthrough is that this AI system can now map

34:30

out protein structures quicker than ever to give cures for diseases.

34:34

So essentially improve drug discovery. Would you like to know exactly how that works?

34:39

Because I spent some time going into the physics and the biology behind it.

34:42

Jonah Maddox: I absolutely would, because I did read the sort of the

34:45

headline of this story and thought that sounds positive, but then I

34:49

read the rest and understood nothing. So I would love to hear Some help there, please.

34:52

Okay, Smera Jayadeva: keep in mind I'm not a doctor by any means.

34:55

If I was, my parents would be so proud of me, but okay.

34:58

Basically, proteins are the workhorses of the cell.

35:00

They're important for everything, and each protein is made up of

35:03

complex amino acid sequences.

35:06

The issue is that these sequences and how they make up the protein

35:10

is governed by these very complex physical and chemical interactions,

35:14

which has meant that humans trying to map it out have taken a lot of time.

35:18

Apparently, it's like a 50 year grand challenge for medicine and biology.

35:22

But now there's a computer that can do it for us.

35:24

And if it means it can map out. proteins, it's the future of drug discovery.

35:29

Why, you ask, is the future of drug discovery?

35:31

It's because drug molecules bind to specific sites on proteins.

35:35

So, if we know where those sites are on a protein to bind the drug molecule to, then

35:39

we find a way to make that drug effective.

35:42

Jonah Maddox: Very nice. Shout out AlphaFold3. Shout out AlphaFold3.

35:45

You like it. So, that's it for this month.

35:50

Thank you very much again to Megan Hughes.

35:52

Our excellent guest, thank you to Jesse behind the scenes, thank you to Smera.

35:58

I should also just mention that Smera, this week I watched you

36:01

perform at the Pint of Science event in London, which, where you were

36:07

performing your imagined future. You came from Mars from the year 2060 or something?

36:12

Smera Jayadeva: Yeah, 2064. Yeah, I came down from Mars.

36:14

It was a very hectic moment of traveling for me.

36:17

I don't usually come back down to terrestrial earth but I luckily got

36:20

the funds from a specific sponsor.

36:23

It was Jonah Maddox: Lidl, right? Yeah, it was really good.

36:25

And yeah, for the, for those interested in that our YouTube will have the

36:29

point of science in the future. That's Mira.

36:31

Well done. Smera Jayadeva: Thank you for everyone who listened this far and we can't wait

36:35

to see you next month with a new set of stories that we will cover in detail.

36:40

Jonah Maddox: Bye.