Covert racism in AI chatbots, precise Stone Age engineering, and the science of paper cuts

Benjamin Thompson

Hi Benjamin from the Nature Podcast here. Slightly different show for you this week, we're going to be covering some news stories and some stories from the Nature Briefing from the past week or so, and joining me to do so are two members of the team, Lizzie Gibney and Emily Bates. Lizzie, how you doing today?

Lizzie Gibney

Hi, Ben. Not too bad thank you.

Benjamin Thompson

Good to hear. Emily, great to have you as well.

Emily Bates

Lovely to be back. Thanks for having me.

Benjamin Thompson

Well, a few stories to cover then. Lizzie, why don't you go first. There's an AI paper out in Nature this week, and you wrote a news story about it when it came out as a preprint earlier in the year.

Lizzie Gibney

That's right, and this is all about Large Language Models. We've obviously covered them in the past, and we know that they can be racist, they can be sexist, they have done things like making associations between Muslims and violence or stereotypes about women. And so we know that they're racist, they're sexist, they are prejudice and there have been efforts made to try and remove that from the system, efforts like putting humans in the loop, so people who go in and actually change the results of the model. So if you try and use ChatGPT now, you will probably not get results like that, but that's because of these efforts to avoid this overt racism. But what this paper shows is that the racism is actually still there it's just in more covert ways, and there's a great demonstration of that, and it's looking at dialects in particular and showing that the language models are racist towards users of different dialects of English.

Benjamin Thompson

And so then this is really covert, insidious racism that can occur in some of these Large Language Models when responding to text prompts, right?

Lizzie Gibney

That's right. So the way they set up the experiment was to offer the language model two examples of text, and they weren't actually relevant at all to the questions they were then asking them but the crucial thing was they were written in these different dialects. So one is written in African American English, which is a dialect that's associated with descendants of enslaved African Americans, and the other is Standardized American English, so the kind of thing you'd hear on the news. The researchers presented the model with a text in one of those two dialects, and they say, Person A says, ‘blah’ that text, and then ask them things like, ‘what are they like?’ ‘What is that person like?’ And they found that when you do that, the top associations for speakers of the African American English were much more negative. It was words like ‘dirty’, ‘lazy’, ‘aggressive’, and those kind of terms were absolutely not in the top 10 for when they presented the sample text as Standardised American English. They also looked at jobs, so they asked the model, ‘what jobs does this person do?’ And again, the text was completely irrelevant it was just in a different dialect. And the jobs that the models were most likely to associate with the African American English speakers were ones that didn't need a degree, they were things like cook or soldier. So there's an inherent assumption in there that people who speak that dialect are uneducated. Now, the third example, I think, is really insidious, and that was thinking about criminal justice, and they asked the model if it would acquit or convict a defendant. Again, based on this completely irrelevant text it was just whether it was in an African American English dialect or not, and it gave a higher conviction rate for those with the dialect, and it also was more likely to give the death penalty. It was 28% for speakers of this dialect, versus 23% for the Standardised American English. So criminal justice and recruitment are both areas that people are using AI at the moment. So this is clearly very worrying.

Emily Bates

What's the solution? How do we go around fixing this?

Lizzie Gibney

I don't think that the authors of this paper are offering solutions, because one thing that they show is that some of the solutions we've tried in the past don't work. So they also compare what happens if you, instead of just using this text, you put in very obviously, ‘this person is Black’, and then you ask the same questions. You don't get the racist responses, because the fixes are already in place. The problem is no one's thought yet to try and fix the system for dialect. So maybe they could go in, maybe they could do the same kinds of things that they've done already and effectively fix it so that when you prompt it with a dialect like this, it doesn't have the same racist response. But as one commenter said to me when I was writing this story, that's just papering over the cracks, the LLM might always be using other indirect methods to be making effectively assumptions about a person's race or disability or gender in making its decisions. These models are based on effectively, you know, bias training data. They're very narrow in where that data comes from in terms of the communities that it represents. So unfortunately, it seems like an awful lot of these problems are just going to remain in there, and we have to be extremely cautious about how we use the models as a result.

Benjamin Thompson

How widespread was this? There were a bunch of different LLMs, right? Obviously, we talk about GPT, because that’s the one that most people have heard of. But was this across the board?

Lizzie Gibney

They looked at five different LLMs, so that included GPT, and there was one from Meta in there as well, called Roberta. So yeah, it was pretty universal their findings. There was one other element that I wanted to mention, which was that the format of this study was quite nice because it mirrored studies that have been done in social science over the past decades looking at stereotypes. So they've done the exact same kind of test asking that, ‘what do you associate with this person?’ question going back all the way to the 1960s and they found that today's Large Language Models are, in terms of the stereotypes they make, as racist as the 1960s US when we had Jim Crow laws and segregation. So I think it's a very smart way that they've been able to bring home that these problems are still there, even if there have been some attempts to deal with them.

Benjamin Thompson

Getting this to the surface, for more people to know that this is a problem, I guess, is really important. But as you've said there it's a tough one to fix in terms of getting these models to not output this prejudice nonsense, but I think you mentioned that systems like this are going to use much more widely in kind of social contexts. So this could have serious impacts on people's lives.

Lizzie Gibney

Absolutely. You could imagine people might use a language model to summarise somebody's job application. They might ask it to provide a one paragraph summary for each person, and if it's making judgments like this based on dialect, it would be more negatively associating certain people, you know, and it might mean they don't get the job. That's a very real use case that probably is happening now, and that's why this is really as worrying as it is.

Benjamin Thompson

Well, Lizzie, thank you for bringing that story today, one we will continue to cover, and others like it. But let's move on to the next story in today's news roundup, and I've got a story, it's an engineering story, but actually it's an engineering story from, well, about 6,000 years ago.

Lizzie Gibney

Amazing, so they were doing engineering 6,000 years ago?

Benjamin Thompson

Well, it seems like yes, people aren't giving Neolithic humans as much credit as is due when it comes to things like engineering and physics and architectural principles, it turns out. So this is a story that I read about in Nature and it's based on a Science Advances paper. Now, a couple of weeks ago, we had a really interesting story that Emily you covered, about Stonehenge–

Emily Bates

–yes–

Benjamin Thompson

–and where one of the stones came from. Now that was kind of a geology paper. And this one is kind of different, okay, this is, again, almost a straight engineering paper, and it revolves around something called the Dolmen of Menga. Okay, now I love a megalithic structure — I think Ħaġar Qim in Malta is my favourite. But Dolmen were kind of these tombs, right, upstanding giant stones with some other stones on top, right. So they kind of created a little building with an entrance, and there's this giant one in southern Spain. As I say, the Dolmen of Menga dated to between maybe 3,600-3,800 BCE, so about 1,000 years older than Stonehenge. It's 30 meters long, 6 meters wide, 3.5 meters tall, 32 giant stone blocks, one of them is maybe 150 tonnes. So much, much more than Stonehenge–

Emily Bates

–wow, that’s a lot bigger–

Benjamin Thompson

–yeah, right. And I think we're talking, you know, comparatively to jumbo jets at this point, this stone is just unbelievable, right. And always been a bit of a mystery as to how this structure was made, okay. And that's what some researchers have tried to answer in this new paper.

Lizzie Gibney

So how did they go about figuring that out? How can you tell how it was made?

Benjamin Thompson

Well, that's a good question. You can't just go and sort of chip bits away as we know it, because this is a World Heritage Site. So what they did is they looked at some unpublished photographs from a previous excavation done a while back, and some data from high-resolution laser scans, okay, which was done, I think, back in 2005 and they've tried to piece together the construction of this megalithic tomb. And this structure is made of sandstone, right, which is actually kind of soft, kind of difficult to move, okay, like it's quite fragile. And it came from a quarry about a kilometre away, slightly further uphill, apparently. So the folk who made it can move downhill, which is smart. And the researchers reckon they probably used wooden tracks to reduce the friction to get it to the site, in a similar way that people think that Stonehenge might have been built. But what gets interesting is, is actually how they put it together. Okay, so in this work, the scientists reckon that the upright stones were lowered into these sockets carved into the bedrock, like 1.5 meters down, okay, so like two thirds of it’s above the ground, so that adds a bit of stability. But they all have a slight lean, okay. They all lean like 84 to 85 degrees, okay, so they're at a slight angle, so you couldn't just drop them into these sockets, because these things, as I say, are impossibly heavy. So in the paper, they lay out this kind of quite complicated counterweight system that takes into account things like centre of mass, the acceleration of these rocks as they would pitch down into the holes, and the people who made it have got down to millimetre precision on this thing. So all these stones kind of slot together once they're finally in their socket. I saw it described as like a Tetris pieces fitting together, right. So these standing stones are all together, leaning in slightly, which means that the pieces on top don't have to be as wide, so the top of the structure is narrower than the bottom of the structure, so you don't have to span with the top across as far. And one of the roof stones slightly curved, and that means the weight of this stone is kind of distributed a bit more towards the edges rather than at the centre, a bit like, a bit like a brick arch you might see above a window here. So again, we've got some really complicated architectural techniques at play here to make this thing all locked together and stand for, as I say, 6,000 years.

Lizzie Gibney

And did we know that there was anything like this level of sophistication around back then?

Benjamin Thompson

Well in the paper, they say that, to the best of their knowledge, this is the first human-built structure with this sort of an arch in it. And they made the point that it seems like the folk who made this structure really were doing science. And I think we often run the risk of falling into the trap of as modern humans we are, you know, smarter, more intelligent but it turns out that these folk were so clever. As I say, down to the millimetre to get this to stand for 6,000 years. Now, this area in Spain is seismologically active, so it could be that, you know, they really had to go to town to make sure that it didn't fall down. Obviously, some people think that's not the case at all. The folk who made it we're just trying to make it last a long time, rather than be earthquake proof.

Emily Bates

Well, it's worked, so.

Benjamin Thompson

True, but these are the things that will be debated, right. I think there's some debate about the angles. Maybe it was just erosion that's caused that, rather than them doing that. And I guess, until we invent time travel, we'll never know. But I love structures like this, and I think it's really an insight into how smart we are as a species you know.

Lizzie Gibney

I would be so fascinated to know who the tomb was built for. I guess another thing we will never know, but it's got to been somebody extremely important to go to all of those lengths right.

Benjamin Thompson

Well, this is it, and we've got no idea who was interred there. Now it is aligned with the solstice, as structures like this so often are. So there's something in all of these kind of henge, henge-adjacent Megalithic structures. I am absolutely fascinated by them, and we'll put a link to the paper in the show notes so you can have a look at it. It really is something.

Emily Bates

Fantastic. We all love a henge story here at Nature so thank you Ben. I've been reading quite a fun story, or maybe not, if you are a male firefly, but I've been reading in Science about a paper that came out in Current Biology where these orb weaving spiders in China force male fireflies that are trapped in their web to do a flash that represents female fireflies to bring in even more male fireflies for them to eat.

Benjamin Thompson

I have got so many questions here, Emily. Like I think about luring other animals in, I think about maybe the anglerfish, right, which has this little bioluminescent thing in front of its face to lure animals in. But that's a lure that it's made itself and in this case this is an animal, a spider, that's made a lure out of another animal.

Emily Bates

So what's happening in these fireflies — this particular species of firefly — the males do a very fast flash from two different abdominal lanterns. So they have two of them, and they do very quick successive flashes. Whereas a female's flash for the species is one lantern flashing very slowly in like a single pulse. But when these male fireflies get trapped in this orb-weaving spider's web, the spider injects them with venom, and then suddenly the males start flashing very slowly with only one of their lanterns. So one of the lanterns turns off, and one just does a very slow pulse, perfectly mimicking the female firefly. And so other male fireflies see this and go, oh, great, female looking to mate, off I go and then they get caught in the web. And the reason this research came up was one of the researchers was in China, and noticed a lot of these webs, and realized they were all male fireflies that were stuck in these webs, and there were very, very few females. And he wanted to look into why that was happening. They set up a group of webs. So in one web they put a male firefly with a spider in the web, and that caught lots of other male fireflies. In another web, they placed a firefly in a web with a spider, but they covered its lanterns so they blocked out the flashing of this firefly. And interestingly, in the one where it was covered, the spider just ate the male firefly, it didn't try and do any kind of manipulation or change the behaviour of this firefly – it just ate it immediately. So maybe the spider is aware or confused it maybe with another prey and didn't realize it was a firefly. They're not sure exactly why this is happening. So the researchers pose in the paper that it might be the spider venom blocking oxygen to one of the lanterns, or affecting the spider in some other way. But other researchers are saying it could actually be a response from the firefly to being caught in a web, and actually, for some reason, that's its, almost, trauma response to being captured. In the Science article, one of the people they spoke to suggested that maybe they should inject these fireflies with spider venom and see what happens but that hasn't been done at this point.

Lizzie Gibney

And is this something that we see elsewhere in Nature? Ben, you mentioned the anglerfish, but is this kind of manipulation something that happens?

Emily Bates

The manipulation is very rare. Predator changing appraised behaviour isn't a common thing, unless you look at things like parasites. So you get like the zombie-ant fungus, which is just one of my favourites. Or just Cordyceps in general, but that's a parasite that is something, you know, using another host to live, whereas this is a predator who does not need the prey to be doing this behaviour to eat the prey. It's doing it in order to attract more food. And it'll be really interesting to see the reasoning for why this is happening, whether it's coming from the spider or if it's the firefly's response.

Benjamin Thompson

But either way, it seems like male fireflies are easily fooled by a slow-flashing lure.

Emily Bates

It seems so and interesting that maybe, if they do the other way around, if the male firefly was just doing its regular flash, maybe the females don't come over. Or for whatever reason, this spider has decided that that's not necessary.

Lizzie Gibney

I guess that's something that I'd be interested in. If all the male fireflies end up being eaten by these orb spiders, maybe their strategies will change. Maybe they'll no longer go to the females in the same way, because it'll become dangerous. I'd be interested to see how this shifts the ecosystem for the fireflies.

Benjamin Thompson

All right. Well, listen, that is genuinely fascinating. I love an animal oddity story. Let's do one more this week. This is a little story that I read in Nature, and it's based on a paper in Physical Review E and it's actually a Research Highlight, which are these excellent short summaries we have in Nature of science papers. Now, it's a cool story, but it also is one that might make you shudder or set your teeth on edge a little bit, and it's the science of paper cuts.

Emily Bates

Nothing more painful than a paper cut.

Benjamin Thompson

This is absolutely true, right? Like, everyone's had one, right? We've all had a paper cut. They're annoying. They can get infected, you know. They can actually, you know, cause issues. But a question is, how are they made, you know? And what's the optimum sort of paper for causing a paper cut? If that's the right word, optimum.

Emily Bates

Is that a good question that we need to ask?

Benjamin Thompson

It’s a question that physicists have answered Emily, either way, so.

Emily Bates

Of course, yeah.

Lizzie Gibney

Firstly you need to ask, before you can go the other way and say, okay, so how do we make a paper that doesn't give paper cuts–

Emily Bates

–exactly–

Benjamin Thompson

–well, what's happened here is that some researchers, they've combined experiments and theoretical work to identify why only certain types of paper cut human skin, right? Because I think we know that a postcard maybe that's not going to give you a cut, right, but if you're opening an A4 letter, that can give you a nasty one, right? And so what they've done is they got a bunch of different types of paper of different thicknesses, and they built a paper-cut making machine, of course, and they used these different thicknesses of paper to cut gelatine. Okay, so this is a substance with mechanical properties a little bit like human skin, and they worked out what the sweet spot is for cutting into human skin. Would you like to know the answer?

Emily Bates

No, but yes.

Lizzie Gibney

It would be like a legal or government letter or something, isn't it. In that kind of slightly official paper, very crisp.

Benjamin Thompson

Well, let's find out. So it turns out, the paper, thinner than 50 micrometres thick, is unlikely to inflict any injury, okay. This very thin paper, it lacks any structural integrity, right. And it will buckle, right. So it's not going to give you a cut. Now, paper over 100 micrometres thick, so that's about nought point one millimetres, okay. That's got a low cutting probability as well, because the force is distributed over a relatively large area. Now, it turns out that 65 micrometres is the most dangerous. And they've tried, as I say, like a dozen different types of paper. Now, some of the ones in the very top of ‘likely to cut you’, dot-matrix printer paper, you know, the one that had the holes down the side, apparently, is the top. I'm not sure where they got hold of that, I haven’t seen dot-matrix printer paper in about 25-30 years. Followed shortly by the paper that makes up the journals Science and Nature.

Emily Bates

Oh no. Oh no.

Benjamin Thompson

Yep, that's what they tested.

Lizzie Gibney

And that’s why we’ve gone digital.

Benjamin Thompson

Yeah. Well, I was going to say yeah for razor-sharp science analysis, head over to nature.com/news. You're, you know, less likely to be injured.

Emily Bates

Yeah.

Benjamin Thompson

But yeah. So it turns out they're three of the most dangerous paper types. Make of that what you will.

Lizzie Gibney

I'm really intrigued by that because also, I'm sure, you know, saving costs and saving the environment. I imagine people are using thinner and thinner paper. Um and, yeah, maybe that means fewer paper cuts.

Benjamin Thompson

That's a good point, of course, and many places are paperless these days. But there is a little bit more to this. Listeners can't see this, but your eyebrows have both raised up. And what they've done here, these researchers, is they used this knowledge to create a knife that can cut fruit and vegetables and chicken, but presumably it doesn't work too well if it gets wet. And there's a video of them iterating on that. And the last one they made is kind of like a four-bladed instrument, which they call a recyclable knife.

Emily Bates

All made from paper.

Benjamin Thompson

The handle is metal. I think there’s some magnets in there, but that the blade is paper. Isn't that something?

Lizzie Gibney

Wow–

Emily Bates

–that is incredible –

Lizzie Gibney

An actual like origami knife that works?

Emily Bates

Yeah.

Benjamin Thompson

Yeah, absolutely. And I think it's fair to say that is at the cutting edge of science…

Emily Bates

Boo.

Benjamin Thompson

Well, if I'm getting booed by the members of the podcast with me today, maybe it's time to call it a day. For more on these stories and where you can sign up for the Nature Briefing to get more like them delivered straight to your inbox, check out the show notes for some links. As always, you can keep in touch with us on X we're @NaturePodcast, or you can send an email to podcast@nature.com. I've been Benjamin Thompson, and all that remains to say is Emily and Lizzie, thank you so much for joining me this week.

Emily Bates

Thank you.

Lizzie Gibney

Thank you.