Massive lost mountain cities revealed by lasers

Nick Petrić Howe

Welcome back to the Nature Podcast, this week: hidden cities in the mountains of Uzbekistan

Emily Bates

And how a digital watermark could help root out AI-generated text. I’m Emily Bates…

Nick Petrić Howe

…and I'm Nick Petrić Howe.

<music>

Nick Petrić Howe

First up on the show this week, reporter Dan Fox has been hearing about some enormous archaeological discoveries…

Dan Fox

An isolated plateau in the highlands of Southeastern Uzbekistan in central Asia, looks like an expanse of grassy rolling-hills. But look closer and a shard of pottery or the stony remnant of an ancient wall might hint at an archaeological secret hidden for hundreds of years.

Farhod Maksudov

It was a huge settlement, urbanized. Urbanized, it's not a village, it's urbanized. Because it has its own citadel where the rulers lived.

Dan Fox

This is Farhod Maksudov from the National Center of Archaeology at the Uzbekistan Academy of Sciences.

Farhod Maksudov

We think it was some kind of mountain civilization, very independent politically and economically from the huge empires which were down there in the lowlands.

Dan Fox

This week, Farhod and his colleagues have published a paper in Nature where they report a detailed survey of two ancient cities hidden in this area. They’re called Tashbulak and Tuganbulak, located 5 km apart, in the mountains close to the Tajik border. The tale of the two cities’ discovery began by chance when the team first stumbled upon evidence of the cities around a decade ago while investigating the regions highland basins, as Michael Frachetti, another member of the team, explains.

Michael Frachetti

The initial strategy was we were looking for locations where ancient pastoralists might have maintained habitations, in this case small camps or kept animals, etc. And it's these basins, because of their environmental situation, they gather more water, they have richer grass, you know, those are the basins we were targeting.

Dan Fox

Computational techniques had suggested these areas might have been fruitful locations to look for evidence of ancient nomadic farmers, and that’s what the team did find, but they also found something else.

Michael Frachetti

We found those small campsites, but we also stumbled upon the first of the two cities, Tashbulak, which presented itself as a much more densely occupied space. There was ceramics all over the surface. We could tell this place was really different from the standard kind of nomadic camps that one would expect in high altitude.

Dan Fox

The team started excavations at the sites, but the full scope of these settlements remained hidden until they could return to the area in 2022, this time kitted out with drone mounted LiDAR equipment. LiDAR or light detection and ranging, is a technique which uses laser pulses to map the precise topography of an area — archaeologists can use it to virtually peel back vegetation and sediment to reveal what’s underneath. The team knew something was there, but the scale of what the LiDAR survey revealed shocked them.

Michael Frachetti

We thought the site might be 25 hectares and that was almost double the size of most known cities or most known settlements in the high mountains of this region. What the LiDAR began to show us, and be able to illustrate in very clear detail was, you know, fortifications, all sorts of other built elements that enclose what we believe to be at least 120 hectares. That's on the order of 10 times the number of any known site that we're aware of in the region.

Dan Fox

In case you need some context – Tuganbulak’s 120 hectares would be more than the area of 100 football pitches, but it’s not just the size that makes these finds unusual, it’s where these cities were built; between 2,000 and 2,200 meters above sea level. As Farhod says, it’s an inhospitable location.

Farhod Maksudov

This elevation of 2,000 metres above the sea level is six months of winter, is a mountain steppe. You cannot do your traditional farming over there in the mountains. Nothing grows over there. That's why archaeologists never went there.

Dan Fox

Even now, it’s not an easy place to work.

Michael Frachetti

The nearest permanent village from Tuganbulak is about 15 to 18 kilometers. And so we rode donkeys up the mountain side and packed all of our gear up into this area on donkeys. We get snow in the summertime. We get really serious wind and storms. We’ve had very serious logistical challenges.

Dan Fox

In fact, today, only 3% of the world’s population live above this altitude. But nomadic peoples had inhabited the area since the late Bronze age, and later perhaps settled these mountain plateaus as they looked to secure economic and political power in the region.

Michael Frachetti

For Turkic nomadic communities, this would have been their bread and butter. This is where they live. And this is a speculation at this point. What we think is happening is that these mountain areas, which would have been outside the comfort realm of the lowland empires, were being consolidated by Turkic empires and Turkic confederacies in order to generate a competitive economic and political advantage.

Dan Fox

While further work will be needed to confirm exactly who built these cities, the fact they exist suggests that these mountainous regions are more historically significant than previously thought, positioning them more centrally in the culture and politics of the time.

Farhod Maksudov

We cannot ignore the significance of nomadic tribes, nomadic politics, nomadic political bodies which were at that time occupying very important strategic spots and dragging the roads towards them.

Dan Fox

These cities sat at the heart of Central Asia’s medieval silk road and were active through many prosperous centuries, so what ultimately caused them to be abandoned? One idea the team has is that iron production made these locations attractive but may also have contributed to their downfall.

Michael Frachetti

We know that they were developing a metallurgical industry and that industry requires resources. It requires both ore as well as resources to make the ore. So we used to burn something in the kilns.

Dan Fox

These mountains are rich in natural outcrops of iron ore, but turning it into useful goods requires fuel, lots of fuel.

Farhod Maksudov

They had started cutting down the trees. This area used to be covered by thick juniper forest, but they cut all of the trees for making iron because this is the only fuel at that time. So they cut the trees, but when they cut the trees, the area became environmentally very unstable because of the flash floods because of the avalanches and things like that.

Dan Fox

The authors speculate that this may have eventually led to a tipping point causing economic collapse, which, combined with political and social changes in the wider area, led to the cities being abandoned. Working out what actually happened will require a lot more research. Extensive archaeological excavations will need to be done to answer basic questions about who lived in the city, and what their lives were like. This should also help identify whether the cities were a patchwork of areas that were actually built at different times, something that LiDAR scans can’t unpick from above. But whatever they uncover, Michael hopes the findings contribute to ideas about the extent of human civilization at high altitudes and how ancient peoples shaped their environment.

Michael Frachetti

We have communities in the Andes, we have communities known in the Himalayas, but there's a lot of mountain areas that are just simply seen as wild territories or areas that are only used for resource extraction. Here, what we see is a community investing in this mountain zone and in a sense, taking ownership of it, building a city for quite a long time. The city seems to last, you know, with various phases of rise and fall for over 300 years, nomadic communities, various social organizations, can have the complexity to maintain major urban centers in places that even today would be quite hard to live in. I mean, if you were to ask your average individual to go up there and figure out how to live over a couple 100 years, they'd have a hard time innovating that city.

Nick Petrić Howe

That was Michael Frachetti from Washington University in St. Louis in the US. You also heard from Farhod Maksudov from the National Center of Archaeology at the Uzbekistan Academy of Sciences. To read their paper look out for a link in the show notes, where you’ll also find a link to a video Dan’s made where you can see what the team found.

Emily Bates

Coming up, a new way to digitally watermark AI-created text, that could help root out the real from the fake. Right now though, it’s time for the Research Highlights, this week read by Helena Kudiabor.

<music>

Helena Kudiabor

What do preschool kids and water vapor have in common? It turns out they both move in similar patterns. Researchers in Florida fitted children with radio emitters that let them track the kid’s movement in real time. The scientists collected data in a variety of settings, from playground to classroom. In the playground, the children were allowed to roam freely. Here, when averaged, their movement was similar to gas molecules. But in the classroom, where their movement was partly restricted, kids formed temporary clusters. This pattern resembles a phase of water in which freely moving individual gas molecules co-exist with liquid droplets. The orientation of the kids’ bodies also showed arrangements similar to atoms in both magnetic and non-magnetic materials. This isn't the first study to track how humans move collectively, but while other studies have explored how pedestrians move on a busy city square, this is one of the first to assess movement at relatively low speeds. Read this research in full, in Physical Review E.

<music>

Helena Kudiabor

It turns out that our coastal waters may be a lot dirtier than previously thought. New research has revealed that around the world, many bodies of water are contaminated by human sewage, but that this is being missed by existing testing methods. The standard way to test for water contamination is to look for the presence of E. coli and Enterococcus bacteria. Although these microorganisms are easy to grow in the lab, they are not unique to human guts. This means they cannot be used to identify waters tainted with human sewage. Researchers used an alternative test that detects certain species of human specific bacteria. They tested water samples from cities across the world and found that nearly half had detectable levels of human fecal contamination. In contrast, the original testing strategy suggested that less than 20% were polluted. Every city had at least one area polluted with human bacteria. Given the effects of contaminated water on human health and coastal ecosystems, the researchers hope their findings will lead to the adoption of more accurate water quality indicators. Read this research in full in Nature Water.

<music>

Emily Bates

AI-generated content is everywhere — text, images, music, podcasts. It’s not a stretch to say that this content is revolutionising different fields. But content is being created by computers that’s almost indistinguishable from content made by humans, and this presents a problem. This week, there’s a paper in Nature looking to ‘digitally watermark’ AI-generated text, to help people tell the real from the robot. Reporter Lizzie Gibney has been following the story, and she spoke with Benjamin Thompson about the work.

Benjamin Thompson

Lizzie, how you doing today? Thank you so much for being here.

Lizzie Gibney

Hello. Thanks for having me.

Benjamin Thompson

So, AI-generated content, then, it's all over the place, right? And you can see, in some situations, this is obviously a problem. For instance, in the creation of misinformation, say.

Lizzie Gibney

That’s right. So definitely, in terms of fake news, you know, you'd like to know if an article was written by a bot or not. And also, there are more technical problems with just not knowing if a text is written by an AI or not. In that there was a phenomenon I think we may even have talked about, called model collapse, which is where if your AI learns on texts which were already written by an AI, it kind of goes into a doom spiral and loses its quality and can completely collapse. So for lots of different reasons, political, academic, technical, we want to be able to label text as being AI generated or not. And when it comes to text, that's really hard. People might be familiar with watermarking on images. So you know, you might buy an image from somewhere and it comes with a watermark until you actually give them the money. Is it harder in text because you just don't have as many dimensions to play with. You know, you can't put a little filter over it, you can't do tweaks to the pixels with text. You’ve just got words, and if you change words, you might change the quality of the output. So it's just something that's a lot harder to do.

Benjamin Thompson

And that's really what the thrust of this new paper is, then. So researchers at DeepMind, that's Google's AI research laboratory, they've developed a new method to watermark AI generated text, but there's a lot that needs to go on with a watermark. I think of them in terms of bank notes, maybe as well, right. They can't be removed, they shouldn't change the content, but they also need to be invisible to the user. And as you say, in AI generated text, there are a lot of issues. So how does it work then? I know this involves Google's AI powered search tool, Gemini.

Lizzie Gibney

All right, so you probably know by now how a large language model works. It trains on loads and loads of text. It builds up these associations between words, and then when it comes out with is what we call a probability distribution. So if you have a word or a string of words which are called “tokens”. It then gives a probability of being the next word to every available token, and then it picks from those. Now your model is that series of associations, the step we rarely talk about, but which is being changed here is the sampling algorithm. So the way in which you pick from your probability distribution which token to use. Like a really simple way of doing it, not for watermark, would be take the top five probabilities and randomly select one of those. And this means that you end up with a model that produces kind of quite natural and variable language. You know, it's not one-to-one each time there's a little bit of randomness in there. So it's that sampling stage that the watermark is applied so it alters the precise words that are selected in a secret but formulaic way. That means that if you know the secret, you're able then to detect it afterwards. And it does add a bit of computational load, like it takes more effort to produce the watermark text. But they also use another clever technique that kind of also speeds up at the same time. So overall, they offset each other, and the watermarking is able to happen without making it a much more laborious process.

Benjamin Thompson

And so behind the scenes then, the algorithm is changing which words are picked, and that is invisible to us, but our computer can say, well, actually, this one is most likely made by an AI.

Lizzie Gibney

Exactly, right. So when you look at the text, it doesn't look strange. It still looks good quality. It's a reasonable word for it to have selected, but there is an imprint that is left, a kind of fingerprint that a computer is then able to detect.

Benjamin Thompson

And how do they go about testing this then, because, presumably, you need to test it with people to see can they tell the difference? Is the watermarks text noticeably worse or more incoherent, or something like that?

Lizzie Gibney

Well, this is what I think is the really big news in this paper is that they've actually deployed this. So if you used Gemini as your chatbot to generate some text in the last few months, it may have been watermarked. So 20 million users got text that was either watermarked or not watermarked. And what Google DeepMind did was compared how favorably the users rated it. So you can usually give a thumbs up or a thumbs down when you get a bit of text, and they saw that there was no statistical difference and a thumbs up thumbs down rate for the watermark text compared to a non-watermark text. So that tells us pretty well that users didn't really notice that the text had this watermark on it, which is good news.

Benjamin Thompson

Yeah, and that's one of the tenants of a watermark, is that it needs to be indelible, but not altering the product that is being used or read or whatever it is in this instance.

Lizzie Gibney

That’s right.

Benjamin Thompson

And so they've shown them that they can mark a text as AI created. People can't see it, but other computers can pick it up. My first thought was, surely one could just strip it out by putting the AI generated text through another AI that doesn't add a watermark, or just by me reading it and paraphrasing it and so changing the way that the words are selected, which essentially removes the watermark.

Lizzie Gibney

Absolutely. So that is a big problem with watermarks, and that still remains, but a few things within that. So first of all, they did test to see how robust which is, where they use, like, how well the watermark sticks when people just change a few words. You know, maybe you have got a text and you just want to put it in your own style. So you change a few words, you cut a bit, you add a bit. In those circumstances, the watermark holds true pretty well. They even tested if you paraphrase with another model, as you suggested, whether the watermark is still detectable. And if it's a long enough text, about 200 tokens, which is roughly 200 words, then yes, you can still. But shorter texts, it's not able to spot the watermark anymore. And what they didn't look at, which is really quite important is whether somebody deliberately attempting to get rid of the watermark would be successful. So with any watermark, that is just a problem that we haven't solved at the moment, right. Like, if you have somebody determined or malicious who wants to get rid of it, then you know, if they're smart enough, they will be able to. That's not to say watermark is useless, because a lot of the time, you know, putting just a little hurdle in people's way will stop them. Also, you know, a lot of people won't necessarily mind that the text is watermarked. They might be happy to go about using it, innocent use will be spotted. So even if we can't completely guarantee that a watermark is going to stick, that it's not going to be possible to remove it that doesn't make it completely useless.

Benjamin Thompson

And it's worth mentioning as well that Google aren't the first folk to try and do this. In terms of watermarking AI text. How does this version differ to others, and how does it stack up in terms of performance do you think?

Lizzie Gibney

In terms of performance, it's got higher detectability than other watermarks that have been shown on text so far, and it's just been shown that it works at scale. Like all the others that we've had so far, I think have been small scale tests, you know, academic runs. This is 20 million users that they've deployed it on. So on those two factors, they've really shown that this is an impressive watermark. And people have been working on these watermarks before. At different stages of the production of the text, you can apply a watermark. The idea for doing it at this stage of sampling, where you know all the probabilities that your model tells you of which token or word should be next, and it's the stage of just picking out from those which to use. That is the stage they've decided to apply the watermark. And that was an idea possibly last year or the year before, so they're building upon that. And they've just made the sampling algorithm more sophisticated. They've got a kind of multi layered process that ends up just leaving a richer signal, and that's what improves the detectability. There's a lot more there for the computer to see.

Benjamin Thompson

And what are researchers saying about this and how useful it might be?

Lizzie Gibney

I think most people are pretty positive. The fact that Google have done this really does seem like others may now follow suit, because there's also a little bit of kind of game theory going on. If only one chatbot comes with a watermark, people might just choose to use another one, right. But if all of them have a watermark, we enter a world where they all have watermarks. So there's a lot of positive kind of feedback to Google, in that sense. On the flip side, a bit more sceptical sources I spoke to said, well, you know, companies like Google have put these things out there into the world, and we now have this problem with not knowing what is AI generated text, and they're now putting out something that they can claim to be a solution – when we know that really it's not yet a solution. It does help in many, many ways. But as we discussed, if you're determined to get rid of it, you can get rid of the watermark. So it really isn't a panacea, and we don't want them to be able to claim now that it is because they've demonstrated that it works and that it works pretty well. So there's a kind of political side to it as well.

Benjamin Thompson

Right. Because, of course, there is a lot of discussion in policy and government circles as to the guardrails that might need to be put on AI and AI generated content and the potential benefits and or threats they might pose. How does this fit into that do we think?

Lizzie Gibney

The EU’s AI Act says that AI generated content should be watermarked, and in the US, President Biden's Executive Order last year also said something similar, we should have standardisation of watermarking. So they obviously have hit upon this as being a possible solution to all the problems we discussed at the start about not knowing what's AI generated and what isn’t is a big challenge ahead, really, for the companies, because this has been ordered, they've been told they need watermarks. They're coming up with watermarks, but obviously the watermarks aren't going to completely hold in every situation. So the hope is just that they keep getting better. So there's no harm in putting out papers like this and like others that just get better and better watermarks. And there's an overarching hope that eventually the entire problem will be solved – we are far from there at the moment.

Benjamin Thompson

Well, Lizzie, thank you so much as always for joining me to talk about all things AI.

Benjamin Thompson

Thanks Ben.

Emily Bates

Lizzie and Benjamin there. To read Lizzie’s News story head over to the show notes for a link.

Nick Petrić Howe

Finally on the show it’s time for the Briefing Chat, where we discuss a couple of articles that have been featured in the Nature Briefing. Emily, what have you been reading this week?

Emily Bates

So, I was drawn to an article in Nature where this scientist called Karina Heller scanned herself in an MRI machine 75 times over the course of a year, and this was all to measure changes in her brain throughout her menstrual cycle, both naturally and when she was on oral contraceptives.

Nick Petrić Howe

What was she looking to find or understand by doing this?

Emily Bates

So I think it's sort of fairly well known now that women's health, and in particular how women function on oral contraceptive, is chronically understudied in science. So she was really keen to see how her brain changed when she was taking the pill versus when she wasn't. So she started off she wasn't on the pill, so she scanned herself 25 times over five weeks during her natural cycle when she wasn't on the pill, this was to get sort of a baseline of changes in brain volume and connectivity between different brain regions naturally throughout her cycle. She also took bloods and did some surveys about her mood after every scan. And then she started taking the pill, waited for three months, continued taking it but did the same thing again over five weeks, scanned herself 25 times. She then stopped taking the pill, waited three months and did it all again. So 75 scans over the course of roughly a year.

Nick Petrić Howe

Oh, wow. And you know, you mentioned there that she was looking at things like changes in brain volume. Did she find such changes between taking the pill or not taking the pill?

Emily Bates

Yeah, so they found that when she was on the pill, her brain volume dipped slightly, and same with connectivity between different brain regions. So there definitely was a noticeable difference between the two states. But what is interesting is, after being off the pill for three months and doing the scans again, everything returned back to the way it was, just proving again that our brain is super adaptable and plastic.

Nick Petrić Howe

Well, you know, I guess I have to say this is N equals 1. So probably caveats abound in terms of interpreting this.

Emily Bates

100%

Nick Petrić Howe

But does it tell us anything more about how the pill is affecting us? Other than these sort of like metrics of brain volume and that sort of thing.

Emily Bates

Yeah, like you say, it's N equals 1. So this is a technique called dense sampling, where you take a lot of data from one participant, rather than minimal data from lots of participants. But 150 million people, more than 150 million people take the pill globally, so it is one of the most used medications in the world. So we need more research in this and what she's hoping is that she can compare her data with a woman with endometriosis, which is a very painful condition that affects up to 10% of women at reproductive age, to see whether hormone fluctuations in the brain could be driving the condition, and whether these hormones that you are ingesting with the pill just how they are affecting your brain, and sort of looking into that a bit further.

Nick Petrić Howe

Well, as you said, this is quite an understudied area of science, and I kind of like the fact that this woman just sort of took it on herself to do it.

Emily Bates

Yeah, absolutely. The article mentions she might be one of the most scanned women in science. Which you know, who doesn't want lots of pictures of their brain?

Nick Petrić Howe

Definitely. Well, for my story this week, I'm taking you to some forests in the mountains of Mexico. And this was an article I was reading in Nature about some attempts to kind of save the future of a monarch butterfly species.

Emily Bates

Oh, lovely. Are they particularly in danger?

Nick Petrić Howe

So this population of Monarch butterflies, it migrates from the US and Canada to winter in Mexico, and since the 90s, its population has drastically declined. Part of that is due to habitat changes, so less plants and things for it to eat, and also because of climate change. And it's climate change that this story really focuses on.

Emily Bates

So, what kind of forest in Mexico are we talking about here?

Nick Petrić Howe

So, these butterflies spend the winter in a particular kind of tree called an oyamel firs. And so these trees are also being affected by climate change. So there's been a sort of slow movement of these trees higher and higher up mountains where it's cooler. And so what this research is about is trying to understand whether it's possible to accelerate that to an extent, to take some of these fir trees and put them even higher than they would have been on the mountain, to sort of ensure that they'll be there for the future, when climate change has made things even warmer, and so that they'll be there as a place, as a refuge for the butterflies to spend the winter in. So it was thought that the natural elevation limit for this tree was 3550 meters. But the team behind this study have planted them at four different elevations and then seen how well they survived.

Emily Bates

And are the butterflies happy going up to these high elevations to find their favourite tree?

Nick Petrić Howe

Well, the reason that they've been doing this study is because they found some butterflies were at these sort of higher elevations, which suggests that possibly, if there were the trees there, it might be a good place for them to spend the winter. And so when they looked at the different elevations and how trees were doing, they found that, in actual fact, despite some of the elevations being above what was thought to be the limit, many of the trees were surviving. So it would be possible to plant trees at these higher elevations and, you know, keep them there as a refuge for the future. I must say, though they're not surviving quite as well. They are growing more slowly, and there are fewer of them surviving in general, but still, there could be some trees put up there to try and help the butterflies have a refuge in the winter.

Emily Bates

So this is just a test so far. Are they going to plant more? Or is this–this it for now?

Nick Petrić Howe

So this is just a feasibility study to see whether it would be possible. To actually do this would require a lot of trees and a lot of effort as well. But as one person who is quoting the article is saying, that may be the only choice down the road. The other thing to say as well is obviously it’s not just climate change that is affecting these butterflies. So an important step would also be to plant more of the plant species that they feed on when they're spending their summers in the US and Canada, and also looking at things like pesticide use and that sort of thing. But yeah, if they were to do this sort of planting to go forwards, they would need around 5000 trees to have reached maturity in Mexico's higher elevations by the 2060s so it would be quite a task if indeed they go forward with it.

Emily Bates

Well, I love a monarch butterfly, so I hope this works out for them. Thank you, Nick. And listeners, for more on those stories and for where you can sign up to get more like them straight to your inbox, check out the show notes for some links.

Nick Petrić Howe

That's all for this week. 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'm Nick Petrić Howe.

Emily Bates

And I'm Emily Bates. Thanks for listening.