Parrot or Prophet: How AI is Shaping Language
48:39 Share

On the Points North podcast, you hear great stories from the Great Lakes. Grew up right on Lake Erie, I should say. The Carp River. That was the last place we went. I'm from the Wukwemcung First Nation. It's kind of universally agreed that Whitefish Point has some of the worst field conditions. This ain't Mesquite Lake or White Lake. This is the Great Lakes, buddy.

Listen to Points North, a podcast from NPR and Interlochen Public Radio. This is The Pulse, stories about the people and places at the heart of health and science. I'm Maiken Scott. Gene therapy researcher Stephan Kadaki was working on a blog post about a recent work trip to Rio de Janeiro. And the blog post was about the collaboration between the Children's Hospital of Philadelphia and the Brazilian National Cancer Institute,

The blog post also mentioned a fun detail, that this collaboration started completely by chance when Stefan took a trip to Rio for Carnival in 2022. So,

So he's typing away on this post. He has a Google AI app on his computer, and he saw that with a few clicks, he could turn this post into a podcast with two hosts powered by artificial voices. So

So he thought, why not? A couple of minutes later, the app delivered. This Brazil project is fascinating. So Dr. Kodacki goes to Rio for Carnival and somehow ends up sparking this whole collaboration. It's wild. Right. It's like something you'd see in a movie. Yeah. It all started with that trip. He saw this opportunity to partner with INCA. That's Brazil's National Cancer Institute. They have the expertise. But like a lot of countries, they struggle with the cost of these. Stefan was impressed.

It actually reasoned through the material, highlighted key ideas, and made the whole story more engaging. But there was also something a little unsettling about it. It was both impressive and a little funny. The AI voices even said, uh, and uh. So it sounded like real people. It was honestly was a little bit creepy how natural it felt. I thought it was creepy too. It sounded way too real.

It's getting so much harder to distinguish between AI-generated language and human language. And you find yourself wondering, am I talking to a person or a bot? Human language is one of the key factors that makes us unique. It's how we express ourselves, how we share and bond. So what does this presence of AI-generated language mean to our existence? Does it challenge what it means to be human?

At the same time, AI is offering new ways to study language and to gain brand new insights into communication. On this episode, AI, language, and are we using a tool or is the tool using us?

To get started, let's hear from a linguist who's been critical of some of the hype around large language models like ChatGPT and what they can and cannot do. Emily Bender. She's a professor of linguistics at the University of Washington. A few years ago, she co-wrote what became known as the Octopus Paper.

It's a thought experiment, a made-up scenario that highlights the abilities and limitations of large language models. The thought experiment involves two people who are English speakers who are stranded on two desert islands...

And separated from each other, but they can see each other. They can see there's someone else over there, and they discover that there is some telegraph equipment. They both know Morse code, so they start sending each other messages. But soon, a hyper-intelligent deep-sea octopus discovers the underwater cable transmitting the messages. And it tracks over a long period of time the patterns that are coming through.

And then one day it gets mischievous and it decides that it is going to snip that cable because it has scissors because it's a thought experiment and start sending the dots and dashes back to one of the English speakers impersonating the other one.

But the thing is, it hasn't understood anything. All it knows are the patterns of what's likely to come next. One of the stranded people sends a message. What a beautiful sunset. And the octopus comes back with the dots and dashes that adds up to, yeah, it reminds me of lava lamps.

But of course, the octopus doesn't know what a sunset is or what a lava lamp is or even that that's what those words mean, right? The two people keep each other updated on their lives and achievements on their island. Hey, look, I created a coconut catapult. Let me show you how. And the octopus comes back with neat, cool, good job and things like that, which are just likely things to have come in that kind of a context. But one day, one of them sends an SOS message.

Help! I'm being chased by a bear. All I have is this stick. What should I do? And now the octopus is stumped. And whatever it sends back won't make any sense. And our joke in the paper is that if the octopus had gotten away with it that far, this is where the person, should they survive the bear attack, would figure out that they were not talking to the person on the other island, but to something that is just sending them back plausible sequences of strings and nothing helpful. And this...

gets to one of the big arguments when it comes to language models and language and what it is, which is the role of meaning and how language has meaning. When does it have meaning? How do you frame that? Where does language get its meaning from?

Yeah. So as a linguist, I see two levels of meaning. On the one hand, there's what we call the conventional meaning. That is, within a speech community, we have shared knowledge about what a word like book or cat or dog or rainbow or sleep, you know, it doesn't have to be just nouns, what these things refer to. And that's a resource that we all have access to when we speak to each other.

And then the second level is really communicative intent. That is what it is we are trying to get across when we choose some words, say them in a certain order, in a certain context. And both of those things are meaning. And the only thing a language model has access to is really a shadow of the first one.

Because words that mean similar things are going to show up in similar contexts. And that's all that the language models are modeling. And I guess even humans talking to each other struggle with this a little bit when we use text messaging, right? Because that second layer can get lost easily.

In those brief written communications, where we sometimes infer meaning into a communication that's not there. So I might read something saying, I'll be right there in a chipper way. Or I could read it like, I'll be right there. But sometimes that meaning can really get lost, even for us humans talking to each other.

Yeah, absolutely. And that's a great example of text messages. So the fundamental use of language is face-to-face co-present communication where we can see each other, we share the same environment, and

And we have everything that goes into our inflection, our tone of voice. Are we smiling or not? Which you can hear, right? Over the radio, you can hear a smile. But in a text message, you can't see it. And so we have to do a lot more inference when all we have is words and maybe some emoji compared to what we do when we are face-to-face, which still involves inference, but it's based on a lot more information.

In another famous paper of yours, you described language models as stochastic parrots. They are basically repeating, they are stitching things together to give them the semblance of meaning and sense. Tell me what exactly you mean by that.

Yeah, so the phrase stochastic parrots really took off, which has been fun to watch. And it is a way to make vivid the notion that these systems are not understanding, having some thoughts, doing some reasoning, and then coming back with an answer. But instead, they are, as you say, stitching together sequences of letters, even, not just words, from their training data in a way that matches the probabilities.

And this is where stochastic is important. It's not just random. It's not just rolling a die and picking whatever comes out, but it's random according to the probability distribution. So all of the information about what words went together informs that choice of what to output, and that's what makes it look plausible, but it is still just probability.

parroting and, you know, no shade on the lovely animals that I'm sure have internal lives. But this is parrot like the English verb to parrot, which means to repeat without understanding. So it is randomly repeating things according to the probabilities in the training data without understanding them. To which some people have argued, well, that's what we are. We're all stochastic parrots, right? I'm repeating something that you have said.

whatever I'm saying I've gotten from somebody else, so I'm also parroting stuff, could be the argument. Yes, and I've heard that argument, and it is an incredibly dehumanizing argument because it's true that we get ideas from other people and we build on them and we learn phrases from other people and we reuse them. But when we do that, we're doing it because they're meaningful to us and we're doing it with communicative intent. And to reduce our experience and our activity to just its observable form

so that you can claim that a language model is as good as a human is really to devalue what it means to be human. To you, is that truly the heart of what is at stake here? Why this is so important? It is an important piece of it. I think we really need to stand up for

What it is to be human and to be human together, as the Dean of Arts here at the University of Washington likes to say, and also to really see the value in our work and our understanding of the world and our creativity. And if we allow the sort of tech company approach of, well, if I can make something that looks like it, that's good enough.

to be the frame of the discussion, then we are really giving up a lot of ground that we should be holding on to and standing firm in. Now, a lot of these chatbots that we interact with are obviously designed to make us believe that they are humans and that they have reason that we are talking to an equal. Why? Why make it this way?

That is a really interesting question. And there's all kinds of design choices behind that, right? There's absolutely no reason that a chatbot should use the "I" pronouns because there's no "I" in there, right? But they are designed to do that. And I think part of it goes back to the history of trying to build systems that are familiar and easy to use. There's this idea that it's good for technology to be frictionless. And because we're used to using language to talk to other people, then if we can just use it the same way to talk to a computer,

then the computer will be easy to use. I think that that is some of the desire. And I think it's really misguided because that is a design choice that makes it hard to understand what the system actually can do and what you can and should not use it for. So I remember that old, do you remember the old paper clip that used to pop up when you would write something as clippy? And I always like perceived clippy to be annoying. Yeah.

But I also knew Clippy was not anything like me, you know, so it just felt like, okay, this thing is kind of annoying, but it was very clear. Our relationship was very clear. And these days I find myself getting just more confused.

Yeah, yeah, exactly. I think it is okay to use anthropomorphizing design choices. So Clippy, you know, wasn't shaped like a person, but it had eyeballs and a smiley face, right? And it certainly, the language that came out of Clippy was, again, speaking as if there were an eye in there, but its functionality was clearly circumscribed and easy to understand what it could and couldn't do.

What we have now with the large language models is things that are trained on so much data on so many topics and with so many different styles that they can mimic just about anything. And so it becomes especially important to make clear what the sort of edges of their functionality is rather than blurring it with this sheen of anthropomorphization.

Do you think we are even capable of talking to something that sounds like us without assuming that it is indeed like us? The tendency to anthropomorphize goes really deep. People frequently point to the fact that we look at power outlets and see a face.

We are constantly seeking other people in the world. And so something that presents itself as if it were a person is dangerous. How so? What is the danger here? What is at stake again?

So at an individual level, there is enormous chances of misinformation, right? If you have something that is speaking authoritatively and taking on the tone of, you know, a Wikipedia article or a medical explainer or something like that, it is very easy to take that as something that is knowledgeable and then make bad decisions based on incorrect information.

So that is sort of a huge problem on an individual level. You go one level up and you think sort of about our information ecosystem. People take this information, hey, ChatGPT told me, and then they put it out into the world. And they don't necessarily say where they got it. And now we have pollution sort of flowing through the information ecosystem. You know, I was trying to, I was writing a paper with my daughter. And of course,

The question nowadays is, do you still have to learn how to write a paper for your English class or not? You know, what's the point in writing the four millionth paper on the Iliad? But there is a chance that she might see something new or interesting or it'll get her thinking as opposed to this thing just like spitting out the paper. Yeah.

Exactly. And there's a really important set of issues in there. We don't ask high school students to write papers in order to keep the world's supply of high school essays topped up. That's not the point of that exercise, right? Learning to write is about learning to engage with ideas and organize thoughts and develop arguments. And if you just take synthetic text and drop it into your document, you haven't done any of that.

But you can copy and paste. You can copy and paste, yeah. And so it really is on the teachers to make clear why it is we're doing this work, what the value of doing it is. And, you know, that's an interesting challenge and maybe a silver lining to the current moment because that's something we should have been doing all along. And now we're forced to do it. I played Emily some of the AI-generated podcasts we heard earlier about the research trip to Brazil.

It's a wild ride from like the heart of Rio's carnival straight into some serious cancer research. Yeah, we're talking CAR-T cell therapy. It's a great example of how sometimes like the best solutions, you got to think outside the box, right? Connect those dots that you wouldn't expect. Totally. So before we get into the whole Samba and science connection, let's make sure everyone's on the same page. Okay, so to my ears...

Here is a machine that took written language, transformed it into a conversation, and it kind of sounds like something I would hear on NPR's Morning Edition. Yeah, yeah. Super cringy in various places, I have to say. You have to think outside of the box, Emily. Right, right. And gosh, the idea that somehow Brazil and science don't go together, of course there's science happening in Brazil. Yeah.

So that part is cringy, but also cringy you might come across from a real person too. And yeah, the voices were very smooth. Like if you think about the text-to-speech technology that we're used to from overhead announcements in train stations, even now they're still using pretty janky text-to-speech there.

This, you can't detect it, right? It sounds super smooth. It's got hesitations like a in it. So it really sounds like it might be coming from a person. And that makes me wonder, you know, what did they use for the training data for this? And if this is Google, almost certainly that's a bunch of YouTube content that's behind that, I would guess, to be able to train it, right? Yeah.

So, all right, so you've got something that's convincing. If you come across it, if somebody listening to this program heard only that part and didn't hear before and after,

they will easily believe that that was real people talking about something. And in that transformation from what the person actually wrote in their blog post draft, where they had accountability for the content into what goes into the podcast, things are going to creep in, right? And one of the issues with automation is that it allows you to do things quickly. And so if anything is wrong, you can sort of cause that harm over and over again. And it

also means that we rarely have the time to actually thoroughly check, do we like what they're saying? Right? It says, did the podcast version actually, was it faithful enough? Did things come in that we don't want to put out in the world? And if it goes out in the world, who's accountable for it? And I think the way to combat this is to get really good at citing our sources and to sort of insist on sources being cited in the sources we're drawing from.

and overtly valuing authenticity, like making it clear that this comes from a real person. This is how I found the information. Here's where you can go follow up if you want to go back and check. And the synthetic ones are going to start mimicking that because it's going to end up in the training data, but they're not actually going to do it well, right? If the source that's cited is even a real source, they're not going to be reliably connecting to it. And that's going to be sort of the remaining thing we can use to distinguish.

Emily has another big concern about these large language models. The way these systems absorb and then amplify biases. So we're taking training data with all of these things, who knows what, because it's not well documented. We don't know what's in the training data, but we can be pretty sure that it's got reflections of all kinds of prejudices in there. And because these systems are just picking up the probabilities, they are going to pick those up too and output them again.

Are there any other ways to protect our language spoken and written from being sucked into this vortex? Like hearing this podcast, I have to say I got a little angry because I'm like, wow, this is based on things that humans have created, right? Have worked hard. The machine sucked it in, is spitting it back out. Poof. Sounds great, right? But is there a way we can defend our own language?

words against this. Yeah, these are interesting questions. And I just want to point out, it's not the machine that sucked it in. It's the people who built the machine. Yes, yes. There's human agency behind this. And it's human agency with a profit motive, although it's not clear that anyone's actually making money with these things yet.

But the next step isn't, there's not clearly profit in there, but everyone seems to think there is. And so that's the motive for stealing all of this data. And I think, you know, the processes are slow, but I understand that there's some court cases now looking at this. So we might be able to do some sort of clarification.

collective pushback through regulation. And then just sort of at an individual level, you might think twice about what you put on the cloud and what you put out into the world. What are some things that you think will happen going forward in terms of language and these large language models? What are some things to look out for?

I think that it's always worth looking out for the different places they are getting integrated into systems and being skeptical. As a linguist, I'm intrigued to see what happens as people deliberately use language in more creative ways to distinguish what we're saying from the output of machines. I think that could be fun. But in general, I think it's going to be important and interesting to watch as we figure out how to distinguish synthetic from authenticated.

authentic texts and speech and really mark and value the authentic ones. Could you give me an example of using language in more creative ways to distinguish ourselves? Yeah, so basically coming up with new metaphors, coming up with new slang, for example, things that aren't in the training data. And one of the things that sociolinguists observe is

is that various kinds of pressure can drive rapid language change. So if you've got a community that is stigmatized and marginalized, they might rapidly develop and turn over slang to make sure that their speech is not comprehensible to outsiders.

And so it's a different kind of a situation, but might lead to similar outcomes, where if we have this homogenizing force of these systems that output the likely next word, and if people start using them as writing assistants, then you're going to get very similar text. And there's then space for people to distinguish themselves by coming up with a new metaphor, by coining new terms, by using words in new ways. And that can be an exciting linguistic moment.

Emily Bender is a professor of linguistics at the University of Washington in Seattle. You're listening to The Pulse. I'm Maiken Scott. We're talking about AI and language. Coming up, how AI is helping to preserve endangered languages. So I taught it. I said, this is the word for water. This is the word for drink. This is how you say, I drink water. Now, how would you say he will drink water?

That's next on The Pulse. This message comes from Stamps.com. Stamps.com simplifies your postage needs and adds valuable time back into your workday so you have more flexibility to focus on what only you can do. Go to Stamps.com and sign up with code NPR for a special offer.

I'm Tanya Mosley, co-host of Fresh Air. At a time of sound bites and short attention spans, our show is all about the deep dive. We do long-form interviews with people behind the best in film, books, TV, music, and journalism. Here our guests open up about their process and their lives in ways you've never heard before. Listen to the Fresh Air podcast from NPR and WHYY.

Catch the Pop Culture Happy Hour podcast from NPR.

When Malcolm Gladwell presented NPR's ThruLine podcast with a Peabody Award, he praised it for its historical and moral clarity. On ThruLine, we take you back in time to the origins of what's in the news, like presidential power, aging, and evangelicalism. Time travel with us every week on the ThruLine podcast from NPR.

This is The Pulse. I'm Maiken Scott. We're talking about AI and language. So far, we've heard about large language models that imitate human speech. Now, let's take a look at how humans are using AI to study language and to gain new insights.

There are about 7,000 languages around the world, and they do more than help people communicate. Languages are a huge part of a group's identity. They are the string that holds together history, culture, and heritage. But today, nearly half of the world's languages are endangered and expected to go extinct by the end of the century.

Native speakers are dying out and indigenous communities are still dealing with the aftermath of political persecution, centuries of being forced by governments to abandon their language. AI could be an important tool in preserving these languages, but there are some barriers that can get in the way. Nicole Curry has more.

When Jared Coleman was in college, he heard his great-great-great-grandfather's voice. It was an amazing experience. Jared is a member of the Big Pine Paiute tribe of the Owens Valley in California. The recording had been passed down through generations, and it's part of a large language archive in Bishop, California.

Jared's distant ancestor is speaking their native language, Paiute. And the first time I listened to it, I had no idea what he was saying. Growing up, you know, everyone knows some words. You know, you know, like the word for water, the word for sit, you know, just here and there are some random words. But

I didn't speak the language at all. Stumbling upon this recording was a pathway to learning the language. Jared started going through this archive, listening to the recordings of elders, interviews conducted by researchers and linguists. He even found a dictionary. So he decided to learn Paiute using those resources.

Jared is a computer scientist, so he also wondered if he could utilize those skills to make this process more accessible to others, kind of like Duolingo. He first built an online dictionary. But it was really with the kind of explosion of chat GPT and large language models that for the first time I saw an opportunity to do research in this area and learn

I saw the potential to use AI for an endangered language revitalization. Big tech companies like Google and OpenAI have begun using machine learning to revitalize endangered languages, hoping to save them from extinction. Essentially, machine learning models would be trained on examples from the internet, but this became an issue for Jared.

Even though Jared's tribe had this archive of recordings, Paiute is considered a low-resource language. It's just not enough. So, for example, Google Translate has used a machine learning model for a while now to do translation. And the way that it works is it's trained on, like, millions and millions and millions of sentences until it learns how to translate.

You can't do that for endangered languages. We just don't have that many sentences. This is an obstacle for many indigenous languages. So Jared tried to work around this for Paiute.

Maybe we don't give it a bunch of examples and train it like a traditional AI approach, but we give it rules about the language. We give it a vocabulary. And then using those rules, because ChatGPT is pretty smart, right? It can figure out how to construct the sentences that we want to construct. And then we'll see how you do. So I taught it. I said, this is the word for water. This is the word for drink. This is how you say, I drink water. Right?

And then this is how you say he. Now, how would you say he will drink water? You know what I mean? So, like, I'm testing it. I'm not giving it the answer, but I'm giving it enough information to where a human would be able to figure this out.

The tool is called a Large Language Model Assisted Rule-Based Translation. Advanced speakers and teachers of the language work with the tool as Jared described, and because of this unique process, there is another advantage of it: it doesn't spit out wrong translations when it doesn't know the answer, unlike other chatbots.

Like when you're talking with Chachupiti, you should never ever assume that it's accurate because it just makes stuff up all the time. Chachupiti can be creative and fun. And so you can never kind of separate when it's being creative and making things up.

Jared says the system knows its limitations. For example, if I ask, how do I say I'm going to eat bread and drink water in Paiute? Like our translation system will basically say, well, I don't know how to say that, right? Because I only know how to translate very simple sentences. But what I can do is tell you how to say, I will eat bread, I will drink water.

Jared says his model is still in the beginning stages, but he hopes the tool can offer accurate translations and help new generations to learn and speak their native language. Across the world in New Zealand, the Māori, another indigenous group, ran into a different issue while using AI to revitalize their language.

the Māori language, which is called Te Reo, started to decline in the 1800s. Māori language declined as a result of the increase of the European population in New Zealand.

And alongside that, there was intense linguistic discrimination. That's Peter Lucas Jones. He's Māori and he's referring to the Native Schools Act of 1867. Which barred the speaking of Māori in schools as part of the white assimilation process.

This meant that children that did attempt to speak the language at school were beaten. It wasn't until the 1980s that the Māori Language Act established Te Reo as an official language in New Zealand, along with English. This was an effort to persuade communities to speak their native tongue. But by then, there were far fewer people who could speak the language fluently.

Peter Lucas is the CEO of a broadcast company called Tehiku that's been assembling an archive of recorded speakers for three decades. And so we had been...

for 30 years running tribal radio, running broadcasting infrastructure for radio, television and our analogue type content creation. But what we quickly learnt was digitising our language was not only a priority for archival purposes, it was a priority to ensure that our language had a place in the digital future of the world.

keeping in mind that our language is counted amongst the endangered languages. Tehiku eventually decided to build a speech-to-text model based on this archive, but it turned out to be an intricate process. For example, sometimes the pronunciation of a word can change the meaning, which can lead to a wrong translation. The phonetic spelling and pronunciation of different words

impacted the model's accuracy. So they needed to benchmark the program, correct it word for word. And that's what we did. We taught our own people how to tag and label phonetical data, how to tag and label language data.

in a way that took into account the context, the cultural context. Peter Lucas stresses that accuracy is crucial when it comes to translation. Yesterday, I wrote something in Māori into a big tech translation tool. And I wrote, What I wrote in Māori was, Men and women, hold on to the land.

It provided me with an English translation and that English translation was white man, white woman, keep the land. So what I'm saying is it's important that as indigenous people, we actually set the benchmark around what the quality is for translation, what the quality is for pronunciation, what the quality is for accuracy and precision.

Because otherwise what we do is we actually excluded in developing what is a result from big tech leading for us. Peter Lucas says their speech-to-text model has reached 92% accuracy, outperforming similar attempts by major tech companies. They also launched an app that improves users' pronunciation of Tariho.

Aside from building accurate tools for their communities, Indigenous groups stress the importance of owning the data that they collect and use. Peter Lucas says it's sort of like controlling your own destiny.

Lots of indigenous groups were marginalized for years and stripped of their culture. And even some of the early language preservation projects kept archives locked up in universities, far away from the people it could benefit. So creating these useful resources and owning them dictates exactly how they will be used in the future. Our tools are only to be used for good.

And we want to ensure that the tools that we create enhance the lives of the people that are members of our communities and do no harm. That is why it's important to own the data. That story was reported by Nicole Currie.

Coming up, a researcher is digging through four decades' worth of audio recordings to find out how spoken language changes over time. In the middle of her conversation, she says, there's always crowds of people because I work 7.30 to 3. She actually pronounced that word work moving her tongue forward a tiny bit compared to most uh.

And you can see how she might modulate her pronunciation of uh in work in one context versus another. That's next on The Pulse.

On the Indicator from Planet Money podcast, we're here to help you make sense of the economic news from Trump's tariffs. It's called in game theory a trigger strategy or sometimes called grim trigger, which sort of has a cowboy-esque ring to it. To what exactly a sovereign wealth fund is. For insight every weekday, listen to NPR's The Indicator from Planet Money. We've all been there, running around a city, looking for a bathroom, but unable to find one.

Do you have a restroom we could use? A very simple free market solution is that we could just pay to use a bathroom, but we can't. On the Planet Money podcast, the story of how we once had thousands of paid toilets and why they got banned. From Planet Money on NPR, wherever you get your podcasts.

Great conversation makes for a great party. But how do you ask the questions that really make the room come alive? Well, here at Life Kit, we've got you. What is a path you almost took but didn't? On our latest episode, how to ask the magical questions that'll make your party sparkle. Listen to the Life Kit podcast from NPR. This is The Pulse. I'm Mike and Scott. We're talking about AI and language.

Large language models are helping researchers gain new insights about language. For example, by digging through a priceless collection of hundreds of hours of recordings captured over the course of four decades. They were collected on the streets of Philadelphia to truly understand how spoken language changes, which happens much faster than what we can gather from studying written text.

A linguistics pioneer started this work in the 1970s, and now a new generation of researchers is picking up where he left off. Grant Hill has more. Joshua Plotkin is a biology professor at the University of Pennsylvania who uses mathematics to study biology, specifically how organisms evolve and what drives those changes.

And he says those same mathematical models can be used to understand changes in vastly different settings. It's just as interesting for me to study viruses evolving, which I do do, or organisms evolving, but also culture and behavior evolving. And the mathematical models that the field has developed for understanding how viruses evolve...

could also be applied and are being applied to cultural evolution. At the end of the day, change, no matter what kind, comes down to probability. And the more data you have to work with, the more accurate your analysis about that probability. That's what drew Joshua to study the evolution of language. The huge amounts of data captured within thousands of years worth of writings and books.

It was like a fossil record unlike anything that could ever be found in nature. And for that reason, people have been doing this kind of work since antiquity, using text to trace how new inventions and mass migrations influenced the development of words and grammar.

Joshua thought his mathematical models could unlock new discoveries, hidden patterns and forces that quietly shaped the way we talk today. But when he told his linguist friends about the kinds of analyses he wanted to perform on this treasure trove of printed materials...

they were not as excited. Maybe that's interesting, but they kind of feel like, well, that's not really language. Language is like what we're doing right now. It's like, I'm talking, you're listening, you're talking, I'm listening. Real language, they said, was spoken, not written down. It's really complicated. It's messy.

It's fast and it's somehow effortless. With language, change happens quickly and every day in countless casual interactions. Like when you first hear your new neighbor say sheesh when the trash isn't picked up yet. The next day, you catch yourself saying the same thing when your grocery store is out of your favorite yogurt.

Or maybe it's pronunciation. A co-worker says "either" and now you say it this way instead of "either." And I thought, "Huh, how great would it be if we just had like some data set of some information about how spoken language has changed over time?" That would be like a goldmine of really studying a real important cultural change.

And then Joshua discovered that this kind of gold mine did exist, just a 15-minute walk right across campus. Right, so we have this whole bookshelf of all of these binders, and you can see each of them has, so if it says like pH 73...

Meredith Taminga is a linguist at the University of Pennsylvania and keeper of the Philadelphia Neighborhood Corpus, a collection of over 400 recordings of everyday Philadelphians taken over the course of four decades.

From 1972 until 2012. When I was a grad student, all of this material lived in its hard form in the tape closet. And at the time, Bill's lab was this old Victorian house. And I just lived in fear of the building burning down. Because it was the only copies of all of these recordings from this epic span in the history of sociolinguistics. The project was the brainchild of Meredith's longtime mentor, William or Bill LeBeauve.

Here he is speaking at a virtual awards ceremony three years ago. I entered linguistics at the age of 33. I'd worked for 10 years as an industrial chemist, and I brought with me the habits of numerical recording. In 1961, Leboeuf quit his job as a chemist to go back to school at Columbia University and, as he put it, enter into the pursuit of the universal properties of human language.

When Leboeuf entered the field, linguists had already begun to turn away from studying written language and toward fieldwork on spoken language, probing real people about how they talked.

Here's Meredith again. The field worker would go and they would try to find representative people in a little village. And they would be like, get me your oldest farmer dude. And then they would be like, here's my word list. And it's all, you know, sort of like farm implement terms. Like, how do you, you know, what do you call this kind of plow? What do you call that kind of plow? Bill LeBeau saw two problems with this methodology. First was all the note-taking.

It occurred to me that the field could profit by the adoption of this new invention, the tape recorder, and preserve what people actually did say. If you could record people as they talked, you could later measure the sound waves they produced and capture a whole new form of data. Data that can be analyzed to determine the exact placement of a speaker's tongue when they talked.

Still, Leboeuf knew these data points would be useless if you couldn't find a solution to the other problem. The fact that being prompted and observed by a researcher changes how people talk. They might feel self-conscious, use different words, or try to hide regional accents. So Leboeuf wanted to accomplish a new level of observation—

record people's everyday language while somehow making them comfortable enough to not think about that language they used. And his early work reared revolutionary results. For the first time, Leboeuf was able to show how class and social attitudes influenced single words, phrases, and even pronunciations.

This might seem obvious now, but Leboeuf put numbers behind it. He could quantify these pressures, measure their impact on how people said certain vowels. But eventually, he wanted to take this work one step further to capture how these expressions change over time. To accomplish all of this, he started a brand new linguistics class at Penn called Ling 560—

Meredith took it as a graduate student. This class is really like no other class that I've ever taken. It was a year long and the assignments were strange.

Students had to team up in groups of four or five and find a single city block in Philadelphia, one with parking only on one side of the street. These were the blocks where people are sort of physically closer together. Fertile ground for interaction and language exchange. The students would draw pictures of the blocks they chose, really study them, all while trying to go relatively unnoticed.

One thing that Bill would sort of teach you to do was don't go out there in your Penn sweatshirt. Try to dress like a normal person. Then, one by one, the students introduced themselves to the neighborhood. Hi, I'm a researcher and I'm from Penn and I have a microphone. Can I record you? They didn't mention that they were doing a study on language. They'd ask about demographic information, people's history on the block.

Questions about childhood games and even whether or not people had ever gotten into fights. The themes were designed to be things that people sort of found interesting to talk about or would get engaged in speaking about.

I started out as a SEALs girl and got demoted to the stockroom. This went on for 40 years, as Bill LeBeau sent generations of grad students all across Philadelphia, getting people to open up, be genuine. They had a SEAL, I'll never forget one day, and I made $116 and they took it away from me and they gave it to this old SEALs lady because I was called a contingent in the store. And I was so mean and I hollered and screamed and yelled and they demoted me to the stockroom. ♪

People have shared intimate details about their lives, their fears. And that's why, at least for now, not many researchers have access to these tapes. One of the chosen few is Joshua Plotkin, the biology professor whose office is only a few minutes across campus.

They have 400 different recordings of different speakers encompassing on the order of 2 million words uttered. When he found this data goldmine next door, he started reading Leboeuf's analysis and conclusions.

LeBeau's researchers had transcribed all of the audio recordings by hand, matched words with sound waves, and charted the exact position of people's tongues as they made these specific sounds. And this work led to new observations. It discovered that certain vowels changed their pronunciation really dramatically over the course of 100 years. Think water and wooder, high school and high school.

But Joshua noticed something different, that frequently people changed how they pronounced the same words in the course of a single conversation. For example, in a recording of an interview with a woman born in 1952. In the middle of her conversation, she says, there's always crowds of people because I work 730 to 3. She actually pronounced that word work differently.

moving her tongue forward a tiny bit compared to most /ə/ pronunciations that she herself made elsewhere in the conversation. There's many, many tens of thousands of other utterances of the vowel phoneme /ə/ and you can see how she might modulate her pronunciation of /ə/ in work in one context versus another.

Why was this speaker suddenly changing how she pronounced the word work from one sentence to the next? Randomness wasn't to blame. This was a systematic thing. Something seemed to push pronunciations back and forth.

And so for me, like the most simple hypothesis, because I'm not a linguist, like the most idiotic hypothesis is that one reason why you might want to pronounce things in a certain way is to avoid miscommunication. I want to pronounce things in a way so that even in the context of a conversation, you're not going to be confused that I've meant some other word. And so we would expect that, you know, from instance to instance,

People would choose, or maybe even subconsciously, I guess choose isn't the right word, just subconsciously modulate their sound production to avoid miscommunication. But now, Josh needed a way to evaluate this hypothesis. And to begin to study that, we have to know, well, what's the chance that when she says work in a certain way, what's the chance that a listener would hear some other word?

Words like wake or walk. Short words that start with a W, sound somewhat similar to work, and given the context, could plausibly be used in that sentence too. This is where ChatGPT comes into play.

And what GTP tells us is the likelihood, given the sentence structure, there's always crowds of people because I work 7:30 to 3:00. GTP tells us the likelihood of all those alternative words. And it turns out that work is actually the most likely word, the one that she actually uttered. But walk is also pretty likely. It's about half as likely as work.

So it's possible that the speaker made a subconscious effort to be extra clear when she said work, as if those same calculations Joshua saw happening under the hood of ChatGPT were also happening somewhere in the deep recesses of this person's brain. Moving her tongue forward actually moves it away from the ah and ahk. So she's definitely kind of

at least in this case, right, her pronunciation of the word work, her production of that sound was, it's as if subconsciously she was moving her vowel phoneme away from the only other word that might have been confusing there, "walk",

Joshua is still in the beginning stages of his research. This example could just be a fluke. But if we see the same thing happening, you know, over one million instances, then we start to think, okay, actually, there is a pressure to modulate your perception.

production of sound, even subconsciously, to avoid miscommunication. That's like the kind of research we want to do. And with the help of ChatGPT digging through these archives, he can do this research, and much more quickly than ever before. For The Pulse, I'm Grant Hill.

That's our show for this week. The Pulse is a production of WHYY in Philadelphia, made possible with support from our founding sponsor, the Sutherland Family, and the Commonwealth Fund. You can follow us wherever you get your podcasts. Our health and science reporters are Alan Yu and Liz Tong. Our intern is Julianne Koch.

Charlie Kyer is our engineer. Our producers are Nicole Curry and Lindsay Lazarski. I'm Maiken Scott. Thank you for listening.

Hey, it's Sarah Gonzalez. The economy has been in the news a lot lately. It's kind of always in the news. And Planet Money is always here to explain it. Each episode, we tell a sometimes quirky, sometimes surprising, always interesting story that helps you better understand the economy. So when you hear something about cryptocurrency or where exactly your taxes go, ya sabes.

Listen to the Planet Money podcast from NPR. Imagine, if you will, a show from NPR that's not like NPR. A show that focuses not on the important but the stupid, which features stories about people smuggling animals in their pants and competent criminals and ridiculous science studies. And call it Wait, Wait, Don't Tell Me because the good names were taken. Listen to NPR's Wait, Wait, Don't Tell Me. Yes, that is what it is called. Wherever you get your podcasts.

Fall in love with new music every Friday at All Songs Considered, that's NPR's music recommendation podcast. Fridays are where we spend our whole show sharing all the greatest new releases of the week. Make the hunt for new music a part of your life again. Tap into New Music Friday from All Songs Considered, available wherever you get your podcasts.