Language and AI Tensions

In this episode, Prof. Li Wei and Dr. Rebecca Curinga contrast human languaging, which is embodied, social, and multilingual, with AI language, which is patterned and disembodied. It highlights translanguaging, creativity, and identity, while raising concerns about bias and standardized norms. Educators are urged to build critical AI literacy, use AI as support, and sustain multilingual classroom practices.

This podcast season was generously sponsored by the Spencer Foundation in partnership with the Adelphi Faculty Center of Professional Excellence.

Transcript

SPEAKER_01 0:05

Language tensions where the messiness of language meets a good conversation. This podcast season was generously sponsored by the Spencer Foundation in partnership with the Adelphi Faculty Center for Professional Excellence. Welcome, I am Clara Bowler, and I'll be hosting Professor Lee Wei and Dr. Rebecca Kuringa today to talk about language and AI tensions. So I welcome our guests. Thank you so much for being here. And first of all, talk a little bit about you and how your work relates to language or languaging and artificial intelligence.

SPEAKER_00 0:49

Thank you and greetings from London. Hello everyone. I'm Li Wei. I'm presently the Dean of the UCL Institute of Education in University College London and Professor of Applied Linguistics. My recent work has made use of the concept of translanguaging. And as many of you know, translanguaging in very simple terms is a dynamic process whereby language users draw on different linguistic, cognitive, semiotic, and modal resources to make meaning and make sense. Transcending, the trans really means transcending the boundaries between named languages and boundaries between language and other meaning and sense-making resources. And translanguaging clearly builds on the concept of languaging, which has existed in the linguistics, anthropology, and cognitive science literature for generations. And languaging, of course, aims to reframe language not as a static fixed system of rules and patterns, but as a dynamic, social and embodied process as a verb, as people say. It emphasizes that we do language, we use it to act in the world. Now, languaging and also translanguaging places particular emphasis on the co-creation of meaning. Meaning isn't just transmitted, it's negotiated between people in specific context. And also embodiment. Language is tied to our physical bodies, our tone of voice, facial expression, gesture, and even our shared physical environment. Identity and relationship building, we use language to perform our identity as friend, parent, colleague, to build, maintain, or challenge relationships. And languaging has imperfections, and the messiness of languaging is also a very interesting and distinctive aspect of it. Languaging is filled with false dots, ambiguities, metaphors, and errors that are often central to creativity and understanding. And also purposeful action. We language to do things, to persuade, to comfort, to lie, to inspire, and to make a promise. My work doesn't specifically address issues around AI, but I've recently published a discussion piece with colleagues in the journal Applied Linguistics, talking about generative AI and its dilemmas from a translanguaging perspective. And I'm happy to talk about them in a minute.

SPEAKER_01 3:39

Thank you, Professor Liwei. And how about you, uh Dr. Karinga? How does your work relate to that? I know you're also involved in translanguaging.

SPEAKER_02 3:50

Thank you, and thank you for having me. My name is Rebecca Karinga. I work for uh the College of Staten Island, which is part of the CUNY system, the City University of New York. And I am a professor in the TESOL program. So my research and work is mostly focused on classroom research for English learners and multilingual learners in New York State. Specifically, I have been looking at adolescent education and newcomers who are English learners who have been in the United States for less than three years, and then a sub-population of those newcomers who have had interrupted or inconsistent education. We also call them SIF in New York State, the acronym. So these students are making up a growing population, especially in New York City, with a lot of immigration from countries in Central America and South America. And a lot of the things I've been looking at is how to develop curricula that employs translanguaging lens and thinking specifically about cross-linguistic skills for language acquisition, literacy acquisition, especially in the high school and adult learning context. Right now I also coordinate the program, so I've been working with my students in lesson planning and the specific classroom skills needed to support these students. Oh, and I will add, sorry, about the research in um artificial intelligence. So the latest research I've been working on is looking at how AI and specifically ChatGPT can or cannot work as a multilingual tutor for high school students and adult English learners when they have to prepare for proficiency exams and exams like the New York State Regents for Content Development.

SPEAKER_01 6:09

Yeah, so you both mentioned the importance of artificial intelligence in the work with multilingual learners and multilingualism, right, in the act of translanguaging. So what do you understand about the language tensions or any kind of language tension that might arise from or out of this work?

SPEAKER_00 6:31

So when we contrast the dynamic, fluid human processes of languaging and translanguaging with how current AI, especially large language models like GPT 4, handles language, several fundamental tensions emerge. AI operates on the statistical model of language. It predicts the next most probable token word or subword or sentence based on patterns in its vast training data. It has no inherent understanding of what words mean in a human sense. It knows how they correlate. Languaging is inherently semantic and intentional. We choose words to convey a specific meaning and intention, which may not be the most statistically common combination. So in our applied linguistics paper, we talk about flexible receiver and stochastic parrot. Generative AI can receive adaptive input and can generate flawless, grammatically correct text, but maybe devoid of genuine meaning because it's mimicking patterns without grounding. AI has no body, no senses, and no lived experience in the world. It has never felt the warmth of the sun, for example, the pain of a stubborn toe, or the comfort of a hug. It's knowledge is abstract and symbolic, derived from text, whereas languaging is fundamentally grounded in embodied experience. We understand the metaphor of a heavy heart because we have felt sadness physically. We understand sharp cheese because we've tasted it. So there is tension between human and machine understanding that is arguably unbridgeable with current architectures. And the tension is about the problem of grounding and embodiment. Another tension is seeing context as calculation versus the lived experience. For AI, context is computational. It's used to inform its predictions. It treats context as more data to be processed. For languaging scholars, context is rich, shared, and often unspoken landscape. It includes the relationship between speakers, physical setting, cultural norms, history of the conversation, and the emotional subtext. I mean there are other tensions like a large language model is a single static model. The database is expanding, of course, but it's finite. When it speaks, it presents a consistent, unified voice because it's based on a finite database, a fixed algorithm. In languaging, humans constantly shift our register, tone, vocabulary, indeed, choice of language. We're talking about multilinguals, depending on who we are talking to. We perform different aspects of our identity in different contexts through different languages. So that's another tension. AI is consistent of a single self versus an emergent property of lived experience and identity formation. And AI is inherently instrumental. Its purpose is to complete a task, answer a question, or summarize a text, to write an email. It's kind of a means to an end. Languaging is almost well, it's mostly relational and fatic. We use language to maintain social bonds, to fill uh silence, to show we're listening, even with no concrete uh information exchange. So fundamentally, there is tension between uh two very different perspectives on creativity. AI creates by recombining, by remixing elements from its training data in some novel ways. It's it's an engine for um fastidia and pattern matching. Human creativity often emerges from a desire to express a new feeling, uh a unique perspective, or to break existing rules. It's driven by emotion, experience, a need to connect something uh internal to the external world, and we do through uh languaging. So AI can produce output that is surprising and novel, but the debate rages over whether it's uh this constitutes true creativity or is merely a sophisticated uh form of interpolation and within within the vast data uh uh space. Those are the things that we talked about in our in our piece in the applied linguistics, but there are other tensions that I'm sure we can uh also talk about.

SPEAKER_01 11:42

Yes, I'm thinking specifically, uh Dr. Kuringa, of your work with youth and how they navigate some of the tensions that Professor Li Wei explored here, because for example, uh this idea of creativity and relationality, right? And then the youth using, let's say, Chat GPT as a friend, or have you seen any of this? And how how do these tensions come up in your work with youth?

SPEAKER_02 12:11

Right, that's a good question. So I think I can add on, and I I definitely agree with Professor Li Wei, that there is a lot of tension between the mismatch of how humans' language and how AI and and Chat GPT has uses language, right? The human languaging, especially for these newcomers and and youth, is fluid, it's multilingual and it's really culturally grounded, but AI is trained mostly on dominant high-resource languages. So there's a struggle there between the languages that the low-incidence languages that many of our newcomers in Cyfe speak and the ones that these large language models have been trained on. So the AI is operating mostly monolingually. It's a series of monolingual codes. And these are translated into systems. They treat each language as a discrete code, and each word is translated directly, uh, but it doesn't have the same fluid repertoire that our our newcomers are using and um in the classroom, and also as as they're approaching and bridging these new contexts of multilingual situations. So there are tensions, and a lot of our research also showed that when we would ask the AI, or we were using Chat GPT for, when we were asking questions as a multilingual tutor, we saw that we tried to train it. We tried to put in some different contexts on how to be culturally responsive, for example. But each time, even when we asked questions in Chinese and Spanish and English, it gave us almost identical explanations across languages. So it was really showing how AI is still operating from a really narrow, standardized model of language. So the tension is also within that cultural identity, and it hasn't caught up to the richness and diversity of human language and human communication. So I think there's a lot of tension there as well. And so a lot of, as I said, our a lot of our newcomers are speaking languages that are not the dominant uh standardized languages of the world. A lot of them are speaking um indigenous languages, let's say from Central America. And so there's not as much context of that within large language models.

SPEAKER_01 14:56

What tensions about languaging, racialization, and disability are emphasized when it comes to AI? And how do you think these tensions might reproduce, sustain, or change deficit perspectives in education? Because you both mentioned that the tension between the artificial nature of uh artificial intelligence, right? The language, and the more natural and human relational uh nature of human uh language, especially translanguaging, right? That is hard to reproduce in uh an artificial intelligence capacity. So, what do you think are some of the tensions that are reproduced negatively in these models? And conversely, what could be positive?

SPEAKER_00 15:45

We live in a world filled with diversity, and that includes race and disability. For many of us, our languaging experiences are intertwined with our experiences with race, disability, and also gender, age, etc. When it comes to AI, we really must remind ourselves that AI doesn't just reflect existing uh racial biases, it can actually amplify and automate them at scale. And that is such an important issue for us to remind ourselves constantly. The common text saying suggests that the biases in AI is due to biased uh training data. The deeper tension is that the data itself is a product of a racially stratified society. We have to recognize that. When an AI is trained on this data, it learns and codifies historical and contemporary injustices, presenting them as objective mathematical outcomes. So we really must be aware of the myth of technological objectiveness. And when it comes to disability, AI, of course, has massive capacity as a powerful assistive technology, but it also has the potential in reinforcing a narrow, able-bodied standard normacy. AI does not represent disabled bodies very well in its large databases, and many disabled people face uh specific challenges using AI technologies. And all of these issues that you mentioned, uh languaging, uh racialization and disability can be interconnected when it comes to AI, which will present very complex challenges. Just imagine a racialized ex-minority speaker of English as an additional language with speech impediment. They would most likely face the so-called perfect storm of AI failures, a voice recognition system that misunderstands their accent and disability, an automated hiring tool that penalizes their non-standard grammar, and a facial analysis system that will misinterpret their expressions. So we really need to develop critical awareness of these issues and critical AI literacy, especially for those of us working in education.

SPEAKER_01 18:37

For sure. There is some background noise uh here. For sure, that's such a great point. Um Dr. Kuringa, have you noticed that in your work as well, these, for example, these uh what Professor Lee, we mentioned as uh even punishment, right, for not sounding standardized or quote unquote perfect? Because even translanguaging is considered, right, uh quote unquote imperfect form of languaging, which we fight to get legitimized and affirmed as a legitimate form of languaging. So, what are some of these uh aspects of uh AI that you witnessed in your work?

SPEAKER_02 19:25

So that's a good question. I think that we have seen just what what we've been talking about already is that the AI has been reinforcing standard or standardized language practices and it marginalizes other language varieties in the way a lot of other technologies do, because it gives preference to normative and standardized language that it's pulling in from all of these sources and across the internet that it's bringing in. So, as I was saying before about the low incidence languages of a lot of our scife and students who are newcomers, these languages are not supported in programs like ChatGPT. And so the students who would be attempting to use these languages would be penalized or they would be misunderstood. And we had some experiences where we were speaking uh non-standard variety of Spanish, and sometimes it was misinterpreted as Italian. One time we, when there was low context, it interpreted our oral language as Japanese, and even when we spoke in English, one time it interpreted as Welsh. So it's still really not quite at the level that we would want it to be in the oral language um comprehension, and then it's responding to us also doing a lot of correcting in the in the standardized language, which is another issue, which is further uh marginalizing the language varieties of our students. So we see this language bias that's carried um not only within the technologies, but it we also see this in the classroom. Um a lot of teachers don't they expect that students that come from, let's say, Mexico or Venezuela, Guatemala, that they would speak Spanish, and then they're surprised when they do not do so well on Spanish assessments. Um, but the reality is that the these students, if they had gone to school part-time, maybe they had learned a little bit of Spanish, but it's not their dominant language. They also don't have the academic skills to participate in school on grade level in that language. So their actual language. The system is much more complex. Their real linguistic repertoires are much more uh fluid, as we've been saying. Um and these AI tools don't really work well for them because their languages simply aren't represented in the training data.

SPEAKER_00 22:15

If I may add uh another point, I think any norm-based system can reproduce and sustain the deficit perspective in the education. I mean, we know large language models are overwhelmingly trained on standardized, prestige uh dialects of a few dominant languages like English and Mandarin Chinese. It creates the tension between the vast with the vast diversity of global languages, dialects, crews, and social acts, with racialized minority learners and people with speech and hearing difficulties, dyslexia, autism, neurodivergence, and other disabilities. So, you know, as I say, AI can reproduce, amplify and automate these biases and give the illusion of neutrality and objectivity. And that's something I I feel really, really important to remind all ourselves. Educators and education policymakers must be aware of these uh issues and really aware of the illusion of neutrality and objectivity and find ways to challenge them in our uh um professional practice. So developing critical AI literacy and critical digital literacy is key in uh teacher education and must be integral to any teacher education programs. Uh of course, we we should promote the use of technology. There's no question, you know, I'm I am here to promote the use of technology in teaching and learning, including language teaching and learning. They do help with teachers' workload, lesson plan, uh, and many other areas of their work. But we must never forget the inherent biases in the databases on which generative AI apps are trained. We need to find ways to challenge these biases and not let the biases be reproduced or reinforced through assessment and evaluation, for example, and really always remind both the learners we teach and ourselves and policymakers of the what I call illusion of neutrality and technical uh and technological objectivity. These the data is not neutral uh and not objective, and therefore what AI produces, reproduces isn't neutral and objective, it's it's not value-free. And that's why we need to develop critical AI literacy and critical digital literacy.

SPEAKER_01 25:04

For sure. Dr. Kuringa, what do you think of this? What are the possible implications and consequences for your work and the teacher education work embedded in your in the research and uh work that you do with the SciF youth? Because you also mentioned that uh they also use these tools for testing, right? To prepare for testing. So I think that's an interesting discussion as well to include.

SPEAKER_02 25:32

Right. So I do think it's really important to take a critical lens. And one thing that we're looking at is how well AI has been responding to our prompts and questions in different cultural contexts and whether it's able to provide any uh differences there. And at the moment it seems not able or not really there, but we're hoping that in the next five to ten years there will be some growth in the bicultural bilingual sector on AI. Um, but one thing I do want to talk a little bit about is assessment. And I think that this is where a lot of our multilinguals in the secondary schools struggle because they have to pass a certain number of exams in English. And sometimes they're given a choice to take a language, to take one of the tests in one of their home languages. In New York State, we have about 15 different languages our tests are translated into, and and maybe more depending on the school and the and the language that the student might speak. But in this case, it's not it's still not a fair assessment, right? So if we have a student that comes in to New York schools speaking um a language like Quiche or Garifana from Honduras, but also went to school in Spanish, they have some understanding of Spanish, but that's not their dominant language. And now they've been in a classroom for a year in English, and they're expected to pass an exam either in Spanish or in English, in let's say history uh or global history, right? So we we don't have the capacity yet to be able to assess them across language or without a specific named language being assessed in a monolingual mode. Um, and so it's really difficult for these students who are used to using their full linguistic repertoire to apply their knowledge and then having to be um in this narrow context of answering questions. So I do think we're looking at trying to figure out how AI can be a more helpful tool in preparing for these exams, um, either by breaking down information just in English into shorter, simpler sentences, and using some of the other languages that both the student and the AI is capable in to help them prepare and understand how to respond to some of these contexts.

SPEAKER_01 28:26

I feel that this is such an important discussion because I see a lot of educators talking about ethical issues with AI, but we don't talk about how inefficient these tools can be, also in other areas, and also what leads to some of the ethical issues, too. Um, I don't know if any of you would like to comment on this because this is something that every time that I go into an AI conversation, it's all I hear about uh the ethical dilemmas, the cheating. So, what do you think?

SPEAKER_02 29:03

I think uh we're still we're still struggling with that, right? So there are policies that we can create about using and using AI to help prepare for something, but then it it is it's is it your own voice coming through? Is it your identity? How far is too far to go with using AI to support you in writing? So I think that is a really good question, and that's something that we're working on in our teacher preparation to have that critical lens on when is AI being used just as a brainstorming tool or as a support, but not as something that's actually doing the writing or doing the work for us. We don't want to lose our own creativity, our own identity in putting forth um the work in the in the classrooms. And we don't want our students to lose that either. So having more of an awareness and I think educating our teacher preparation programs, but also our students in being critical on that and and not using it as a crutch, I think is another area that we have to think about that maybe it's something to get started, to generate ideas, but not to use for a final product. So I think there are a lot of things to think about, and and also with our newcomers, there's another area that I'm a little bit worried about with the AI because a lot of them have struggled with trauma in their home countries, in their travel, and they need a lot of social emotional support, and they've gone through a lot of stress. So I think that AI does not have the capacity to deal with that in the context of learning language and literacy at the same time. So I think there's a lot more human interaction that needs to happen, and I think that we also have to be careful about that on the ethical side of things.

SPEAKER_01 31:13

For sure. Oh, I wish we had two more hours to talk about these things because these are so important and I feel they are so relatable and current. I think everyone is thinking about AI now and how it's helpful and when it's not. And right now we are going to a special section of our show, student questions. We have three questions here that we are going to explore. The first one is from Emily Ledesma. She's a teacher education student at Adelphi. And she asks, is language facing difficulties because of AI, or is AI making languaging easier in communicating with different languages? Are people uncomfortable with AI because it's a newer element and that is causing tensions?

SPEAKER_02 32:00

Okay, so this is a good question. I think that AI does make cross-language communication a little bit easier, but it doesn't automatically make it more equitable or more human. And that's the problem there. I think it's a little bit, I think that's a little bit what people are scared about. Um, AI is going to become more and more normalized because it's transmitted so easily and quickly across platforms. I worry that this may lead to fewer language variations in the future. For example, minoritized language varieties might become less common. We've already seen languages and dialects dying very quickly, and um many languages have fewer than 300 speakers, many language varieties, and with this normalization and standardization across AI and across technologies, we may see fewer language variations. And that said, on the other hand, I believe humans will continue to be creative with language, will and it will always be tied with their identity and their culture. This is something that machines cannot replicate or keep up with in real-time change. But the input and the carrying across platforms like TikTok and YouTube and all the reels that the youth are watching, a lot of the expressions across languages become more generalized and less localized. And the same thing with autocorrect features in these formalized languages. There's a lot of pressure to write correctly or to use the standardized form. And I wonder if youth are feeling that pressure from these technologies.

SPEAKER_01 33:51

I wonder about that too. Professor Lee, we have a question here from Madalena De Rizi. She's also a STEP student at Adelphi, and she asks, since AI opens up people to many various perceptions, what do you think is the biggest no to using AI technology for communicating and learning languages?

SPEAKER_00 34:13

That's a really interesting question. There is lots of debate and discussion about the uh ethical uh issues and ethical challenges AI has uh has uh raised, uh, and I won't go into that. In terms of practical uh issues to do with language teaching and learning, I think we must be aware of the uh what I would call the correctness trap or the illusion of competence. AI would give people the impression there is a single correct way of using language, and that is quite the opposite to the uh spirit of languaging and translanguaging. Uh, and AI can give you the impression that there is uh a single correct pronunciation, grammar, or even choice of words, and then you you you know will remember the native speakerism uh debate. You know, again, AI would give you the impression that native speakers have the full competence that everybody else should model on. And that is clearly something that we have been uh uh fighting uh against, and we don't want that to be reinforced through the use of technology in uh teaching and learning uh languages and using it for communication purposes either. So that's something that I think we absolutely need to be aware of the correctness trap and the illusion that somehow there is a native speaker normal standard way of using language. And that is something that you know the translanguaging and languaging scholarship has been really, really fighting against.

SPEAKER_01 36:05

Yes, for sure. We have been trying so much to get to normalize multilingualism and let's say imperfection, right? And then AI comes and then everything needs to be perfect again, and that's not human. So I fear that as well. And a question for the both of you from Julia Carpio, she is a new teacher. In what ways can educators use AI as a tool to promote multilingualism in the classroom while ensuring that speakers of all languages are able to effectively engage with these tools?

SPEAKER_00 36:44

Um for me, I think uh AI can be used to raise multilingual awareness, uh metalinguistic awareness as well, to uh use it uh for uh to compare and contrast, but focus on creativity and the creative aspects of language use and language practice. We can use AI as a tool for problem solving, not purely transactional or just as a translation tool. It needs to AI in fact can translate really, really well, but quite often it it uh you know gets into difficulties of uh kind of understanding and interpreting the real meaning, uh meaning intention behind what uh an interlocutor articulates or utters. So you we get into a situation where there's communication without understanding, and that's that's the kind of thing AI does and does quite well. But we can use that in the classroom, especially uh linguistically diverse classrooms, to help with problem solving, with uh comparison and contrast between different uh languages and raise metalinguistic awareness and the uh the you know all the interesting aspects of linguistic diversity.

SPEAKER_02 38:13

I agree with that. I think there are a lot of ways to raise metalinguistic awareness, and I think that AI can be a good tool for that. I also think on the surface level with translation, um, quick translations at least help some of our uh newcomers and emergent bilinguals have access to some of the classroom content, um, getting used to classroom norms and culture on the bigger picture and easily quick with the with the translation and with those kinds of connections, they can feel more comfortable right away. Um, because we don't have when we have 10 different languages spoken in a classroom, we don't have teachers that are equipped to help students on every in their home languages on every level. They may speak one or two languages of their students. Uh so in that case, I think that's a quick access point. I also think um AI can be a good tutor in English when it's able to help break down answers and directions and explain things in a simple sentences, and if it's prompted to do so, it can also add in um other multimodal responses so that students can quickly gain access. Although I will mention that there is still this big problem with the marginalization of the non-standard dialects and non-standard um languages. It has difficulty interpreting those. And um, so that might be one of the issues in the classroom to really engage effectively engage with all of the learners in the classroom.

SPEAKER_01 40:07

Yes, let's hope that we can also teach. I think what I learned from the conversation today is that we can also teach AI to be uh to address these issues and hopefully it becomes more humane than what it is right now, especially in terms of linguistic diversity and multilingualism. So I would like to thank both of you for coming to the show and engaging in this dialogue. I hope it's the first of many because I loved hearing both of you, and I'm a fan of both of your work, and it's such a privilege to be here with you, joining in conversation, and let's hope for more.

SPEAKER_00 40:49

Thank you for the opportunity, and I look forward to hearing all the other episodes of your podcast.

SPEAKER_02 40:56

Thank you so much. It was really nice to be here and speak with everyone.