Portal de Periódicos da CAPES

Twenty‐first century technologies and language education: Charting a path forward

2024; Wiley; Volume: 108; Issue: 2 Linguagem: Inglês

10.1111/modl.12924

1540-4781

Richard Kern,

AI in Service Interactions

The COVID-19 pandemic, necessitating remote instruction in the vast majority of educational institutions around the world, put technology squarely at the center of language learning and teaching. Virtually all teaching and learning was mediated by Zoom or similar videoconferencing platforms. Students could see one another's faces (rather than the backs of their peers' heads), but they could also completely hide themselves from view. They used the chat window as a communicative backchannel for both authorized and unauthorized comments. They engaged in small group work in breakout rooms that were truly out of earshot of the teacher. Interacting with native speakers online was almost as easy as holding a regular class, once a mutually convenient meeting time could be negotiated. Web content such as videos, songs, news reports, artwork, and historical documents could be seamlessly integrated into lessons without the need for data projectors. However, in many quarters, this technological mediation weakened students' feelings of belonging, of camaraderie, of esprit de corps. Many teachers (and students) believe that learning remotely was not nearly as effective as face-to-face instruction (Muscanell, 2023). We have now returned to those face-to-face classrooms, and students (at least at my institution) are overjoyed to be in co-present community with one another. But the world is not quite the same as it was prepandemic. The fact that we were able to teach students remotely with some degree of success raises new questions about how we manage instruction going forward. Should we, in an effort to increase lagging enrollments, modify our 5-day-a-week teaching schedule so that we have 3 days in person and 2 days of synchronous or asynchronous online instruction? In theory, this would reduce conflicts with other courses, especially multihour labs. Should we, in an effort to "flip" our classrooms, shift the bulk of presentation of new material to online work that students do at home and then come prepared to apply that material to interactive and reflective activities in the classroom? Should we, to support less commonly taught languages and to broaden our teaching impact more generally, negotiate interinstitutional agreements that would allow sharing of resources (e.g., adding remote students from other institutions to in-person classes)? Should we, to expand our students' opportunities to use the language they are studying, have them do virtual internships abroad (or in the United States), promoting not only their language and cultural knowledge but also their professional skill set? These and many other questions are raised by new technological possibilities that invite us to reflect on what is important in what we do and to think creatively about what curricular experiments we might wish to try. But the biggest and most dangerous question that has arisen from the pandemic is whether we really need to go back to face-to-face language teaching at all. Contributing to this question is a recent strand of folk wisdom claiming that digital tutorials like Duolingo or Rosetta Stone, machine translation (MT) platforms like Google Translate, and generative artificial intelligence (AI) programs like ChatGPT have all but rendered obsolete the need for personal instruction in languages other than English in US schools and in higher education. After all, MT allows us to instantly give and receive informational messages in another language; ChatGPT can write coherent, grammatically correct essays for us on any topic and in all major languages; and tutorials can give us the basic vocabulary and grammar we might want to fill in the gaps to get by if we are traveling to another country. This growing perception of irrelevance is reinforced by the fact that language enrollments are dropping in just about every language except Korean and Hawaiian (Berg et al., 2023; MLA, 2022). This is the rationale upon which administrators at West Virginia University based their recent decision to discontinue funding their Department of World Languages, Literatures, and Linguistics (Anderson, 2023b). And, of course, all of this is superimposed on the misguided—yet longstanding and persistent—belief held by some administrators (and some faculty colleagues from other disciplines) that language teaching is essentially remedial activity and is therefore marginal to the mission of a research university. Our profession is truly at an inflection point. In the face of technologies that seem to provide the support people need to function reasonably well in another language, we urgently need to articulate and communicate the value of language study in a social context, identify what technology offers that is positive for language education, rethink how we organize our teaching in light of technology's affordances, and be clear about what technology cannot do. In this essay, I will discuss some of technology's affordances and limitations, focusing primarily on recent generative AI applications (MT and ChatGPT), then chart a pedagogical pathway harmonious with several theoretical frameworks, emphasizing critical engagement with technology guided by human teachers. Before launching into these considerations, however, let us first situate our current moment in the long history of relationships between language and technology. Language has been intertwined with technology for millennia. Despite innatists' claims that human language is an instinct (e.g., Pinker, 1994), some anthropologists (e.g., Finnegan, 1989) argue that speech is itself a technology. Writing is indisputably a technology, as it separated language from speakers' bodies, transforming it into visible form, and making it preservable beyond the moment of utterance. This in turn made linguistic analysis possible, generating metalinguistic notions like words and rules, leading to dictionaries and grammars that became technologies of language standardization. It also made it possible to scrutinize and critique ideas expressed in texts, essential to the advancement of knowledge. The need to document and preserve records gave rise to archival technologies and libraries (institutional forms of technology). Print technology made it practical to reproduce and disseminate texts on a much larger scale than manuscript technology could possibly have achieved and allowed libraries to vastly increase and diversify their holdings. The codex book and schools made the technology of reading gradually available to the masses, revolutionizing learning. More recently, sound technologies like the telephone, the microphone, radio broadcasting, and the tape recorder allowed the human voice to be heard well beyond its natural range and to be preserved so that being in a speaker's physical presence was no longer essential. Film, television, and video technologies added a visual dimension (gestures, facial expressions, and postures) as well as sound and music to recorded speech. More recently still, digital technologies and the Internet have integrated all of the above systems into a common underlying binary data structure that has allowed unprecedented mixing and manipulation of media. Speech can be instantly transformed into text, and text into speech. Even brainwaves can be transformed into speech or text (Guglielmi, 2019). Today, the content of popular sites such as Wikipedia, Facebook, X (Twitter), YouTube, TikTok, and Instagram is dynamically produced by literally billions of users, and for the first time in history, ordinary individuals enjoy the possibility of being read or heard by potentially hundreds, thousands, or even millions of people around the world. All of these technologies have evolved and been enabled by symbiotic and synergistic relations to cultural practices and societal needs and values—and have transformed how we do things with language (Kern, 2015). From this perspective, generative AI, momentous as it is, is far from the first (r)evolutionary development in technology that has had a major impact on language learning and use (and on education more broadly). These relations are reflected in the etymology of the word "technology," derived via Latin from the Greek "technê [art, craft, system, method]" (which in turn derives from the reconstructed Proto-Indo-European root "*teks- [to weave, fabricate]")—and the Greek "logos [word, speech]." In its earliest uses in English, technology referred to a systematic treatment of grammar. Today it is easy to lose track of this connection, as we think of technology as being associated with science and industrial innovation, and we typically think of writing and reading and language analysis as natural processes, rather than the cultural technologies that they are. From this perspective, we, as language educators, are all technologists, whether we like to think of ourselves that way or not. We need to attend to technology not because it is either a panacea or a peril, but because it affects how we use language every day as we move from one medium and setting to another, and we need to sensitize students to those differences. It is clear that the digital era offers a rich palette of resources, experiences, and opportunities for language learners. Online reference materials provide not only the definitions, examples, rules, and textual information found in bound dictionaries, grammars, and encyclopedias, but also provide audio for pronunciations, images and video for illustration, and hyperlinks to other sources for elaboration. Tutorials on linguistic and cultural topics abound, and are often presented as highly engaging videos on YouTube. Learners now have instant access to songs, films, current news, podcasts, radio broadcasts, and virtual museum exhibits in the target language that provide models of language use as well as enjoyment, cultural information, and stimulation of the imagination. And very significantly, the Internet affords engagement with other speakers of the language through social media, discussion forums, virtual exchanges and internships, and online game environments. Such engagements can broaden learners' exposure to sociopragmatic situations and provide opportunities to develop their L2 pragmatic competence (González-Lloret, 2019). Mobile technologies, such as smartphones, tablets, watches, glasses, and other forms of wearable technology—combined with location services that allow a device to "know" where it is—make it possible to annotate real-life objects with sound, text, images, and animations (as in viewing a painting in a museum and activating an audio narrative as well as superimposed visual markers or images of other paintings by the same artist for comparison). Mobile technologies can also be used to "rewild" approaches to language learning, integrating language use in and outside the classroom (Hellermann & Thorne, 2022; Thorne et al., 2021), or to explore linguistic landscapes (Bruzos, 2020; Malinowski et al., 2020) of multilingual communities at home or abroad. Preliminary studies (e.g., Dizon et al., 2022) suggest that intelligent personal assistants like Amazon Alexa, Apple Siri, Google Assistant, and Microsoft Cortana may also be useful, especially for learning vocabulary and pronunciation. The most recent and controversial forms of technology fall under the rubric of generative AI. These include MT applications like Google Translate and ChatGPT, which will, on command, produce texts tailored to task, genre, and register or provide feedback, correction, or translation of students' writing. Some teachers view these as threats to student learning and have attempted to ban their use. Other teachers fully embrace them and are excited about how they will empower learners. Many teachers fall somewhere in between with mixed feelings. We will consider their affordances and limitations below. While early forms of MT go back to the 1950s, it was the arrival of Google Translate in 2006 and the introduction of neural network technology (replacing statistical models) that propelled its use on a massive scale. Google Translate allows users to get instant translations in some 133 languages by typing or handwriting text, by speaking into their phone (thus allowing translated conversations in real time), or by pointing their cell phone camera at written text or even an image.1 Google Translate was joined by other platforms such as Microsoft (Bing) Translator in 2007, Reverso Context in 2013 (providing examples of use from a wide range of registers), Pairaphrase in 2014 (offering retention of original document formatting), and DeepL in 2017, claiming on their website (https://www.deepl.com/whydeepl) to offer "the world's most accurate and nuanced machine translation." A flurry of research on MT has appeared recently, and an excellent starting place is the special issue of L2 Journal edited by Vinall and Hellmich (2022). For language educators, MT presents a number of issues. Students can theoretically compose an essay in their native language and then automatically translate it into the target language. Short of such wholesale application, inattentive use of MT on words, phrases, and sentences is likely to contribute to incomprehensibility of student writing (increasing correction time) and may short-circuit the learning process. Consequently, some instructors or whole departments have decided to ban the use of MT. But MT is readily available and attractive to students, and O'Neill (2019a) reported that almost 90% of the Spanish and French students he surveyed used MT programs at least occasionally, even when they were prohibited from doing so by their teachers. While teachers' attempts to determine if students have used MT in their writing are likely to be futile, teachers can engage students in discussions about how they use MT and how they might be able to refine their strategies. They can point out that because translation algorithms rely on context to predict the most appropriate translation, MT is the least accurate when only a single word is entered (what learners most commonly do), according to Hellmich and Vinall (2023). They can show students how reference materials of all sorts—traditional bilingual and monolingual dictionaries, thesauri, and grammars as well as electronic resources like WordReference.com, search engines, and language corpora—can be used in tandem with MT to crosscheck with other sources and compare options. They can show them how subtle modifications in input into MT systems can lead to quite different search results, and, most importantly, they can treat students as co-researchers in exploring best practices. Hellmich and Vinall (2023) pointed out that instructors' beliefs and policies have an impact on students' success in making effective use of MT, but that instructors themselves may not have adequate understanding of how MT works in order to provide useful guidance. Such training may optimize students' use of MT, but it should not be expected to improve students' language learning per se. In one training study, O'Neill (2019b) compared the composition scores of 310 intermediate-level Spanish and French students using Google Translate or the online dictionary WordReference.com (each with or without training) with those using neither resource. The 40-minute training involved review of the tool's functions (Google Translate or WordReference.com), demonstration and discussion of the potential advantages and disadvantages of using the tool, practice in using the tool to translate sample words, expressions, and sentences in Spanish or French, and a self-monitoring phase in which students were to decide how much or how little to use the tool. O'Neill found that students who used Google Translate after training scored the highest on two letter-writing tasks, followed by those who used Google Translate without training. Next were the WordReference groups with and without training, respectively. The control group, which did not have access to either tool, had the lowest scores. In order to see if tool use had any long-term negative effect on students' unassisted writing ability, a posttest was administered 1 week later and a delayed posttest was administered 4−5 weeks later with very similar prompts but with no online tool use allowed. On the 1-week posttest, the WordReference groups scored no worse than the control group, but the trained Google Translate group did worse than any of the other groups. On the delayed posttest, there were no statistically significant differences across groups. These findings suggest then, that although use of online translation or online dictionaries will likely improve students' performance on certain kinds of writing tasks, they do not make those students better writers when they are not using those tools. As O'Neill pointed out, the tools may not contribute to student learning, but they may allow students to achieve more effective communication in the act of composing. In a more recent study, Lo (2023) explored whether editing with MT improved Chinese English-as-a-foreign-language learners' vocabulary learning and retention and found an immediate recall effect, but that 2 weeks later, lower proficiency level students had significantly declined in recall compared to their higher proficiency peers, suggesting that language proficiency may be an important factor in the degree of vocabulary learning from MT use. In any event, it remains unclear what long-term learning accrues from using MT in L2 writing. The AI project that has captured public attention like none other in recent years is generative pretrained transformers (GPT).2 Unlike a search engine, which finds existing information on the Internet, generative AI systems create new text, images, audio, music, video, and other content by recognizing patterns and relationships in massive datasets on which they have been "trained."3 Based on those semantic, syntactic, and genre-based patterns, GPT can (in response to human prompts) quickly write an essay or speech in a particular style, summarize large bodies of research, solve problems, make analogies, or produce a poem, scenario, or prose narrative—among many other things. Other generative AI platforms can also create artwork and musical compositions using text prompts. Generative AI is thus a technology that potentially competes with human creative efforts rather than merely assisting or complementing them. ChatGPT, introduced in 2018 by OpenAI, has now been joined by other similar chatbots such as Bard (Google), Bing (Microsoft), LLaMA (Meta), Perplexity AI, and YouChat. GPT-4, the most advanced transformer at the time of this writing, was released in March 2023, and is a large multimodal model that processes images and sounds as well as text input. Much more powerful than GPT-3.5 (available for free), GPT-4 requires a paid subscription. While GPT-4 can pass the multistate bar examination taken by US law students (Katz et al., 2024), even GPT-3.5 can pass the National Board of Medical Examiners test at the level of a third-year medical student (Gilson et al., 2023). ChatGPT has already been integrated into Duolingo Max (which adds AI role play and "explain my answer" features), and will soon be integrated into Microsoft Word, PowerPoint, and Outlook (Warren, 2023). The mechanization of text writing (and thus the mechanization of knowledge) has been satirized at least since the time of Jonathan Swift. In Gulliver's Travels (1726), we encounter a professor at the grand academy of Lagado who shows Gulliver his writing machine, with which "the most ignorant person (…) might write books in philosophy, poetry, politics, laws, mathematics, and theology, without the least assistance from genius or study" (Swift, 1963/1726, p. 188). Now, it seems, we are faced with this reality. GPT does not really "know" anything about language—or the world—rather, it predicts sequences of characters (and then larger chunks of language) based on their statistical likelihood of occurrence within the corpora on which it has been trained (see Bhatia, 2023, for an online demonstration). When paired with chatbots, GPT systems can interact with users in a conversational way, allowing follow-up questions and challenges from the user—thus creating the impression of having some intelligence. But GPT systems' impersonation is based on no person, and interaction produces no intersubjectivity. Whereas we generally come to know people the more we interact with them, ChatGPT is unknowable. As my colleague Mark Kaiser (personal communication, October 21, 2023) has mused, ChatGPT might be sufficient if language use were only transactional in nature. But if our purpose is a deeper understanding of others and ourselves, ChatGPT is of no help whatsoever. Although ChatGPT will sometimes identify sources (especially when asked), it does not reliably do so, and moreover its reported sources are not always reliable. Citing ChatGPT as a source is problematic because as it aggregates information, ChatGPT may be essentially averaging information across various sources, removing valuable contextual connections that might be essential for accurate interpretation of the information. ChatGPT's inability to "know" if a source is reliable or not is obviously a major issue, and even worse is its tendency to fabricate or "hallucinate" (when it cannot locate enough information) in order to create a convincing response. Clearly the onus to assess reliability remains with the user. Similarly, while ChatGPT can predict the probability of linguistic items and strings co-occurring in different contexts, it "knows" nothing about content matter, much less theoretical frameworks or disciplinary practices (even though it may appear to, based on its ability to detect and reproduce patterned relationships in data). For that reason, GPT systems also cannot "know" when information is malevolent, offensive, or biased.4 Large language models require extensive filtering by humans to override racism, sexism, homophobia, and other offensive ideological qualities of source texts. But the fact that these human filter generators are hired for pitiful wages in developing countries raises further ethical issues (Perrigo, 2023). [Chatbots'] deepest flaw is the absence of the most critical capacity of any intelligence: to say not only what is the case, what was the case and what will be the case—that's description and prediction—but also what is not the case and what could and could not be the case. Those are the ingredients of explanation, the mark of true intelligence (… .) In short, [they] are constitutionally unable to balance creativity with constraint. They either overgenerate (producing both truths and falsehoods, endorsing ethical and unethical decisions alike) or undergenerate (exhibiting noncommitment to any decisions and indifference to consequences). (para. 7, p. 17) Regarding the epistemic frame, Cope and Kalantzis prompted feedback in relation to eight "knowledge processes" of their epistemological theory of learning, as well as two measures of academic communication (related to "expression" and "genre"), yielding 10 cycles of analysis of student work and 10 epistemically focused sets of feedback. With respect to the empirical frame, the idea is not to ask anything factual of GPT systems (since the results are not reliable) but to feed GPT systems a set of relevant (and already verified) facts to work with. Similarly, calibrating the ontological frame essentially involves feeding the system "widely agreed definitions and taxonomically well-formed schemas that define the domain" (Peters et al., 2023, p. 16). Cope and Kalantzis commented that despite the harm that GPT systems can inflict on education, they see value in the "neatly formed narrative responses" (Peters et al., 2023, p. 16) they offer—if they can be intelligently constrained by users. Of key importance to language educators is that although ChatGPT currently supports some 95 languages, it has been trained predominantly with English-language texts, so its capabilities are most extensive in English.5 Nevertheless, on the level of grammatical accuracy, ChatGPT seems to excel in the most widely taught languages. Where its performance is more likely to vary is in the accuracy of content and its ability to deal with genre forms that differ from culture to culture. Languages not written in Roman script are generally disadvantaged, and languages like Urdu (written in Nastaliq calligraphic script, whose letter forms vary depending on the following letter, and which did not develop a typographic system until the late 20th century) present a special challenge, as they have relatively little digital content available (Irfan, 2023). This means that ChatGPT will have little to train on, will be conspicuously lacking in historical texts, and will thus perform comparatively poorly in such languages. Furthermore, because GPT systems are trained on monolingual models, they do not deal well with code switching or translanguaging. ChatGPT has been trained on information available online up until September 2021 and, as of this writing, only ChatGPT Enterprise subscribers and Bing GPT-4 users can access the Internet in real time. Moreover, paywalled items (including most academic journals) cannot be reliably accessed, especially if users do not have premium subscriptions. This hampers ChatGPT's ability to provide the most up-to-date and accurate information. Furthermore, whatever bias or misinformation that exists in the training corpus will also exist in ChatGPT's output, although OpenAI is making efforts to reduce the risk of toxic speech or misinformation. Despite these limitations, generative AI tools are now a fact of life, and language educators have a responsibility to think through the linguistic, social, and ethical issues related to AI—and other forms of technological mediation—with their students. The rest of this article will propose a way forward. The Internet, social media, mobile technology, and generative AI applications all ostensibly foster the development of language learners' agency and autonomy. But if they are used as the central pillar of learning, or if they allow students to sidestep the real work of developing proficiency in the new language, they risk encouraging a dependency on the technology itself and ultimately reducing students' learning to little more than manipulation of external resources to create an illusion of linguistic ability. What I believe we want to encourage is a genuine agency and autonomy that is born of an integration of language, culture, mind, and body. An integration that allows connection with other people, creative expression of new identities in multiple modalities, and critical remove from monolingual and monocultural perspectives. It is important to recognize that technology can facilitate such integration; what is needed is discernment in how to creatively employ technological resources toward that end. Arguably, and beyond language learning per se, another purpose of using technology in language education is to explore the very ways that technology mediates language use, communication, cultural expression, and social meaning. That is, to adopt a semiotic approach that raises learners' consciousness of the myriad mediations that are part and parcel of our daily communication, and especially our online communication, where the making and taking of meaning is influenced by multiple layers of mediation. Such an approach is grounded in the idea that people do not use technologies as much as they interact with and through them (Shaffer & Clinton, 2006, p. 289). For example, a poem will be experienced differently and involve different activities when it is memorized and recited, or read aloud expressively from a book, or read silently, or searched for on the Internet (Erstad & Wertsch, 2008). When we write a message in an email, in a tweet, in an online forum, on a post-it note, on a piece of fine stationary, on a wall, or on a tombstone, each medium and its associated social conventions in a particular culture—combined with our notion of the intended readership—shape our linguistic choices. As Jones and Hafner (2012) put it, "The process of mediation (…) is not just a matter of media controlling people or people controlling media. It is a matter of the tension between what technology wants us to do and what we want to do with it, between the limitations it imposes on us and our ability to get around these limitations by 'hacking' it" (p. 101; emphasis in original). Technology, like art, offers a means to make the familiar unfamiliar, to reframe, rethink, and rework our conceptions of language, communication, and social conventions. If we as language educators are technologists in a broad sense, our students need to be as well. A number of theoretical frameworks already associated with second language acquisition (SLA) can be helpful in framing an approach focused on mediation. Sociocultural theory, grounded in the work of Lev Vygotsky and other Soviet psychologists, considers human development, cognition, and activity within cultural, institutional, and historical contexts. It posits that cognition is socially mediated and distributed, rather than isolated within people's heads. Mental or physical activity is mediated by cultural artifacts serving as physical or symbolic tools, which are often modified as they are passed from one generation to the next (Lantolf, 2000). Activity produces and reproduces culture, and culture provides the resources (material and psychological mediation) that make new forms of activity (i.e., innovation) possible, generating a perpetual cycle. One of the interesting characteristics of mediation is its participation in human agency—it does not just facilitate mental functions, it transforms them. For example, when we communicate something, the channel