1. Introduction

What if digitalization is not a specific condition required for digital humanities but "a general condition of language and life" (508) that will prevail in the future of humanities research and education? This is what Matthew Kirschenbaum and Rita Raley alert us to in their essay "AI and the University as a Service" published in the most recent issue of PMLA.1 As they suggest, the computational approaches that characterized the journal's previous Theories and Methodologies section on digital humanities in January 2020 (135.1)2 "are increasingly regarded as altogether ordinary" (508), while the imminent structural changes in the humanities with the advent of LLMs, such as ChatGPT in November 2022, look unprecedented.

I would like to respond to their call to rethink the future of humanities research and education in the era of large language models and generative Al from my position as an English PhD who studied writing (and thinking!) in English through both undergraduate and graduate education in Korea and the United States. As someone who first learned English in Korea and later in the United States, I will focus on the digitalization of the university and language, especially from a global perspective. By the digitalization of the university, I mean mostly the courses recorded, uploaded, and shared on digital platforms such as Coursera, edX, Udacity, etc. In the case of the digitalization of language, the meaning of the word "digitalization" refers to not only the proliferation of online resources for language learning but also, and more importantly, to the reformulation of language into machine-readable units through tokenization, word embedding, etc., and the production of sentences by prediction.3 This is the language pervading LLMs such as ChatGPT, Gemini, and HyperCLOVA. My positionality in interpreting these kinds of digitalization is global, involving multilingual, multicultural contexts.

My focus is mostly on writing pedagogy, especially because what distinguishes language education in the university from that of the private sector is the academic goal shared by all disciplines in the humanities to develop the capacity for critical thinking and writing. Writing in one's native language requires a lot training, and L2 writing, or writing in a language other than one's native language, requires not only the acquisition of language skills but also the reformulation of the thinking process in terms of culture. With the digitalization of language, university education calls for new pedagogical adaptation. As Kirschenbaum and Raley (2024) write, "Language in the form of executable code, machine-readable data, and humanreadable text subject to tokenization for word embeddings is not merely product but also managerial apparatus (yes, dispositif) that produces humans as subjects of, and subjects to, the extractive economies that are the new scene of writing" (512).

2. The Global Digitalization of the University and Language: Possibilities and Challenges

The pandemic and its enforcement of online lectures accelerated the digitalization of the university as a service, which is now at hand with the rise of shared online platforms that store and stream recorded lectures. Many universities in Korea as well as those abroad have created online platforms to share their lectures within their university communities and sometimes outside of them. Besides universal online platforms that stream and archive courses such as Coursera, edX, Udacity, Stanford Online, MIT OpenCourseWare (OCW), Harvard Online Learning, there are also Korean versions such as SNU OpenCourseWare (OCW), Korea University OpenCourseWare (KUOCW), SKKU Online, to name a few. While these online platforms transform the originally in-person, communicative, communal forms of the class setting into repetitive, receptive, and isolating forms, they also make higher education accessible beyond the barriers of physical distance. For my graduate seminar on academic writing in English last semester, for example, I used the recorded workshops from UCLA graduate writing center (https://gwc.gsrc.ucla.edu/workshop-videos) to teach about the academic norms of writing circulating in Anglophone academia in the form of flipped learning. Students watched the recorded workshops in advance of class and came to class with drafts to peer-review and revise. This kind of global digitalization of university lectures enhances international accessibility to higher education and makes the academia inclusive of diverse cultures and nations.

On the other hand, the digitalization of language at first glance seems to refer to the same trend of global accessibility to language resources that are abundant in vernacular environments available on the internet, especially with the growing use of LLMs for language learning in the 2020s. This changing accessibility to language learning, however, is underpinned by the digitalization of language in another sense: the computational processing of language into machine-readable units, known as natural language processing (NLP). NLP is a field of computer science focused on enabling computers to analyze and understand human language. It involves technologies such as text preprocessing (e.g. tokenization, cleaning, normalization, stemming, padding) and linguistic representations at lexical, syntactic, and semantic levels (e.g. part-of-speech tagging, topic modeling). These NLP methods are foundational to language models (LMs), which are systems trained to predict sequences of words. Language models have evolved from N-gram models-a sequence of n-words-and word-vectors, then to Recurrent Neural Networks (RNN), which were capable of using "memory" of previous inputs for sequence prediction and text classification. Most recently, Large Language Models (LLMs) of Transformer architecture like ChatGPT, released in November 2022, use self-attention mechanism to perform more complex tasks like summarization and question-answering.

While the digitalization of language seems helpful to students who learn foreign languages, especially English, across the globe, it poses some challenges. First, is it justifiable to train students on the language of LLM outputs based on statistical probability rather than communicative intent, which, according to Knapp and Michaels (1982) in their essay "Against Theory,"4 constitutes meaning in human language? Given that an LLM is "a stochastic parrot" producing words "without any reference to meaning" (Bender et al. 2021, 617), it is questionable whether an L2 learner who studies the language solely figured as probable patterns could attain, or should attain, the same capacity of contextualizing knowledge and communicative interaction. As the authors of "On the Dangers of Stochastic Parrots" write, "languages are systems of signs, i.e. pairings of form and meaning. But the training data for LMs is only form; they do not have access to meaning" (Bender et al. 2021, 615). LM outputs, which consist of "linguistic forms" produced by statistical probability without any reference to meaning, may trick us into believing that they constitute real language and demand human understanding. Bender et al. (2021) caution that an LM is "a stochastic parrot" that spreads out a string of words without meanings, and humans who try to read and interpret it may be "led down to the garden path" (617, 616). A more serious concern is that models only trained on LLM outputs could collapse,5 suggesting that machine-produced texts are not good enough to be the source of language learning in the case of language models themselves, which may or may not apply to human brains as well. It has been discovered that machine-produced texts show different patterns of probability among word-distributions.6 Will humans linguistic capacity trained on inputs produced by LLMs collapse as well?

Secondly, the vast size of English datasets and the limited access to texts in other languages in most big-tech LLMs may reinforce Anglocentrism and the English monopoly on datasets rather than promote multilingualism in higher education. I second Eduardo Ledesma's argument that Al and LLMs could exacerbate "the existing marginalization of languages 'other than English, by buttressing English dominance and further eroding linguistic diversity in higher education and beyond" (534).7 Although his review of the effects of English monolingualism through Al and LLMs mostly relates to the Global South, this issue is not significantly different in East Asia and for Asian languages. When the "HyperCLOVA X Technical Report" was first released by NAVER in April 2024,8 people expected it to have good multilingual skills as said in the report,9 but many found HyperCLOVA X's Korean performance very disappointing. When I used it, it produced sentences that stop with subwords due to the word limit, and its style, tone, and diction were too casual. When I asked Clova how it learned Korean, it said it only read casual or informative writings such as blog posts or news articles, not the academic articles available in DBpia. The limitations of Korean LLMs were already confirmed by Joo et al. in 2023,10 When LLaMA and Polyglot become available as open source, Korean models were developed using pre-trained LLMs as foundation models: Polyglot-ko modeled on Polyglot, KoAlpaca-Polyglot and KULLM modeled on Polyglot-ko and KoAlpaca modeled on LLaMa, and KoVicuna modeled on Vicuna.11 However, the size of their Korean parameters as well as that of HyperCLOVA X, however, is far smaller than the size of English parameters available for GPT 40.12 Users of GPT may all have experienced that its answers about Korean culture and history are often inaccurate or fabricated. To resist this "new form of coloniality" (which Ledesma uses to refer to this problem of English dominance over other languages 354), it is important for each nation to develop a vernacular language model that is attuned to its vernacular culture and history, but the kind and size of current datasets used by Korean LLMs for these decolonizing tasks are far from enough.

3. Moving Forward Writing with Al, or APT (AI Personal Training)

What should we do to move forward in the face of this global digitalization of the university and language driven by Al? Echoing Elkins (2024), I would say that the question is not how or whether we "serve" this new apparatus called Al that restructures higher education but about how to "advocat[e] for an Al that would serve us" (546).13 Responding authors to Kirschenbaum and Raley's thought-provoking essay reflect on the question of language and offer various approaches and suggestions for "AI and the University as a Service." As Vadde (2024) tells us, human-centered writing is no longer the default when more and more students are tempted to use generative Al to complete their writing assignments and as automatic text-generation proliferate in daily life. Vadde adds that we humanists, as called upon to help establish the university's task force and guidelines for writing, are reminded that "our domain is language" (557). Martin (2024) presents "Command Lines for the Humanities," calling for collaboration, communication, method, and collective approaches to datasets and data reviewing.14 Perlow (2024) urges us to restore "aesthetic criticism" to define the contours of "value." Similarly, Laquintano and Vee (2024) highlight the potential "linguistic imperialism" and the validity of the "human voice as a marker of value," inviting us to reflect on the history of discrimination and hierarchies to better reshape "writing voice" of or between humans and machines. Their emphasis on the value-judgements that define what should be counted as "good" output, good writing, good analysis is absolutely needed in the contemporary textual vista where "textpocalypse" is near.15 As I mentioned earlier, Eduardo Ledesma invites non-English studies scholars to engage in shaping critical Al studies from diverse perspectives in order to decolonize the Anglocentrism that pervades not only LLMs and literary studies but also other disciplines, implying that Al may be "another cultural form altogether" (538). Huang (2024) discusses writing as a "therapeutic practice" drawing upon Wittgenstein's idea of language as an active, practical utterance in the time of crisis.16 Dugan and Smith (2024) refuses to debate which reading is better between humanist reading and computational reading and instead welcomes "the cultivation of the lag that this pedagogy introduces and the critique that it hopefully inspires in its wake: a hiatus or suspension" (521).

Hoping on this train of thoughts and suggestions for Al's role in higher education, I will suggest what I think about the challenges posed by the global digitalization of the university and language, which I have identified as the problem of the future higher education anticipated in Kirschenbaum and Raley's caricature. The second question about coloniality requires an ongoing reformulation of policies and cooperation between the national governments and LLM developers. In the United States, many current literary archives storing texts protected by intellectual property are closed to public access in the name of fair use, as seen in Warhol v. Goldsmith, as Perlow (2024) laments in his response essay because enough training in large data is necessary for the consecutive development of Al models.17 Access to library resources and news archives in Korea, however, is relatively flexible and open because most textual data in Korea are funded by the government and are open to the public or members of academic institutions for fair use. The Korean Association for Digital Humanities (KADH) has a webpage for data available in Korean or other languages: https://www.kadh.org/category/resources/data/. Besides those tailored datasets, current Korean textual data available in ancient and contemporary Korean include the following:18

Of course, not all Korean data are available or free of copyright laws, and I am just pointing out the relative generosity of public accessibility in South Korea compared to the aggressive copyright laws in the States. Korean LLMs such as HyperCLOVA X should leverage this large amount of original Korean textual data widely accessible and available as well as academic language data available in DBpia (https://www.dbpia.co.kr/), just as GPT uses ScholarAl and other plug-in services (though some of them are to be paid additionally) for that purpose. Another reason that Korean LLMs are less developed than English LLMs may probably be the more complex morphology required for tokenization, including the use of particles (josa) that classifies word phrases into grammatical categories such as noun phrases, verb phrases, etc. More complex NLP tools such as KIWI (Korean Intelligent Word Identifier) should be used. Lee's (2023) article "Kiwi: Developing a Korean Morphological Analyzer Based on Statistical Language Models and Skip-Bigram," which examines how to resolve ambiguities in morphological analysis, has recorded more than 700 views, showing high interest in Korean linguistics and textual data.19

As for the first challenge about whether it is justifiable to train students using the language of LLM outputs based on statistical probability rather than communicative intent, I would argue we humans are born to read and the question is not about justification but about how to develop new hermeneutics that will help us navigate this new landscape of reading and writing called "textpocalyse," where texts produced by humans and/or machines without intent proliferate and blend together.20 Citing Kittler's words on "hermeneutics-a pilot's understanding of signals, whether they stem from gods, machines, or sources of noise," Slater (2023) argues that it is necessary to develop the capacity to interpret, or to make sense of, multiple forms of signals that are not explicitly generated by human intention (quoted from Kittler. "Signal-to-Noise Ratio" 177).21 We humans "read [...] literally" anyway when we see a text, but this reading is different from "interpretation" an activity that gauges the meaning of the text.22 We should distinguish between mechanical reading that machines can do as well and subjective interpretation that only belongs to humans, as Shoemaker (2023) alerts us to in his essay "Machines, Reading." All in all, in the age of textpocalypse and with LLMs being commercialized for language learning business, humans will read anyway and they will keep interpreting whether or not the text they see is produced by humans with intentions or machines without intentions. By reading it and interpreting it anyhow, we humans can transform the string of linguistic forms absent of reference into the language pairing forms and meanings. This act of interpretation out of nothing, however, was also what humanists have done for decades, as post-structuralists like Roland Barthes free interpretation from the author's intention in the "Death of the Author."23 Reading and interpreting machine-written texts as learning materials would work as long as humans are aware of it and know how to identify "disinformation" (Slater 2023, 1258), because, unlike machines that process linguistic patterns without a reference, humans can dissect the language and re-pair its forms with meanings.24

Equipped with the capacity of interpretation, we humanist scholars and students can read and write with Al, as if we have Al Personal Training (APT) that helps us with writing, rather than either totally disregarding or relying on it. Much has been said about how to read with Al by scholars who developed diverse methods of distant reading, and their conclusion is to keep both close-reading and distant-reading as main methods and not to lose our humanistic perspectives that are attuned to localized contexts. But not much has been said about how to write with Al in academia yet, especially in L2 context. Though not many academic articles are available on this topic, some have reviewed and recommended ways of collaborating with Al in academic writing.25 In Korea, this new task of how to write with AI, especially with LLMs on the level of higher education, is more actively discussed in publishing industries that promote practical approaches. Kim (2023) provides a variety of approaches and prompts that could be used to proofread and correct grammatical errors.26 Last semester when I taught a graduate writing seminar, for example, I let my students use ChatGPT to correct their grammar, and their learning was more effective when their English was less fluent. GPT was mostly correct in its revisions except in a few cases when a nuanced understanding of the context was required, which attests to the fact that it still requires human attention and proofreading. My students confirmed that GPT 3.5, which was free, judged ungrammatical sentences as correct and did not correct anything, while GPT 4.0, which required subscription fees, made multiple corrections on the same text.

The partnership with Al and LLMs to figure out probable distributions of words for precise diction predates the advent of GPT. Language learners in academic settings and university writing centers have used large corpora like the Corpus of Contemporary American English (COCA), the British National Corpus (BNC), or WordNeighbors and electronic software by checking n-strings of words and their probable collocations. Mobile-Assisted Language Learning (MALL), which refers to the type of language learning that uses laptops, tablets, and phones to access online resources, was initially recognized in the late 1990s and has been actively discussed in contemporary education studies. ESL students in American universities showed great development in L2 writing sills by using grammatical applications, especially in their use of verbs and prepositions (Li and Hegelheimer 2013).27 Kwon's (2024) study of academic trends in MALL studies in KCI articles, which used LDA topic modeling and word cloud to read 247 LACI articles on MALL, show that the word "university" or "college" appears significantly more (eighty-six, sixty-eight times respectively) than words of secondary or elementary or pre-elementary schools (twenty-one times respectively), indicating that MALL is more used in higher education.28 Taecharungroj (2023) and Sallam (2023) even encourage the use of GPT for the whole process of writing including brain storming, researching, editing, and proofreading, arguing that it is tremendously useful. Yet a more nuanced sense of academic integrity29 is absolutely required, if the use of LLMs exceeds beyond subtle grammatical corrections.

In this context, I propose APT, by which I mean "Al Personal Training," a practice involving humans partnering with Al to train, or improve, their writing.30 My suggestions regarding APT include, first, the practical realm of developing writing pedagogy, which encompasses multiple steps of thinking or brainstorming and improving L2 writing skills, as outlined above. People now find LLMs such as GPT and Gemini very helpful for brainstorming ideas, finding research topics, and conducting literature reviews, though the summaries they provide are not always accurate. While it may be acceptable to get help from APT for brainstorming and editing, however, I believe that drafting should solely be done by humans, relying on their own capacities.

Secondly, and more importantly, while it is crucial to maintain human agency in writing, in the era of APT, we humans should embrace co-evolution with Al in the realm of language. When the languages humans read in contemporary environments are already influenced by chatbots, machine-translated subtitles on Zoom, and generative text models, human and machine languages will evolve by influencing one another. The co-evolution between different species, which Haraway (2003) has introduced when the binary between nature and culture no longer holds in the co-habitable space shared by humans and dogs,31 is applicable to humans' relationships with AL. If we humans dare to believe in our hermeneutic capacity, we may venture into this new arena where human and machine-generated languages co-evolve with each other, shaping our co-constitutive relationships and languages.

4. Coda

The meaning of the word "Prompt" in the subtitle of my essay is twofold. Firstly, it is an ambiguously belated yet prompt response to Kirschenbaum and Raley's essay, as my ideas for this essay were brewing but were never put into words because of my busy schedule when I was asked to write a quick response to their essay in early March with a very short deadline. I hope my response, though belated, is prompt enough to be printed in the same year as their essay's publication. Secondly, I hope my essay becomes another prompt for future responses from the non-Anglophone worlds, reflecting on the possibilities and challenges that the LLMs and Al bring to other languages and universities spoken, written, or located outside Anglophone academia. Unlike Kirschenbaum and and Raley's concluding remarks"Al is not the prompt, it is the punctuation"-I would say, "Al is the prompt," especially for APT (AI Personal Training), which will dominate or become common in future higher education.32 Perhaps, what Kazuo Ishiguro (2021) described in his 2021 novel Klara and the Sun is becoming or has already become real. Don't we all have an "Artificial Friend" like Klara, though in the form of a disembodied voice like ChatGPT, installed on our cellphones? Alas, not everyone but only those who can afford to pay the subscription will have APT.