AI and epistemology in applied linguistics

Epistemology is an inherently fuzzy concept. Broadly conceived, epistemology theorizes how we construct and make sense of our knowledge. Epistemology thus governs the values by which we validate and justify knowledge and the conventions we follow and reconstruct within our wider knowledge-making community and communities of practice. Owing to its diverse subfields and varied theoretical frameworks and methodological approaches (Kuteeva & Andersson, Reference Kuteeva and Andersson2024), applied linguistics is arguably not governed by any single epistemology. Rather, it is host to a plurality of epistemologies (Dewaele, Reference Dewaele, Wright, Harvey and Simpson2019; Pennycook, Reference Pennycook2018) that shape how we come to know things in our field. For example, research in applied linguistics often relies on local, context-dependent knowledge (Kuteeva & Andersson, Reference Kuteeva and Andersson2024). Such knowledge is mediated by language, with resultant knowledge cultures varying across languages, modalities, cultures, time, and contexts of production (Curry, Reference Curry2024; Pérez-Paredes & Curry, Reference Pérez-Paredes and Curry2024; Rymes et al., Reference Rymes, Lee and Negus2024). This breadth renders applied linguistics a heavily reflexive field that sees strength in reflection, positionality, and criticality (Consoli & Ganassin, Reference Consoli, Ganassin, Consoli and Ganassin2023), and one that moves iteratively between epistemological perspectives.

In research on GenAI, in particular, there is potential for epistemological blurring between AI and applied linguistics. For example, GenAI, with its ability to process vast volumes of data and produce outputs in various modalities, can be seen as a tool for knowledge generation (Creely, Reference Creely2024); it should be noted, however, that this term does not correspond to “knowledge generation” in the likes of Wei et al. (Reference Wei, Chen, Yu, Fei, Liu, Al-Onaizan, Bansal and Chen2024) where it is applies specifically to the validity of LLMs. The nature of such knowledge is a key concern, as what it means to know something for AI versus humans has long been a matter of discussion (Pitrat, Reference Pitrat1995; Schank, Reference Schank1980; Wang, Reference Wang2007). For GenAI, knowing is a product of processes built on probabilities and pattern recognition, derived from training data (Cope & Kalantzis, Reference Cope and Kalantzis2024). Such a conceptualization of knowing may have little to do with the knowledge-making processes attributed to the human researcher in applied linguistics. Human capacity for creativity, the role of human accountability in research, and knowledge-making as a reflexive, flexible, and non-linear process are among the ways in which such differences are realized. As such, we may wonder whether a machine can indeed know anything at all. In our view, we can understand AI approach(es) to epistemology or knowing – broadly conceived – as its way of creating knowledge, though we may wonder whether this knowledge belongs to the AI or to the sources on which its LLM is trained. Ultimately, we do not see AI-driven processes as the same as those involved in human knowledge-making for applied linguistics – at least for now.

For AI, these processes of knowledge-making can be difficult and, at times impossible, to access. Consequently, GenAI tools and their LLMs have been found to lack transparency and have been likened to a conceptual black box (Curry et al., Reference Curry, Baker and Brookes2024; Wiener, Reference Wiener1961). It is interesting to note that Weiner’s attribution of a black box quality to machines pertained mainly to the unknowns within machine systems and their processes that needed to be understood or clarified through system identification. In the context of AI, similar notions of the unknowns within machines remain relevant as, for some, the metaphorical treatment of GenAI tools as representing a black box is their means of critiquing the acceptability of AI in academic work. This critique of AI’s unknowns has led to a focus on so-called explainable AI, which centers on delivering greater transparency in all facets of AI development (Xu et al., Reference Xu, Uszkoreit, Du, Fan, Zhao and Zhu2019). Yet, despite such advances, issues remain, as the data on which AI is trained are inevitably limited – there may not be enough data available to perfect the models, should perfection be possible. This problem is exacerbated by AI generated text starting to populate the web pages from which data for training models are drawn, meaning that GenAI may poison the source of data it draws from in such a way that it reinforces its own limitations and biases.

GenAI models thus present a significant epistemological challenge as they reflect biases that can homogenize language, propagate limited cultural narratives (Choi, Reference Choi2022; Putland et al., Reference Putland, Chikodzore-Paterson and Brookes2023), and limit our capacity to move beyond normative worldviews represented in the training data. As such, the epistemology of AI, if such a thing exists, lacks a conceptual understanding of the world or human-like sentience (Rodríguez, Reference Rodríguez2023). Thus, AI’s approach to knowing contrasts sharply with human knowledge in applied linguistics, which is rooted in criticality, lived experience, cultural context, and interaction with the physical world (Creely, Reference Creely2024; Tang & Cooper, Reference Tang and Cooper2024). Recognition of such differences has begun to have a ripple effect in AI research, with some shifting from language models toward world models (e.g. Xiang et al., Reference Xiang, Tao, Gu, Shu, Wang, Yang and Hu2023) that are designed to reflect the complex models of reality on which humans draw. These models do not see language as divisible from context and offer wider world contextualization of data, as opposed to solely a linguistic contextualization of data.

As applied linguistics embodies a plurality of epistemologies, we should query, as Kuteeva and Andersson (Reference Kuteeva and Andersson2024) do, how AI-assisted tools can respond to the differing epistemological stances employed within and across applied linguistics studies. Likewise, in recognizing the importance of localized and situated epistemologies in studies across cultures and languages, we might consider what the limitations of AI’s approach to knowledge-building means for applied linguistics. For example, GenAI may not capture the nuances of human language (Sardinha, Reference Sardinha2024) and has been found to fabricate data, a process AI researchers call hallucination (e.g. Baker & Kanade, Reference Baker and Kanade2000; Curry et al., Reference Curry, Baker and Brookes2024). This was initially seen as an asset, as AI could hallucinate and enhance the quality of images and supplement missing detail in them, for example. Yet, as AI has gained more prominence in fields such as applied linguistics, hallucinations have come to be seen more negatively as they have the potential to create content that reproduces biases and eliminates diverse cultural and linguistic contexts (Brandt & Hazel, Reference Brandt and Hazel2024; Choi, Reference Choi2022; Putland et al., Reference Putland, Chikodzore-Paterson and Brookes2023). Likewise, fears of epistemicide – the destruction of knowledge systems that do not align to the dominant paradigm – emerge owing to GenAI’s inclination to standardize and simplify knowledge (Pragya, Reference Pragya2024). Many of these issues with GenAI are antithetical to wider social justice movements that shape contemporary applied linguistics (e.g., Badwan, Reference Badwan2021). Moreover, given that language is seen in applied linguistics as a means of mediating the creation of new knowledge, a question arises regarding the extent to which GenAI can support this creative practice if its operationalization requires the reproduction of existing, learned patterns (Kuteeva & Andersson, Reference Kuteeva and Andersson2024). Thus, with reference to epistemologies, we can begin to see the alignment problem (i.e. the potential (mis)alignment of research and practice in AI and applied linguistics) as an active issue.

To begin to address the problem, there is a need for critical understanding of GenAI that involves evaluating its capabilities and limitations and safeguarding the integrity of culturally rich applied linguistics research (Creely, Reference Creely2024). To support this effort, we must seek means to ensure that any knowledge GenAI appears to produce can be interrogated within the heterogeneous epistemological paradigm of applied linguistics (Pennycook, Reference Pennycook2018). This should include a focus on preserving researchers’ abilities to critically select, analyse, and interpret information as well as their ability to challenge normative views constructed through discourse. As researchers, we implicitly accept the accountability that comes with conducting research and the importance of sustained, critical, and reflective practices that underpin the research process and the development of a discipline as well as a wider community of practice. This accountability will always remain with the researcher, and any use of AI tools in research will need to be considered in line with the potential impact of AI tools both on knowledge as a product and on knowledge-making as a process.

To address the alignment problem, we must also balance GenAI’s capabilities with human agency and creativity. We must augment rather than replace human skills (e.g. using GenAI for routine and automatable tasks) and develop criteria for evaluating the quality of knowledge produced by GenAI. This will be a challenge, given the growing body of work revealing that AI-conducted research does not necessarily outperform research conducted by humans (e.g., Vaccaro et al., Reference Vaccaro, Almaatouq and Malone2024). This challenge is echoed by the evident epistemological divide between applied linguistics and GenAI. To date, few studies at this interface directly engage with notions of epistemology. Those that do have consistently problematized the capacity of GenAI to work within the epistemologies that govern applied linguistics. Research in applied linguistics is built on the central and indispensable role of human culture in knowledge construction (Tang & Cooper, Reference Tang and Cooper2024). This is why the alignment problem is so important to applied linguistics. As we move forward, it is imperative that our approach to knowledge construction in applied linguistics remains inseparable from our interaction with material objects in a world increasingly influenced by AI. AI as a tool must be put in service to epistemologies in applied linguistics – the epistemologies must not change simply to fit the affordances of AI.

AI and ontology in applied linguistics

The alignment problem stretches through from epistemology to ontology. Ontology is concerned with the nature of being and existence (Hall & Wicaksono, Reference Hall, Wicaksono, Hall and Wicaksono2020). Engaging in ontological work as part of the applied linguistics research process essentially involves asking what kinds of things exist, how they relate to each other, and whether these things exist independently of human minds and language (Hall & Wicaksono, Reference Hall, Wicaksono, Hall and Wicaksono2020). Ontology is foundational to much of what applied linguists study, as language is typically seen as the primary means through which immaterial entities and social institutions are socially constructed. In applied linguistics, ontologies are not fixed but are actively negotiated and performed through social practice (Demuro & Gurney, Reference Demuro and Gurney2021; Dewaele, Reference Dewaele, Wright, Harvey and Simpson2019). The plurality of ontologies in applied linguistics mirrors that of epistemologies – an unsurprising linkage, given that ontologies can be shaped by epistemologies and vice versa (Demuro & Gurney, Reference Demuro and Gurney2021; Hall & Wicaksono, Reference Hall, Wicaksono, Hall and Wicaksono2020). Consequently, in applied linguistics, language can be seen as an object of study (e.g., Curry & Pérez-Paredes, Reference Curry and Pérez-Paredes2021) or a practice (e.g., León et al., Reference León, Lemmi, Sedlacek, Ortiz and Feldman2024), inter alia, depending on the focus of any given analysis. From a social justice perspective, ontologies are shaped by researchers, language users, and disciplinary practices (Schalley, Reference Schalley2019). In this way, ontologies can be used to challenge monolithic and dominant views of language and promote awareness of the diverse ways in which language is understood and used (Hall & Wicaksono, Reference Hall, Wicaksono, Hall and Wicaksono2020).

In AI research, whether it be symbolic or statistical, ontology takes a very different form as there is a strong tendency to select and fix ontologies. Researchers seek to control ontologies as a way of simplifying the problems facing AI system developers. As such, AI ontologies are often designed for specific domains or purposes and aim to represent reality in a formal, structured, machine-readable format (Machado et al., Reference Machado, Almeida and Souza2020). A good example of this was the EAGLES initiative, which sought to standardize part-of-speech tag sets across languages (Leech & Wilson, Reference Leech and Wilson1994). This practice of setting ontologies also underlies large-scale AI resources such as WordNet (Miller et al., Reference Miller, Beckwith, Fellbaum and Gross1990). Yet, fixing ontologies is largely a convenience – making one set of decisions to pare down the complexity of the legitimate choices that, in fact, exist (see McEnery & Brezina, Reference McEnery and Brezina2022). Thus, although AI draws on the same broad philosophical tradition and notion of being as applied linguistics, ontology, when localized in AI research, looks very different. A key goal of ontologies in AI is to achieve semantic interoperability, thus allowing different systems to exchange and understand data while providing a common vocabulary and rules for structuring data (Machado et al., Reference Machado, Almeida and Souza2020; Spyns & De Bo, Reference Spyns and De Bo2004). These ontologies are typically organized, representing knowledge through hierarchical structures or taxonomies (Iliadis, Reference Iliadis2019; Schalley, Reference Schalley2019) designed to support practical tasks, like reasoning, inference, and knowledge retrieval (Machado et al., Reference Machado, Almeida and Souza2020; Spyns & De Bo, Reference Spyns and De Bo2004). They are also data-driven, that is, developed to organize and label data (Iliadis, Reference Iliadis2019). Reflecting upon both this and the previous paragraph, it becomes clear that ontology and the notion of being in applied linguistics and AI represent a further dimension of the alignment problem.

Applied linguists use ontology as a critical lens to examine how language and its use are shaped by social, cultural, and political forces. AI researchers use ontologies to create formal representations of knowledge for machine processing. Thus, while applied linguistics prioritizes philosophical understandings of ontology, emphasizing the social and cultural dimensions of language, AI adopts a practical, engineering approach, using ontologies as tools to model and manage information for specific applications in a way that typically explicitly simplifies the complexity of the system described (Hall & Wicaksono, Reference Hall, Wicaksono, Hall and Wicaksono2020) and promotes interoperability (Iliadis, Reference Iliadis2019). Unlike applied linguistics research, wherein the complexity of ontologies is not only accepted but seen as a core strength, in AI research, simplification in search of standardization is a key research practice. Part of the simplification is usually normative – the data on which the systems are developed and the categories chosen are oriented toward the typical.

As with epistemology, applied linguists need to be aware of misalignment between the ontological goals of applied linguistics and AI. While the appeal of being able to apply AI ontologies to a research problem is clearly great, this advantage should not be gained at the expense of distorting our view of the object of study. For example, AI language models can struggle with the nuances of human language, including idioms, cultural meanings, and expressions deeply embedded in cultural and historical contexts (Creely, Reference Creely2024). A good example is McEnery and Baker’s (Reference McEnery and Baker2017) attempt to use a semantic tagger, based on a semantic ontology, to explore the meaning of words involved in the representation of prostitutes. The ontology was developed for work on present-day English, yet the researchers were applying it to early-modern English. The attempt was deemed a failure on ontological grounds. The role of religion in the modern world had relegated religion to a single semantic field in the system used. However, in the early-modern period, religion was a frequent and complex feature of public discourse – around half the texts in the billion-word corpus used in the study concerned religion (McEnery & Baker, Reference McEnery and Baker2017, p. 163, see note 10). The descriptive poverty of the ontology rendered it unusable – it eliminated rather than illuminated nuance. Where this happens, researchers should abandon the tool, not the object of study.

Compared to the issue of epistemology, to-date an even narrower selection of studies in applied linguistics on the topic of AI or GenAI engages in depth with the notion of ontology. An initial means of advancing thinking in this area would thus be the inclusion of reflections on ontologies in research using GenAI. Those studies that do exist indicate that the worldview of applied linguistics and the limited linguistic models of GenAI may be incommensurable. While blending human input with machine learning in so-called “human in the loop” approaches may seem to address this issue, in practice such approaches have proved to be difficult to implement (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bibes-Bascarán and Fernández-Leal2023). Based on this, we argue that AI cannot replace the human analyst in the research process, especially if we continue to care about issues such as equality and fairness in socially situated research. Challenges such as the well-documented tendency of AI algorithms to perpetuate negative racial stereotypes (see Baker & Potts, Reference Baker and Potts2013) should cause any applied linguistics researcher to pause before using AI uncritically. It is essential, in determining the potential use of GenAI, that applied linguists take critical approaches to technological determinism and solutionism (McKnight & Shipp, Reference McKnight and Shipp2024), the so-called Eliza effect (whereby humans place implicit trust in machines, McKnight & Shipp, Reference McKnight and Shipp2024) and the wider digital divide that governs access to GenAI tools (Li, Reference Li2023). With these divergent perspectives on the role of ontology, we once again find ourselves searching for a means through which applied linguistics and AI can align.

As a potential move toward epistemological and ontological alignment, users may now, for example, try generating training data in which a linguistic analysis is encoded in annotation, to get a GenAI tool to learn how to undertake that analysis and then apply it to new data. The tool learns and uses the ontology provided by the user. While this may on occasion prove helpful, it is littered with problems as a research process. Firstly, users may not be able to generate sufficient training data to produce an analysis that is accurate enough to be useful. Secondly, in trying to do so, they may distort their research goals, investing a great deal of time on a Grail quest for an automated analysis rather than actually undertaking the research they set out to complete. This in part may be behind findings emerging in industry that AI assistants, for example, cost rather than save time (Monahan & Burlacu, Reference Monahan and Burlacu2024). Thirdly, the results will almost certainly exhibit errors – users must be aware of this and invest time in both quantifying and understanding those errors in an attempt to appreciate the distortions that those errors may produce. It is unlikely the errors will be smoothly distributed, and it is much more likely that they will be focused on the non-normative. Fourthly, the approach clearly makes daunting demands of the researcher in terms of permitting others to repeat their study – while the researcher may record the prompts and the version of the GenAI tool used, and, in some circumstances, the seed (a random number) used to initiate the generation process, it remains to be seen whether, in the long term, this will be sufficient to permit researchers to critically evaluate the findings of papers based on such research practices. Finally, it may be that the automated approach fails – in such a case, the danger of point two above is real as time is poured into trying to get an LLM to undertake an annotation that it is simply not capable of doing effectively. This final point is most likely to apply at the more subjective end of linguistic analyses or in areas in which an understanding of real-world social context is key, for example, in pragmatics. As a rough rule of thumb, in our experience, where a research question relates very closely to lexis, the likelihood of an LLM being trained to perform an analysis is good (e.g., Curry et al., Reference Curry, Baker and Brookes2024; Yu et al., Reference Yu, Li, Su and Fuoli2024). However, as we push through into a reliance on the non-lexical and non-linguistic, the likelihood of an LLM being of use reduces, and the greater the danger of point two above becomes. So, while LLMs may, in principle, assist with the ontological alignment problem, they are not a catch-all solution to it.

A further way of approaching the question of alignment could be through the issue of representation. Both applied linguistics and AI are interested in questions of representation, albeit from different perspectives. Applied linguistics is interested in how language represents concepts and how this representation varies across social space (e.g., speakers, cultures) and time. AI is interested in how knowledge can be formally represented for computational purposes. Yet, there is potential for dialogue here. As Goodwin and Hein (Reference Goodwin and Hein1982) noted, applied linguisticsFootnote ² could bring a critical perspective to AI research by challenging the implicit ontological assumptions embedded in AI systems. It can be difficult to bridge disciplinary ontologies; however, this is arguably the case for any interdisciplinary endeavor and, optimistically, some advances have already been made in this area (e.g., in the development of semantic web technologies, Machado et al., Reference Machado, Almeida and Souza2020; Schalley, Reference Schalley2019; Spyns & De Bo, Reference Spyns and De Bo2004). Looking forward, it is imperative that we seek to further share our foundational views of language across disciplinary boundaries to ensure that GenAI development and deployment are sensitive to the diverse ways in which language and knowledge are understood and situated within our lived realities.

AI and ethics in applied linguistics

Applied linguistics research is shaped by macroethics, concerned with institutional forms of ethics, and microethics, concerned with situated ethical challenges within and across the research process (De Costa et al., Reference De Costa, Sterling, Lee, Li and Rawal2021; Yaw et al., Reference Yaw, Plonsky, Larsson, Sterling and Kytö2023). Both facets of ethics are generally guided by respect for people and the goal of yielding optimal benefits while minimizing harm and responding to issues of (social) justice (De Costa, Reference De Costa and De Costa2015). Applied linguistics has shifted over time toward a context-dependent ethics and the view that ethical considerations are not static but are negotiated within specific rhetorical situations (Vetter et al., Reference Vetter, Lucia, Jiang and Othman2024). Ethics has moved beyond a focus on individuals toward a socially constructed view of evaluating ethical behavior.

Macro- and microethics work to protect research participants and ensure the responsible application of linguistic knowledge to social challenges. In this way, ethics in applied linguistics shapes and is shaped by its varying epistemologies and ontologies, linking ethics firmly to the discussion of alignment in the previous sections. The macro- and microethical approach of applied linguistics typically focuses on research practices, addressing issues such as informed consent, data privacy, and the potential for harm to participants (De Costa et al., Reference De Costa, Sterling, Lee, Li and Rawal2021). In each case, ethics is seen as being highly dependent on the specific research context, researchers’ positionality, and participants’ cultural values. Though learned societies can and do proffer ethical guidelines (e.g., BAAL, 2021), they often do so with reflexivity, acknowledging that there is no one universally accepted approach to conducting ethical research in applied linguistics. That is why issues of ethics are often explored within specific subfields of applied linguistics, for example language teaching (Anderson, Reference Anderson and Mirhosseini2017) and corpus linguistics (Brookes & McEnery, Reference Brookes, McEnery, De Costa, Rabie-Ahmed and Cinaglia2024). This compartmentalization of ethics provides space to localize ethical concerns amid our varied practices. Within and across applied linguistics, researchers must then identify ethical concerns and translate idealized discussions of ethics into practice. This transformation can prove challenging when working in complex contexts, with industry partners, and with or on marginalized groups.

While our exploration of epistemologies and ontologies in AI and applied linguistics offered insight into areas of misalignment, ethics in AI research is arguably more easily aligned to ethics in applied linguistics. Ethics is a central facet of responsible AI research, with researchers sharing a focus on evaluating issues such as algorithmic bias, the impact of AI on society, and the complex relationship between AI and the climate crisis (Jabotinsky & Sarel, Reference Jabotinsky and Sarel2024; Kirova et al., Reference Kirova, Ku, Laracy and Marlowe2023). However, as in the case of applied linguistics, ethics in AI research represents an ideal which is not always realized in practice, in particular in the AI industry in which ethical concerns regarding the alignment of GenAI with applied linguistics arise.

Ethics in AI research is largely concerned with the moral principles that govern the development and practical implementation of AI systems, particularly with regard to a lack of transparency and fairness, and algorithmic bias (Kirova et al., Reference Kirova, Ku, Laracy and Marlowe2023). However, there have been instances in which AI companies have been accused of training their models on copyrighted materials without permission (e.g., Milmo, Reference Milmo2024) and this has raised questions surrounding the legality of certain AI models (Gromova et al., Reference Gromova, Ferreira and Begishev2023). Further critiques of training data contend that models used in popular GenAI tools may not offer a nuanced picture of reality (Creely, Reference Creely2024; Farrelly & Baker, Reference Farrelly and Baker2023) – at least from an applied linguistics perspective. What AI can produce is, as noted, confined to its ontology, that is, to what the AI has been trained on. Typically, AI models are trained on texts from the web and therefore reflect the style and tone attributed to journalistic language (Nesi, Reference Nesi2024). As mentioned previously, the increasing prominence of AI-generated texts on the web (Europol, 2022) means that new AI models will inevitably be trained on AI-produced texts, which risks the exponential proliferation of biases, prejudices, and dominant perspectives through GenAI tools. Indeed, Thompson et al. (Reference Thompson, Dhaliwal, Frisch, Domhan, Federico, Ku, Martins and Srikumar2024) argue that already 57% of online data is the result of AI generation or machine translation.

Thus, the question of bias in AI-produced texts may not be an issue of algorithmic bias per se. It may be that AI and GenAI tools are effectively reconstructing reality based on norms and tendencies in the data as it exists. When evaluated from an algorithmic perspective, we will then likely conclude that such tools are effective at doing their job and at ethically reflecting such social tendencies, even if the kinds of biases and stereotypes that are reproduced are views that many of us in (critical) applied linguistics would challenge. As discussed earlier, these problems with training data can result in GenAI tools that reproduce knowledge from uncredited sources (Creely, Reference Creely2024; Stahl & Eke, Reference Stahl and Eke2024) and AI models that perpetuate existing societal prejudices and biases related to gender, race, and culture (Choi, Reference Choi2022; Putland et al., Reference Putland, Chikodzore-Paterson and Brookes2023). Such issues in AI development are antithetical to values in applied linguistics and, from an ethical perspective, can undermine advances in social justice initiatives, researcher autonomy, and representations of identities, inter alia.

While black box research (Casal & Kessler, Reference Casal and Kessler2023; Curry et al., Reference Curry, Baker and Brookes2024) and reasoning processes (Lodge et al., Reference Lodge, Thompson and Corrin2023) in modern AI tools present epistemological and ontological challenges, they also create a particular ethical challenge for researchers in terms of the explanatory power of results derived from GenAI (for more on this, see Egbert et al., Reference Egbert, Larsson and Biber2020; McEnery & Brezina, Reference McEnery and Brezina2022). An analogy would be like seeing the answer to a mathematical puzzle without being able to see how the answer was derived. Being unable to explain the answer is, in some ways, as problematic as a wrong answer, as it is the process of discovery, as well as the discovery itself, that forms the contribution to knowledge. In particular, understanding errors is rendered difficult-to-impossible in a context in which explanatory adequacy is low. As discussed in the “AI and epistemology in applied linguistics” section, research on explainable AI is now coming to the fore (see Angelov et al., Reference Angelov, Soares, Jiang, Arnold and Atkinson2021, for an overview) and offering some solutions to ethical concerns in the research process. However, some of the techniques used in modern AI and other areas, in particular, opaque models such as random forests and deep neural networks (Rudin, Reference Rudin2019), provide a substantial, possibly insurmountable, barrier to explanation. More recent developments, such as ChatGPT o1, have endeavored to shed more light on these reasoning processes. However, the reliability, consistency, and depth of such reasoning still needs to be tested. In this context, it is hardly surprising that questions of over-reliance on GenAI in the research process (Creely, Reference Creely2024; Lodge et al., Reference Lodge, Thompson and Corrin2023), the impact of GenAI data processing and hallucinations on data integrity and security in applied linguistic research (e.g., Curry et al., Reference Curry, Baker and Brookes2024; Muñoz-Basols et al., Reference Muñoz-Basols, Neville, Lafford and Godev2023), and the role of GenAI in producing and reporting research findings (Casal & Kessler, Reference Casal and Kessler2023) abound in the literature. These concerns represent a notable misalignment between GenAI tools and the ethical values on which, in principle, research in applied linguistics rests.

With this overarching review in mind, it becomes clear that despite academic research in AI demonstrating shared ethical values with applied linguistics research, in practice, commercial AI has not always enacted these values and some of the black box algorithms used in both industrial and academic research are poorly aligned with epistemological and ontological practices in applied linguistics. From an applied linguistics perspective, it is reasonable to assert that the use of GenAI tools that reinforce bias and negatively impact society may undermine fairness and equity in language education (Choi, Reference Choi2022), perpetuate normative or ethnocentric views (Spennemann, Reference Spennemann2024), and mis- and underrepresent society’s most vulnerable (Nguyen et al., Reference Nguyen, Nguyen, Ludovise and Santagata2024; Putland et al., Reference Putland, Chikodzore-Paterson and Brookes2023). Likewise, algorithms that obscure, simplify, or hide the processes by which analyses are undertaken also align poorly with macroethical research practices in applied linguistics. To develop a means for applied linguists to engage with AI ethically, such concerns must be addressed and attenuated.

To respond to this misalignment, we can turn to a growing body of work dedicated to enhancing GenAI users’ critical AI literacy. This research seeks to make GenAI users aware of how AI works, its limitations, and the ethical implications of its use (e.g., Casal-Otero et al., Reference Casal-Otero, Catala, Fernández-Morante, Taboada, Cebreiro and Barro2023; Strauß, Reference Strauß2021; Walter, Reference Walter2024). Yet, while research designed to develop guidelines for the ethical use of AI is already available, there remains a need to develop disciplinary-situated guidelines that localize ethical AI use or non-use within and across applied linguistics. We must tease apart and juxtapose contextually situated ethics in applied linguistics and AI research and determine their compatibility. We must adopt a critical use of AI and question the environmental and social impacts of AI, particularly in the global south where limited resources, such as water and energy are used to power and cool large server farms (Bashir et al., Reference Bashir, Donti, Cuff, Sroka, Sze, Delimitrou and Olivetti2024). There are movements toward sustainable AI (see van Wynsberghe, Reference van Wynsberghe2021), which is promising, and any user of GenAI should give considerable thought to the models on which they draw. Supporting the use of sustainable AI when we use it may be the best means of combatting such unethical practices. More importantly however, we must decide whether or not the expediency that AI promises is worth any such costs or negative impacts. Grieve et al. (Reference Grieve, Bartl, Fuoli, Grafmiller, Huang, Jawerbaum and Winter2025) note that language models could be improved through engagement with sociolinguistics research. Arguably, drawing on such social perspectives on language when developing AI could help to lead us towards critical and responsible use of AI in applied linguistics – a use that prioritizes the human values, transparency, and accountability that sit at the heart of applied linguistics.