The AI ‘alignment problem’ in language education and applied linguistics

A recurring theme in many conversations about artificial intelligence (AI) is what might be termed the alignment problem. This refers to the question of whether the AI systems we use, however powerful, actually reflect our (human) values, intentions, or needs. This is not a new idea, or one that uniquely associated with AI: it is exactly what we experience, in a simple form, whenever the autocomplete function produces something different from what we might have wanted. However, the stakes are higher now that generative AI systems are capable of producing fluent, persuasive, and sometimes authoritative-sounding text at scale.

A recent article by Neil Curry, Tony McEnery, and Gavin Brookes, in the Annual Review of Applied Linguistics (2025), brings this debate into our own disciplinary backyards of applied linguistics and language education. The paper, A Question of Alignment: AI, GenAI and Applied Linguistics, is both a warning and a call to action. It urges applied linguists, language teachers, and researchers to think carefully about how AI technologies are reshaping our field, and what it would mean to ensure that these tools truly align with our epistemological, ontological, and ethical commitments.

In what follows, I will highlight some of the paper’s central arguments and reflect on why they matter for those of us who work at the intersection of languages, research, and education.

---

What do we mean by ‘alignment’?

At its simplest, the alignment problem in AI refers to the gap between what humans intend a system to do and what the system actually does. This is potentially a problem, because a misaligned AI might optimise for the wrong goals. For example, if you ask a system to make students’ essays “better,” it might substitute synonyms mechanically, or worse, produce plausible-sounding but fabricated content. The output may appear polished, but it may not align with the educational purpose of encouraging critical thinking or authentic expression.

Curry, McEnery and Brookes argue that alignment is not merely a technical challenge. It is also about how AI systems intersect with the epistemologies (i.e., our beliefs about knowing), ontologies (i.e., our beliefs about beign and existence), and ethics (i.e., our beliefs about values, right and wrong) that underpin applied linguistics and language education.

Put differently: even if an AI system performs flawlessly according to its internal logic, it might still misalign with what applied linguists take to be meaningful knowledge, legitimate data, or fair practice.

Epistemology:
What counts as knowledge?

Applied linguistics is a field deeply concerned with how knowledge is generated and validated. We spend considerable time debating the merits of different research methods, the scope of evidence that can support claims, and the interpretive frameworks that give findings meaning.

GenAI complicates these processes. On the one hand, AI tools can facilitate research by handling large datasets, generating transcriptions, or suggesting connections across vast literatures. On the other hand, they may obscure the very basis on which knowledge rests. A text generated by an AI may look persuasive, but it might not have been produced through processes that we recognise as legitimate research practice: systematic data collection, transparent analysis, and critical peer review.

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.

Curry et al. (2025, original emphasis)

There is a risk, in other words, of normalising epistemic shortcuts, that is to say, outputs that look like knowledge but lack the underpinning rigour. If we take such shortcuts at face value, we risk diluting the very standards of evidence that define our discipline.

This echoes concerns I have raised in earlier reflections on research literacy, where the challenge is not simply to access information, but to evaluate its quality and provenance.

Ontology:
What counts as language?

Ontology, in this context, concerns the nature of what we study (language, interaction, meaning etc.) and how it exists in the world. There is a long-standing debate in applied linguistics whether language should be understood primarily as a cognitive system, a social practice, or something else entirely.

GenAI forces us to revisit these debates. Large Language Models (LLMs) mimic language by building on statistical correlations across immense datasets. They do not “know” language in the sense that humans do; like an auto-corrector on steroids, they generate plausible continuations of text without reference to communicative intent, embodied interaction, or lived experience.

This raises difficult questions. If AI-produced text is increasingly part of the linguistic landscape, should we treat it as data for applied linguistics? If so, what assumptions are we making about the ontology of “language”? Conversely, if we exclude such text as ‘inauthentic’, do we risk ignoring a growing dimension of how people actually encounter language in their daily lives?

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.

Curry et al. (2025).

There is no easy answer. What is clear is that AI unsettles ontological assumptions that many of us have tacitly relied upon, forcing us to clarify what we mean when we say we study “language.”

Ethics:
What counts as fair and responsible practice?

Finally, there is the ethical dimension. Applied linguistics and language education have long been attentive to the moral implications of our work, including issues of of linguistic inequality, language rights, access to education, and the politics of representation.

AI tools, as Curry and colleagues note, are never neutral. They encode biases present in their training data, reflect the interests of the corporations that build them, and risk exacerbating inequalities of access and representation. For instance, an AI trained on dominant language varieties may marginalise non-standard or minority forms (bias). Another example of inequality would be when a well-resourced institutions is able to adopt AI at scale, while under-resourced settings fall further behind.

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.

Curry et al. (2025).

The ethical alignment problem, then, is not just whether AI avoids causing harm, but whether it actively contributes to a more just, inclusive practice of applied linguistics. This resonates with wider debates about AI and social justice in education, where the risk is not just technological disruption, but the reinforcement of old inequities under a new guise.

---

Language education: Opportunities and tensions

The authors devote special attention to language education, where GenAI is already making inroads. Automated writing feedback, adaptive textbooks, AI-driven tutoring platforms are just some examples of how artificial intelligence has been brought to bear on language education that promise efficiency, scalability, and personalisation.

But such systems also raise tensions. For example, if a machine gives instant grammar feedback, what becomes of the teacher’s role in fostering critical engagement with language? If students outsource essay writing to AI, how do we maintain the integrity of assessment? If adaptive tools decide what learners “need,” whose pedagogical values will the learning experience reflect?

The temptation is to see these as technical issues that will likely be resolved by better prompts, smarter detectors, or improved interfaces. But Curry et al. insist these are fundamentally alignment issues, in the sense that they are questions of how AI systems embody the values of education rather than the delivery of correct answers.

---

Towards a constructive path

If the risks are serious, so too are the opportunities. The article is not anti-AI; it is a call for a more reflective, responsible integration of AI into applied linguistics. Three directions seem especially important:

Interdisciplinary collaboration

The development of AI systems for applied linguistics and language education cannot be left solely to computer scientists. Or, conversely, applied linguists and language teachers should not approach it as passive consumers. Collaboration across disciplines can help shape technologies that respect both computational possibilities and linguistic realities.

Ethical stewardship

As a community, we must articulate clearer ethical frameworks for how AI should be used in research and education. This includes not only avoiding harm but actively promoting inclusivity, transparency, and fairness. Our ongoing work in the ECML-funded AI Lang project, is an example of such systematic guidance for language teachers.

Critical reflexivity

Perhaps most importantly, we need to remain self-critical about our own assumptions. AI has exposed the fact that applied linguists and language education are not in perfect alignment, even among ourselves: we hold diverse epistemologies, ontologies, and ethical commitments. Engaging with AI is an opportunity to surface these differences and negotiate them more openly.

---

Why this matters now

The timing of this article is significant. In the past two years, GenAI tools have moved from the margins to the mainstream of educational and research practice. Many of us are already experimenting —cautiously or enthusiastically— with ChatGPT, Claude, Gemini, or other applications that fill niches in language teaching. Universities are drafting guidelines; publishers are setting policies; students are quietly (?) using AI in their assignments.

The risk is that practice will outrun reflection. By the time we pause to ask how AI aligns with our values, it may be too late to influence its trajectory. Curry, McEnery and Brookes remind us that applied linguistics has a long tradition of engaging critically with language technologies, from corpora to CALL to automated translation. GenAI is the latest, and perhaps the most disruptive, in this lineage.

If we take the alignment problem seriously, we can approach this disruption not with fear, but with a sense of responsibility: to ensure that the tools we adopt do not simply reshape our field, but do so in ways that remain true to its deepest commitments.

---

Closing thoughts

Reading “A Question of Alignment” left me with two impressions. First, that the alignment problem is real and pressing, not just in Silicon Valley think-pieces but in the heart of applied linguistics. Second, that we are uniquely positioned to address it: our expertise in language, context, and human meaning equips us to ask the questions that technologists often overlook.

In other words, applied linguistics and language education must not simply adapt to AI; rather, we must ensure that AI adapts to us.

The challenge, then, is one of agency. Do we accept AI as it arrives, retrofitting our practices to accommodate its quirks? Or do we assert our disciplinary voice, insisting that if AI is to have a place in applied linguistics, it must align with the values of rigour, inclusivity, and ethical responsibility that have long defined the field?

That, I think, is the real question of alignment.

---

If you are experimenting with AI in your teaching or research, what opportunities or tensions have you noticed? Sharing experiences in the comments may help us think more collectively about what responsible alignment could look like.