Who Is the Human in the Loop? AI and the Future of Multilingual Education
Anandana Kapur, Independent Researcher, Educator, Filmmaker, US
Introduction
Generative AI has arrived in multilingual education, promising frictionless access. Models can draft, translate and give feedback across many languages. Yet when AI sits between learners, teachers, community and knowledge, it is not a neutral pipe. It shapes what counts as acceptable language, which registers feel natural, and whose knowledge travels easily.
If AI becomes the intermediary, who is responsible for teaching what cannot be translated or automated—the hesitation, the search for the right culturally specific word, the repair of a misunderstanding that is itself a form of learning? The next frontier is not simply better multilingual AI. It is the governance of AI-mediated multilingualism: whose language, knowledge and cultural assumptions the systems privilege; who owns the data generated in multilingual classrooms; and what forms of human linguistic and pedagogical agency must remain non-automated.
The Assumptions Behind “Multilingual AI”
Because models can generate text in many languages, they appear naturally suited to multilingual education. Language, however, is not the same as translation.
Multilingual practice is social and situated. A teacher still needs to decide: “Even if this translation is grammatically correct, it feels culturally off.” Or “The register is wrong for these learners.” Or “The model has turned a locally meaningful expression into a generic English definition.” Or “I need to ask someone who knows this community rather than prompt the model again.”
A useful framework for assessing AI-mediated multilingualism is:
- Linguistic equivalence: Can the AI produce a parallel sentence?
- Communicative equivalence: Will the recipient understand it in the intended way?
- Relational equivalence: Does it carry the same sense of trust, familiarity, dignity and social relationship?
Most technical evaluations stop at the first. Multilingual pedagogy lives in the second and third.
Authorship and Translingual Composition
The familiar question “Did the student write this?” becomes more complicated in multilingual settings. A learner may think in one language, discuss an idea orally in another, use AI to move it into a third, ask the model to polish academic English, and then submit. The ideas and the evaluative judgement may be the student’s; the final surface form may not be.
Institutions therefore need finer distinctions—idea generation, interlingual mediation, stylistic scaffolding, substantive generation, assessment substitution—rather than a simple plagiarism binary. Traditional plagiarism detection software interprets multilingual syntactic mediation as automated fabrication, disproportionately penalizing English-as-an-additional-language (EAL) learners. Multilingual learners often make these boundaries more visible than monolingual ones.
The COVID Lesson: Community as the Human in the Loop
During the early COVID response, the author led a citizen initiative in India that produced public-health videos in 22 languages (U.S. Embassy India, 2020). The contributors were domestic workers, homemakers, students and retirees—people who knew how their communities actually spoke. They understood which register sounded urgent without sounding panic-inducing, which official terms alienated daily-wage workers, which local analogies explained transmission, and when literal translation would break trust.
This pre-dated everyday LLM translation tools. It showed that effective multilingual communication is relational and contextual. Linguistic competence is embedded in social knowledge. A native speaker does not merely know word-for-word correspondences; they know what sounds patronizing, what builds solidarity, and what should not be translated literally—because they belong to the community being addressed.
In technical discourse, human-in-the-loop usually means an operator who checks for hallucinations. If the only human checking school outputs is an English-speaking administrator, the deeper problem of AI-mediated multilingualism remains unsolved. The relevant human may need linguistic, cultural, community, pedagogical and lived knowledge. That expertise can reside with the domestic worker, the retiree who knows local idioms, the student who understands peer talk, or the security guard who knows the social context of the audience. Policy must recognize whose linguistic expertise counts.
Reframing Critical Multilingual AI Literacy
Standard AI literacy focuses on prompting, spotting hallucinations and basic bias awareness. Critical multilingual AI literacy asks a sharper question: who is the human in the loop, and what knowledge do they bring?
Large models are trained disproportionately on high-resource, standardized varieties, especially English. They can introduce linguistic and cultural bias, flatten culturally specific meaning, and treat dialect or register as noise (Baumgarten & Tieber, 2025; Bender et al., 2021). Functional literacy asks whether a learner can use the tool. Critical agency asks whether teachers and learners can understand, question and influence the conditions under which the system operates (Pack & Maloney, 2024).
What Policy Must Anticipate
Three structural questions press for attention:
- Representation: Who is represented in the training and evaluation data?
- Agency: Who may correct and shape AI-mediated language in educational settings?
- Governance and data sovereignty: When learners upload voice, writing, family expressions or classroom talk, who owns that linguistic data? Can it be reused to train future models? What happens if the commercial platform changes terms or drops support for a language?
Commercial dependence is not a side issue; it is integral to governance. Public education systems that rely on proprietary platforms for core multilingual functions risk long-term loss of control over curricular standards and community knowledge (Pack & Maloney, 2024; UNESCO, 2023).
Concrete, implementable moves that flow directly from these questions include:
- Procurement and vendor contracts: Require linguistic-data sovereignty clauses so that learner speech, writing and classroom recordings stay under school, district or state control. Any commercial reuse or model training must need explicit, time-limited consent and local oversight.
- Assessment and academic integrity redesign: Adopt clearer stage-based distinctions: idea generation, interlingual mediation, stylistic scaffolding, substantive generation, and assessment substitution.
- Teacher education and human-in-the-loop protocols: Build the five literacies (AI, sociolinguistic, cultural-pragmatic, data-stewardship, and governance) into pre-service and in-service programs so teachers know both how the systems work and when human judgement must intervene.
- School- and system-level governance: Create formal roles or modest paid micro-contracts for community language mediators i.e. parents, elders, bilingual assistants and local speakers, as recognized reviewers of AI-generated materials for register, trust and cultural safety.
- Public infrastructure and open alternatives: Prefer or co-develop open-weight or publicly governed multilingual models for core educational uses, and avoid sole reliance on proprietary platforms for minority-language support.
These five steps keep human linguistic agency at the center while remaining practical for ministries and school systems already negotiating technology contracts.
Classroom Tensions and Future Competencies
Integrating generative tools surfaces recurring tensions:
[Insert Figure 1]
Table 1: Possibilities and perils of AI-mediated multilingualism
Design cannot stay trapped in a simple optimism/rejection binary (Kapur, 2023). Educators will need overlapping competencies: AI literacy (how the system works), linguistic literacy (what it does differently across languages), cultural literacy (what contexts it misunderstands or flattens), data literacy (whose knowledge is being used), and AI governance literacy (who makes the rules).
The classroom is becoming a site of technology governance, not only technology consumption. AI-supported multilingual education is ultimately a question of linguistic, cultural and institutional agency.
References
Baumgarten, S., & Tieber, M. (Eds.). (2025). The Routledge handbook of translation technology and society. Routledge. https://doi.org/10.4324/9781003271314
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? 🦜. Proceedings of the 2021 ACM FAccT Conference, 610–623. https://doi.org/10.1145/3442188.3445922
Kapur, A. (2023). Embracing uncertainty: A paradigm shift for anti-racist technology policy design. Anti-Racism Policy Journal, Harvard Kennedy School.
Pack, A., & Maloney, J. (2024). Using artificial intelligence in TESOL: Some ethical and pedagogical considerations. TESOL Quarterly, 58(2), 1007–1018. https://doi.org/10.1002/tesq.3320
UNESCO. (2023). Guidance for generative AI in education and research. UNESCO Publishing. https://unesdoc.unesco.org/ark:/48223/pf0000386693
U.S. Embassy India. (2020, May 27). #DostiUnitesUs: “You may appeal in the language of one’s education and intellect but you can convince only in the language of the heart!”, says Fulbright-Nehru alumna and filmmaker Anandana Kapur… [Facebook post]. Facebook. https://www.facebook.com/India.usembassy/photos/dostiunitesus-you-may-appeal-in-the-language-of-ones-education-and-intellect-but/3205966286186045/
Dr. Anandana Kapur is a filmmaker, researcher and educator working at the intersection of emerging media, culture and public-interest technologies. A Harvard alumna and former Fulbright Fellow at the MIT Open Documentary Lab, she writes on AI governance, inclusive media and community-led knowledge systems.