Generative AI tools like ChatGPT and machine translation have made it easier than ever for students to produce fluent language, but a new opinion article in Frontiers in Education argues this creates a fundamental problem for assessment: a polished final text no longer reliably shows how much a learner actually knows.
J Y Jun Yan of the University of Bath's Department of Education writes that the central question for teachers is not simply whether AI produced a given sentence, but whether the learner can explain how meaning was produced, revised, culturally mediated, and responsibly conveyed. The piece, published on 16 November 2026, proposes a framework called "accountable judgement" for AI-mediated language, translation, and multilingual assessment.
According to Yan, the issue extends well beyond specialist translation programmes. In foreign language, English for Academic Purposes (EAP), English as an Additional Language (EAL), and multilingual classrooms, teachers routinely use language products — translated sentences, bilingual explanations, target-language summaries, AI-revised drafts, and polished academic paragraphs — as evidence of learning. Generative AI and neural machine translation destabilise the evidential relationship between text and learning, Yan argues, because a competent-looking output may reveal little about the linguistic, communicative, cultural, or ethical judgement the student actually exercised.
From product verification to accountable judgement
Yan defines accountable judgement as "learners' capacity to explain, provide evidence, and take responsibility for decisions made when AI has shaped a language task, a translation task, or a multilingual text." The concept builds on evaluative judgement — the ability to assess the quality of one's own and others' work — but adds a requirement that learners make visible the reasoning behind their decisions, disclose how AI contributed, and take responsibility for the consequences of the final text.
The framework identifies four dimensions of accountable judgement, each with its own assessment question and possible forms of evidence:
- Linguistic judgement: Can the learner justify their choices concerning accuracy, grammar, vocabulary, register, and fluency? Evidence might include annotated drafts or language-choice notes.
- Communicative judgement: Can the learner adapt the language to the audience, purpose, relationship, and situation? Oral justifications and reflective commentary could demonstrate this.
- Cultural mediation judgement: Can the learner explain how cultural meanings, identities, values, humour, politeness, and traces of minority languages are retained, adapted, explained, or made visible? Cultural translation notes and bilingual commentary are suggested evidence.
- Ethical judgement: Can the learner disclose AI use and take responsibility for the final text? AI-use statements and decision logs could serve this purpose.
The limits of AI-supported fluency
Yan acknowledges that generative AI and neural machine translation offer clear benefits in language and translation classrooms. Research cited in the article notes that AI can provide vocabulary support, reformulations, grammar suggestions, and alternative registers. In translation pedagogy, post-editing, comparison, and revision have become central, requiring learners to identify risk and decide whether to accept, modify, or reject machine-generated suggestions.
However, AI-supported fluency creates a specific assessment problem, Yan writes. In a language-learning setting, a statement can be syntactically correct yet pragmatically unsuitable—overly blunt, excessively polite, culturally jarring, or mismatched to the speaker-listener dynamic. In a translation-learning setting, a smooth machine-assisted output can conceal unresolved issues concerning tone, cultural adaptation, and accountability. The evaluation question, Yan argues, should shift from whether the final product is accurate or fluent to why particular AI suggestions were accepted, rejected, or revised, and what cultural meaning was preserved, adapted, explained, or resisted.
From feedback uptake to judgement uptake
The article also connects accountable judgement to broader debates about feedback in education. Yan notes that recent discussion of AI feedback in writing instruction has shifted attention from feedback quality to learners' engagement with feedback. Citing work by Helm and Hesse (2026), Yan argues that AI feedback should not be treated as an endpoint but as a catalyst for classroom communication and revision.
Yan pushes this argument further: feedback uptake alone is not enough unless it becomes "judgement uptake" — the process through which learners transform AI suggestions, teacher comments, peer feedback, or automated corrections into their own explainable decisions. Accountable judgement, Yan writes, differs from simple compliance with feedback: students are not only expected to revise but to account for why a revision is linguistically appropriate, communicatively effective, culturally responsible, and ethically disclosed.
Cultural mediation as meaning responsibility
A central theme of the article is that the question "Who owns the text?" is too narrow for AI-mediated assessment. Yan proposes a more significant question: "Who carries the meaning?" AI can generate words, Yan argues, but it does not carry educational, cultural, or ethical responsibility for those words.
Drawing on translation studies scholarship, Yan notes that translation is not only linguistic transfer but cultural mediation. Thus, translation done by AI involves not just linguistic work but also a cultural evaluative process. Even if a text produced by a machine is correct in terms of vocabulary and grammar, it may lack elements such as humor, politeness, contextual relationships, local expertise, historical context, ideological stances, or representation of minority languages. Cultural mediation goes beyond mere fluency or naturalness., Yan writes. It requires learners to decide what should be retained, adapted, explained, softened, resisted, or made visible for a particular audience. These decisions, Yan argues, are not merely stylistic; they are part of the learner's accountable judgement.
Implications for students and educators
For university students studying languages, translation, or any subject involving written assessment, the article's argument has practical implications. Yan suggests that assessment designs should increasingly require students to demonstrate their reasoning process — through revision trails, decision logs, reflective commentaries, annotated drafts, oral justifications, and bilingual explanations — rather than relying solely on the polished final product.
Yan also argues that such an approach is especially important for multilingual equity. An assessment design that recognises only polished final products, Yan writes, may miss precisely the judgement that multilingual learners are well placed to show.
The article situates this argument within wider higher education debates that are moving beyond simply catching AI use towards assessment designs that make authentic learning, process transparency, and assessable judgement visible. For language-intensive assessment, Yan argues, this shift is particularly urgent because language is not only a product to be corrected; it is also a medium through which learners express meaning, identity, cultural positioning, and disciplinary understanding.