On this page

Synthesis: Yan (2026) argues that generative AI and machine translation have broken the evidential link between a fluent text and the learning behind it: a polished paragraph, translation or summary may now be compatible with almost any amount of learner judgment. In language, translation and multilingual assessment, the question worth asking is therefore not "did AI produce this?" but "who carries the meaning?" His answer is accountable judgment, an extension of evaluative judgment that requires learners to explain the reasoning behind their linguistic, communicative, cultural and ethical decisions, disclose how AI contributed, and take responsibility for the result. The piece sets out four dimensions with the evidence each would need, and argues that assessment should shift from feedback uptake to judgment uptake.

Key Findings

  1. Fluency is no longer evidence of learning. AI-supported text can be accurate at the sentence level yet pragmatically inappropriate, culturally flattened or ethically undisclosed.
  2. Accountable judgment extends evaluative judgment. Beyond judging quality, learners must make the reasoning behind their choices visible, disclose AI contribution and accept responsibility for meaning.
  3. Four dimensions carry different evidence. Linguistic, communicative, cultural-mediation and ethical judgment each need their own assessment question and their own artifact rather than one holistic score.
  4. Feedback uptake is the wrong target. What matters is whether learners transform AI suggestions, teacher comments and peer feedback into explainable decisions of their own.
  5. Disclosure alone proves nothing. A student can declare AI use honestly and still be unable to account for the final text, which is why explanation rather than declaration is the evidential core.
  6. Cultural mediation is the dimension most easily missed. Deciding what to retain, adapt, explain, soften or make visible for a particular audience is a judgment that fluency metrics do not capture.

What the four dimensions ask of a learner

Linguistic judgment asks whether the learner can justify choices about accuracy, grammar, vocabulary, register and fluency, evidenced by annotated drafts and language-choice notes. Communicative judgment asks whether they can adapt language to audience, purpose, relationship and situation, evidenced by oral justification or reflective commentary. Cultural mediation asks whether they can explain how cultural meanings, humor, politeness, identities and minority-language traces were retained, adapted or made visible, evidenced by cultural translation notes and bilingual commentary. Ethical judgment asks whether they can disclose AI use and take responsibility for the text, evidenced by decision logs and the AI-use statement.

The distinction from evaluative judgment matters because the two are often treated as one capability. Evaluative judgment concerns recognising quality in one's own and others' work, including AI output. Accountable judgment adds the responsibility to justify and evidence that judgment at the point where agency is distributed between learner and machine. In assessment terms, the learner must substantiate the decisions behind an AI-supported text rather than let the fluency of the final product stand as proof of competence.

Why language assessment feels this first

Language is both a product to be corrected and a medium through which learners express meaning, identity and cultural positioning, which is why the evidential collapse shows up here before it shows up in other disciplines. In a language classroom a sentence can be grammatically correct and pragmatically wrong: too direct, too formal, culturally awkward for the relationship between speakers. In a translation classroom a fluent target text can hide unresolved register, cultural-mediation and responsibility problems. In EAP, EAL and multilingual classrooms, the risk is that an assessment recognising only polished final products misses precisely the judgment that multilingual learners are well placed to display.

Yan's practical proposals are deliberately small. A language-writing task might require students to explain three AI-suggested changes they accepted, two they rejected, and one cultural or pragmatic issue they had to resolve. A translation assignment might pair an AI-output comparison with a risk log, a cultural-mediation note and a final reflective commentary. A multilingual task might accept bilingual commentary or an oral explanation as legitimate evidence.

He is also careful about the limits of process evidence. Decision logs and reflective commentaries may themselves be AI-generated, so their evidential value depends on how they are elicited; situated classroom activity, dialogue and selective oral justification provide complementary evidence, and no single mode establishes authenticity on its own. Reviewing detailed logs in large classes also raises workload questions, which is why he suggests asking learners to justify a small number of consequential decisions through short structured templates.

What this means for practice

  • Language and translation instructors. Require an account of a few consequential choices — what was accepted, rejected or adapted — rather than a general AI-use declaration.
  • Assessment designers. Score the four dimensions separately where the outcomes warrant it; a single holistic mark hides which judgment is missing.
  • Instructors. Build oral or dialogic checks into written tasks, since written process evidence can be generated by the same tools it is meant to evidence.
  • Academic integrity teams. Move the policy question from detection to what evidence shows that learning, judgment and responsibility took place.

Limitations

  • No empirical evidence is reported. This is an opinion piece: the framework, dimensions and task examples are proposed, not tested with students or raters.
  • Process evidence can be AI-generated. The author acknowledges that decision logs and reflective commentaries may be produced by the same tools, and that no single evidence mode establishes authenticity.
  • Workload is acknowledged but unresolved. Reviewing decision logs and oral justifications at scale is not costed, and the proportionate model is a suggestion rather than an evaluated design.
  • Scope is largely higher education. The examples come from university language, translation and EAP teaching, and applicability to school-level language assessment is not developed.

Citation

Yan, J. (2026). Who carries the meaning? Accountable judgement in AI-mediated language, translation and multilingual assessment. Frontiers in Education, 11, 1929156.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.