How Can AI Support Multilingual Learners Without Misjudging Their Knowledge or Erasing Their Voice?

A student understands the concept and needs four extra minutes to say so in English. The AI can translate the prompt, define the terminology, rehearse the explanation with them, or write the whole answer. Three of those are teaching. One is doing the assignment.

Short answer to give a colleague: decide what the language is doing in this task before you decide what the tool may do. Where language is incidental, AI can remove a barrier that has nothing to do with your construct. Where the language is the construct, the same tool bypasses the learning. The risk runs both ways at once: language support can improve access, and automated scoring can still misread the student. In one physics study, explanations of lower linguistic quality were systematically graded below expert judgment by every AI scoring approach tested, nine machine-learning and two language-model systems across 116 secondary-school students (Feser and Tschisgale, 2026). Fluency and understanding were being scored as one thing.

The short version

1. State the role of language in the task: incidental, central, or one criterion among several.

2. Write permitted support into the instructions and criteria, and say what students must still do themselves.

3. Give targeted help before whole rewrites. Ask the tool to identify unclear meaning and offer options.

4. Require the student to compare the suggested wording against their own meaning and course sources.

5. Score conceptual reasoning and language separately when both are intended outcomes.

6. Never make an untested tool the only route to support.

Decide what the language is doing

In a biology explanation, grammar is mostly incidental to the concept being assessed. In a language course, producing and revising particular structures is the learning. In professional communication, disciplinary accuracy, audience and register all count. The same translation tool is appropriate in the first case, questionable in the second and negotiable in the third.

Write the decision down where students can read it. If students may translate directions but must produce a target-language explanation unaided, say so explicitly. If terminology support is allowed, specify what the student still has to explain and verify. Then check the arrangement against accommodation requirements and institutional policy rather than treating every form of language assistance as interchangeable. A student with a documented accommodation and a student avoiding practice can request the same thing, and only the second is your teaching problem.

What the research supports, and what it does not

Lopez and Arabadzhy (2026) reviewed 24 empirical studies of generative chatbots in language learning. Their contribution is taxonomic: every sub-affordance catalogued for pre-generative chatbots persists, and three are new, including translanguaging, the deliberate movement between languages as part of learning. That is a map of reported uses and experiences rather than a pooled estimate, and the corpus is geographically concentrated, with China, Iran, Taiwan and Türkiye accounting for 19 of the 24 studies and all but one set inside formal institutions. Support for using cross-language explanation as a learning move is not support for translating every assignment.

Assessment is where the evidence gets sharper and the stakes rise. Feser and Tschisgale (2026) compared eleven automated scoring approaches against human expert scores for 116 secondary-school students' physics explanations and found systematic underestimation of conceptual understanding for linguistically weaker explanations, in every single approach. The design isolates the mechanism: the same conceptual content, expressed in weaker prose, scored lower. This is not a claim about any particular country's multilingual students, and it is exactly the pattern that makes overall scoring agreement a poor diagnostic. A scorer can agree with human markers on average while erring in a consistent direction for the students whose language is still developing.

The Multilingual Learning synthesis describes writing-screening research showing multilingual predictive-bias patterns that resemble human scoring biases. Read the three findings together and the operational lesson is narrow: neither "AI removes human bias" nor "AI adds linguistic bias" is a useful rule. What you need is a check on the task, the population and the language, at the point where a score becomes a grade.

Give the tool a specific supporting role

Start with the directions: have the tool translate or restate them in plain language, check that output against the original, and leave the student to work out what the task requires and to ask about any genuine ambiguity.

Disciplinary vocabulary works differently. Set the tool to compare meanings across languages and to generate contextual examples, then have the student explain the concept and identify the misleading equivalents.

Speaking practice calls for a bounded rehearsal with retries, with the student producing the response and demonstrating the target skill.

When revising writing, ask for a few named problems with possible fixes and keep the decision and its justification with the student.

Reading course content is the widest case. A bilingual explanation grounded in approved materials helps, provided the student checks the key claims and demonstrates understanding.

These are uses to pilot, not validated interventions. A tool that performs well in one language or one genre is not thereby reliable in another, and the review's geographic concentration means your language pair may barely be represented in the evidence.

Protect the student's meaning and voice

Ask for targeted help before asking for a rewrite, because a whole rewrite is where the student's thinking leaves the page. A request worth teaching: "Identify two places where my meaning is unclear. Explain the problem and offer options. Keep my examples and do not add claims." The student then chooses.

A sequence that keeps the work with the student:

1. Develop the explanation in whatever language or mix of languages lets the student reason, where the task permits.
2. Use approved support to examine terminology or expression.
3. Compare the suggested wording against the intended meaning and course sources.
4. Produce the required explanation and note the consequential revisions.

In a sociology assignment, a student might reason through a family-role conflict bilingually, check a short list of course terms, then write the submitted analysis in the required language, with the instructor assessing the sociological reasoning and any separately stated language criteria. In a language-learning task the same sequence would give away the assessment, and would need to preserve independent target-language production.

Avoid making "sounds like a native speaker" the standard, since it measures the student's relationship to a dialect rather than their grasp of your subject. Specify clarity, accuracy, genre or audience instead. A polished rewrite can shift stance, drop a culturally meaningful expression, or add a claim the student never intended, and the student signs it.

Check the tool and the assessment

Review meaning, not fluency. Sample translations and explanations for technical terms, omissions, invented detail, cultural assumptions and the varieties of language your students actually use. Involve qualified language colleagues or support staff; do not make students the sole auditor of a tool you require them to use.

Separate the criteria. When both reasoning and language are intended outcomes, score them separately and check whether an AI-assisted marker is penalizing language twice, or missing correct reasoning expressed unconventionally. This is the practical form of validity for a multilingual cohort, and the subgroup question in differential effects.

Keep a route to clarification. A brief conversation, an annotated diagram or an explanation of one key decision can show what a student understands when the prose will not carry it. Choose the format that fits the construct rather than the one that is easiest to grade.

Provide workable alternatives. Students need a clear path to human language support and accessible course resources. Requiring an untested tool can add a barrier while intending inclusion, which is the same failure support for disabled and neurodivergent learners has to avoid.

What will come back at you

"Translation means the student is avoiding the work." It depends on what the work is. Translation can enable engagement with disciplinary content it has no bearing on, and it can replace practice when independent target-language production is the outcome. Decide which you are assessing and say so.

"The AI fixed the language without changing anything else." Check it. A rewrite that improves grammar often changes emphasis, hedges a claim, or invents support. Ask the student to name one meaningful change and say whether it preserves their intended point, which also tells you whether they read the suggestion.

"Human markers are fairer." Human scoring carries its own language-related bias, and the multilingual writing-screening research finds automated systems reproducing patterns of it. Fairness here comes from explicit criteria, sampled subgroup evidence and a real chance to reconsider a judgment, not from who or what holds the pen.

"Our students are fine in English, so this does not apply." The finding that matters is not a deficit claim about any group. It is that conceptual understanding expressed in weaker prose is systematically under-scored, in every method tested. That applies to rushed writers, second-language writers, and anyone whose register differs from the examiner's.

What to do this week

1. Take one assignment and write a sentence stating what the language is doing in it.

2. Rewrite the instructions to say what support is permitted and what students must produce independently.

3. Test a translation and a terminology explanation in the languages your students actually use, and read the output for invented detail and dropped meaning.

4. Add one check that records the student's own meaning or decision, such as a two-sentence note on a change they accepted or rejected.

5. If your rubric scores writing quality, separate that criterion from the conceptual one and re-read two submissions of each with the split in mind.

Where to go next

Multilingual Learning is the concept page for language, translanguaging and assessment in AI-mediated teaching, and assessment validity is the standard for whether a score supports the inference you are making. Differential Effects Across Learner Groups covers how to look for subgroup patterns rather than assume them.

For the writing-instruction case where language is the construct, see best practices for writing instruction. For whether automated scoring can be trusted at all in your setting, see AI grading and checking whether an educational AI works.