AI for Differentiated Instruction: ESL Feedback Guide

The hardest part of differentiated instruction isn’t designing tiered tasks or grouping students by proficiency—it’s what happens after they submit their work. When a teacher receives writing samples from thirty students at four different levels, the clock starts ticking. Give everyone identical feedback and advanced learners are left unchallenged; write fully individualized responses for every student and you’re working until midnight. This gap between ambitious differentiation and realistic marking time is where most ESL teachers quietly compromise. AI is beginning to change that equation.

person holding on red pen while writing on book

Why Feedback Is the Hardest Part to Differentiate

Most differentiated instruction guides focus on inputs: how to write tiered tasks, how to scaffold instructions, how to group students effectively. Input differentiation is visible and plannable. Feedback differentiation is invisible and happens under time pressure after the lesson ends. A weak writer needs encouragement alongside targeted correction of their most recurring errors. A strong writer needs stretch comments that push their expression further. An intermediate learner needs something in between—validation of what’s working alongside clear guidance on their next step.

Writing that kind of feedback three different ways, for thirty students, every week, is not sustainable without support. This is the response gap: the distance between the feedback each student needs and the feedback each student actually receives. In most ESL classrooms, that gap is wide—not because teachers don’t care, but because time doesn’t stretch.

Research in second language acquisition consistently shows that the type of corrective feedback matters as much as its presence. Lower-level learners benefit most from explicit, targeted correction of a small number of high-frequency errors—too many corrections at once overwhelms and discourages. Advanced learners gain more from implicit feedback that nudges them toward natural, native-like expression. A single feedback template applied to every student in the class is, by definition, wrong for most of them. AI offers a path out of that compromise.

What AI-Assisted Feedback Looks Like in Practice

AI writing tools—including ChatGPT, Claude, and education-focused platforms like Grammarly for Education—can function as a first-pass feedback engine. The teacher’s role shifts from generating feedback from scratch to reviewing, adjusting, and personalizing AI-generated suggestions. The time saving is significant; the quality improvement is often surprising.

Two women working together, both are looking at the laptop screen.

Here is what a practical workflow looks like. After students submit a short paragraph, the teacher pastes the text into an AI tool with a level-specific prompt: “This student is a low-intermediate ESL learner. Give three corrective feedback points focused on subject-verb agreement and article use. Keep the language simple and encouraging.” The AI returns a draft response. The teacher reviews it in ten to fifteen seconds, adjusts the tone, adds a personal note, and sends it along. What used to take three to four minutes per student now takes under a minute—and the feedback is often more consistent because AI doesn’t get tired on student twenty-eight.

The key to making this work is writing strong prompts that specify the student’s level and the focus area. A prompt for an advanced student looks entirely different: “This student is at a B2 level. Identify one area where their writing sounds unnatural or overly formal for the context. Give a single example with a suggested revision.” Level-specific prompts produce level-specific feedback. Generic prompts produce generic responses that feel no different from a spell-checker—they satisfy no one and help no one.

Building a Prompt Library for Your Class Levels

The quality of AI feedback scales directly with the quality of the prompt. Teachers who get the best results develop a small library of level-tagged prompt templates—one for beginners, one for intermediate, one for advanced—and paste them in as needed. Over a few weeks, these templates improve as the teacher notices what kinds of AI feedback land well with each group and refines the language accordingly.

A slick business card mockup (www.brando.ltd)

A useful calibration exercise: take three student writing samples at different proficiency levels and run each through the same generic prompt, then through a level-specific one. Compare the outputs side by side. The level-specific version almost always produces feedback that is more relevant, more appropriately challenging, and closer to what the teacher would have written manually. After running this comparison even once, most teachers find they never return to generic prompts.

Differentiating Oral and Speaking Feedback

Written feedback is the easier half of the equation. Speaking feedback is where most differentiated instruction systems break down entirely, because recording and reviewing oral output at scale feels impossible in a normal working week. The result is that speaking practice is assessed largely through in-class observation, which is impressionistic, difficult to differentiate by level, and impossible to revisit.

Male on phone outside working on his computer.

AI transcription tools—Whisper (via OpenAI), Otter.ai, and similar platforms—make speaking feedback accessible in a way that wasn’t practical five years ago. Students record a two-minute spoken response on their phone after a lesson or as homework; the teacher uploads the audio for transcription; the resulting text becomes raw material for AI-assisted feedback using the same level-specific prompt approach used for writing tasks.

For lower-level students, the feedback focus is on intelligibility and basic fluency: Did they complete full sentences? Did they use the target vocabulary from the lesson? For intermediate learners, the focus shifts to accuracy and cohesion—are ideas connected logically, are transitions used correctly? For advanced students, feedback addresses register, precision, and natural expression. The transcript doesn’t capture prosody or intonation, but it captures enough for the teacher to generate meaningful written feedback on spoken production—something most ESL students rarely receive in any systematic way.

Building Adaptive Rubrics for Differentiated Assessment

Beyond student-by-student responses, AI can help teachers build rubrics that are differentiated by design. A standard rubric applies the same criteria to every student in the class. An adaptive rubric gives each level group criteria that match where they are and what they need to focus on next—making the assessment instrument itself a form of differentiation.

Go and vote! This is a german voting paper for a local election.

Prompting an AI to generate a tiered rubric takes less than two minutes: “Create a writing rubric for an ESL class with three groups: beginner (A1–A2), intermediate (B1), and upper-intermediate (B2). Include criteria for grammar, vocabulary, organization, and coherence. Write the language in each column so that students at that level can read and understand it themselves.” The AI produces a working draft. The teacher edits it for their specific context, shares it with students before the task begins, and uses it as the consistent frame for all subsequent feedback in that unit.

When students can see what good performance looks like at their level—not at some abstract universal standard—the feedback loop closes faster. They know what they are aiming for, and they are more likely to read and act on feedback because it speaks directly to their current stage of development rather than to an idealized endpoint they can’t yet picture.

Student Self-Assessment as a Differentiation Multiplier

There is a deeper opportunity here that goes beyond teachers using AI to generate feedback. Students themselves can be taught to use AI as a self-assessment partner before they submit work—shifting some of the differentiation work to the learner, where it arguably belongs for long-term language development.

A student writing on paper at a desk in a classroom

A lower-intermediate student finishes a paragraph and pastes it into an AI chat with the prompt: “I am a B1 English learner. What are the two most important grammar problems in this paragraph?” The AI responds with specific, targeted observations. The student fixes the issues and re-submits a stronger draft. The teacher receives work that has already been through one revision cycle—and the student has practiced metacognitive self-editing in the process, building a skill that transfers to every future writing task.

This approach requires some classroom setup. Students need a brief orientation on how to prompt for useful feedback—not just “fix my English” but “I am at this level, identify this specific type of problem.” They need to understand that the goal is to learn from the feedback and not simply apply corrections mechanically without engaging with why the change was needed. Many teachers ask students to submit their AI exchange alongside their final draft, so the thinking process is visible and can itself become a discussion point during class feedback sessions.

Done well, AI-assisted self-assessment builds student autonomy and reduces the teacher’s correction load simultaneously. It is one of the few places in differentiated instruction where the workload becomes more sustainable over time rather than more demanding—because students are developing a skill they apply independently, in any class, at any proficiency level.

Keeping the Human Element at the Center

AI feedback works best as infrastructure, not as the relationship. Students know when a comment comes from a teacher who read their work carefully versus when it came from an algorithm, and the motivational difference is real. The goal of AI-assisted feedback is not to replace teacher judgment but to free up teacher attention for the responses that matter most—the ones that require knowing the student, not just the text.

A practical division of labor: use AI to handle the mechanical correction load—grammar errors, vocabulary gaps, structural problems—and reserve your own writing time for responses to the student’s ideas, effort, and progress. “Your argument in the second paragraph is much stronger than last week—you’re starting to build from evidence instead of assertion” is a comment that no AI writes well, because it requires longitudinal knowledge of a specific learner. That is precisely where teacher time belongs.

The same principle applies to speaking feedback. AI can flag that a student’s transcript shows consistent omission of third-person singular -s. Only the teacher knows whether that student has been working on that exact pattern for three weeks, whether it is a new error, or whether the recording context introduced performance anxiety. Context is the teacher’s irreplaceable contribution—and with AI handling the mechanical load, there is more space to bring that context to bear.

A Starting Point for Skeptical Teachers

The biggest barrier to AI-assisted feedback is not technology—it is habit change. Teachers who have been marking papers the same way for ten years do not need a complete workflow overhaul; they need one contained experiment with enough structure to generate a real comparison.

A good entry point: pick one class, one writing task, and try AI-assisted feedback on just the grammar and vocabulary tier. Take the amount of time you would normally spend on that set of papers, use half of it for AI-assisted marking, and spend the other half writing personal responses to student ideas and visible progress. Compare the experience—and compare what your students say when they receive the feedback—to a week when you mark the same assignment without AI support.

Most teachers who run this experiment do not go back. The combination of reduced marking time and improved feedback specificity is difficult to walk away from once you have experienced it. The differentiation that felt unsustainable starts to feel manageable—and then, gradually, like the obvious default. The response gap does not have to stay wide.