What 2026 taught us about AI writing feedback for language learners

2026-07-09

TrustAI Writing Lab is a writing platform for English learners. It's built around the CEFR (the A1-to-C2 proficiency framework that language schools, coursebooks, and exams like IELTS and Cambridge English already run on) so that feedback, grading, and curriculum all share one model of what a learner can do.

Before deciding what to build next, we reviewed the market in depth: the AI writing-feedback tools schools are adopting, the platforms built specifically for language learners, what teachers say in their own communities about these tools, and where academic integrity policy is heading. This report summarizes what we found and the principles it left us with.

Key findings

• Generic AI essay feedback became free at platform scale. Google now ships AI-suggested feedback inside Google Classroom at no cost. What has not been commoditized is feedback matched to a particular learner's proficiency level.
• AI-writing detection failed language learners specifically. A 2023 Stanford study found detectors flagged an average of 61% of essays by non-native English writers as AI-generated while almost never flagging native speakers. Universities have been disabling detectors, and vendors are pivoting from verdicts to process evidence.
• Teachers keep AI as a drafting assistant and abandon it as a replacement. Surveys find most teachers now use AI somewhere in their work, but in one 2025 survey only 13% of AI-using teachers let AI grade even low-stakes work.
• No product we reviewed adapts its feedback to the learner's level, neither in what it targets nor in how it's worded. This is the gap we're building against.

How we did this research

This is a desk review conducted over the first half of 2026, drawing on four kinds of public sources: product documentation and announcements from writing-feedback and academic-integrity vendors; peer-reviewed and preprint research on AI feedback and AI-text detection; published surveys of teachers and students (RAND, College Board, EdWeek); and teachers' own accounts in professional media and community discussions. We did not run primary studies of our own. Every empirical claim below is footnoted to its source; limitations are discussed at the end.

Generic AI feedback became free this year

In 2026, "AI can comment on an essay" stopped being a product. Google now ships AI-suggested feedback on written assignments directly inside Google Classroom, at no cost, to the largest install base in education.¹ Dozens of tools draft rubric comments in seconds.

What did not get commoditized is feedback that's right for a particular learner. Reporting and teacher accounts keep describing the same failure modes in generic AI graders: comments on sections that don't exist in the essay, scores that drift across a class set, and measurable bias against students who don't write like native English speakers.² A tool that can't tell a B1 learner from a C1 learner gives both of them the same feedback, pitched at neither.

AI detection collapsed, and language learners paid the price

The most consequential shift we found is what happened to AI-writing detection. In 2023, Stanford researchers (Liang et al.) tested seven widely used GPT detectors and found they misclassified an average of 61% of TOEFL essays written by non-native English speakers as AI-generated (97% of those essays were flagged by at least one detector), while essays by native-speaking US students were almost never flagged.³ The mechanism matters: detectors punish exactly the vocabulary and sentence patterns language learners are taught to use. Turnitin acknowledged cases of higher false-positive rates in its own detector,⁴ and a growing list of universities, with Vanderbilt among the first in August 2023, disabled AI detectors entirely.⁵⁶

What's replacing detection is process evidence: showing how a piece of writing came to be, rather than scoring how "AI-like" it reads. Turnitin now sells Clarity, a composition environment that captures draft history,⁷ GPTZero pivoted to writing replay,⁸ Cadmus markets process-based integrity to universities,⁹ and Grammarly built its Authorship feature into Canvas.¹⁰

In our view this is the right direction, and it matters doubly for English learners, the population detectors misclassified most. A draft history a student and teacher can look at together is evidence. A percentage score is an accusation.

Teachers keep assistants and abandon replacements

Teacher communities were the most clarifying source in our review. RAND's survey panels find AI use in schools rising quickly while guidance lags behind,¹¹ and College Board research reports that a majority of high-school students already use generative AI for schoolwork.¹² But teachers police a bright line: AI may draft feedback; it may not own grades or speak in the teacher's place. In a 2025 EdWeek survey, only 13% of AI-using teachers said they let AI grade even low-stakes work.² What teachers say they want is guidance, not outsourcing: AI that highlights evidence against their rubric and leaves both judgment and voice to them.¹³

The tools teachers abandon are the ones that break this contract, waste their time fixing bad output, or, as one teacher argued in a widely shared essay, corrode trust between them and their students.¹⁴ The tools they keep make one thing affordable that never was before: a real revision loop. Write, get feedback, revise, resubmit: the pedagogy every writing teacher believes in and few have time to run for 120 students.

Feedback has a reading level too

Here's the gap that surprised us most. Across every product we reviewed, including the tools built specifically for English learners, feedback comes in exactly one register. A beginner and an advanced learner get comments written in the same English, at the same complexity, about the same kinds of issues. Even Cambridge Write & Improve, the most CEFR-native tool we reviewed, gives writers a level estimate but is thinnest exactly where learners need the most support: depth of feedback for A2–B1 writers, and feedback on whether the writing actually accomplished the task, not just whether the grammar is clean.¹⁵

Meanwhile, current research, including work from Cambridge's own labs on assessing which vocabulary a learner deploys at which CEFR level,¹⁶ points at what proficiency-aware feedback could look like: knowing which words and grammar structures a learner attempted versus mastered, in context, and meeting them one step ahead of where they are. In our review, no commercial product puts that into a classroom feedback loop today. That's the product we're building.

The principles we're building by

The findings above translate directly into commitments, and we expect to be held to them:

• A teacher reviews AI feedback before students see it. On graded work, AI drafts and the teacher decides. The teacher's name is on the feedback because the teacher owns it.
• Feedback quotes the student's actual writing. No comments on paragraphs that don't exist. If the AI can't point to the sentence, it doesn't get to make the claim.
• Feedback is pitched at the learner's level, in what it targets and in how it's worded, and covers whether the writing achieved the task, not just its grammar.
• Provenance, never verdicts. We show draft history and how a piece of writing developed. We will not ship an "AI-probability" score, and we won't do keystroke surveillance.
• Growth is the point. Revision is the default assignment structure, and learners can see the structures and vocabulary they've mastered grow over time.
• We'll show our work. We're building an evaluation of our scoring against experienced teachers' ratings, and we'll publish the numbers, agreement rates included, the way serious assessment organizations do.
• Student writing is not training data. Clear retention terms, clear deletion, no training models on learners' work.

Limitations

This is a review of public sources, not primary research, and it has the weaknesses of one. Product capabilities are drawn partly from vendor documentation, which is self-reported. The teacher-community evidence is qualitative and subject to selection effects: teachers with strong experiences write about them. Our sources skew US-centric, while our users don't. The space is moving quickly, so the findings are dated to mid-2026. And we are a vendor with a stake in the conclusions; we've cited primary sources throughout so readers can check our reading against theirs.

We'll be writing more here as we build. If you teach English learners and want to shape what this becomes, we'd like to hear from you.