AI activities suit ESL students best as extra practice around teacher-led instruction: rehearsing conversations, getting feedback on a draft the learner wrote first, and building quizzes from the learner's own word list. Research on chatbots for English learning is still mostly descriptive, so no activity here has a proven size of benefit. One finding is solid and matters for classrooms: a 2023 Stanford-led study found that seven AI-text detectors wrongly labeled 61.3 percent of TOEFL essays by non-native writers as AI-generated on average.
Key takeaways
- Activities work when the learner produces language first and the AI responds, not when the AI produces the language.
- A role-play partner, a draft reviewer, and a quiz builder are practical uses; each needs a human check on accuracy.
- AI-text detectors flagged non-native English writing far more often than native writing in a published test, so they are a poor basis for judging ESL students.
- Canada's adult ESL standard, the Canadian Language Benchmarks, has 12 levels, and a chatbot only knows the level it is told.
- Personal details such as immigration documents and addresses do not belong in prompts.
Activities that keep the learner producing language
Language learning depends on the learner speaking and writing, so the useful question about any AI activity is who does the producing. The table lists seven activities in which the learner does, with the AI cast as a partner, reviewer, or quiz writer. Each pairs a task with a check, because chatbots make mistakes in grammar explanations, invent idioms, and can change the meaning of a text they simplify.
| Activity | Skill | What the AI does | What the learner does | Check |
|---|---|---|---|---|
| Role-play rehearsal | Speaking, listening | Plays a pharmacist, landlord, or interviewer at a stated level | Answers aloud or in text, then lists phrases that caused trouble | Teacher or tutor reviews the list |
| Draft feedback | Writing | Points out up to three errors with short explanations | Writes the draft first, then rewrites it unaided | Compare with the teacher's comments |
| Personal error log | Grammar | Writes new practice sentences on a repeated error | Completes them and keeps the log | Confirm rules in a grammar reference |
| Own-word cloze quiz | Vocabulary | Turns the learner's word list into fill-in-the-blank items | Answers from memory | Verify answers in a dictionary |
| Level-adjusted reading | Reading | Rewrites a short passage in simpler English | Reads both versions and notes what changed | Meaning must match the original |
| Question rehearsal | Speaking | Asks follow-up questions on a topic the learner chose | Answers in full sentences | Record and replay |
| Peer comparison | Writing, discussion | Suggests corrections for two learners' versions of one text | Pairs argue which correction is right and why | Teacher settles disagreements |
The table is a planning aid, not a ranking of tools or methods, and it names no product on purpose.
Matching activities to a learner's level
A chatbot does not know whether a learner is a beginner or close to fluent unless the prompt says so. In Canada, the Canadian Language Benchmarks give teachers a shared vocabulary for this. The organization that maintains the standard, CCLB, describes it as a descriptive scale of ESL ability for adults, written as 12 benchmarks from basic to advanced, and a national standard for planning curricula. It also states that the benchmarks are not a curriculum, not tied to a teaching method, and not an assessment. An instructor can therefore place the benchmark in the prompt ("keep sentences short and use everyday vocabulary") but the benchmark number alone does not tell a chatbot what to produce. Whatever framework a school or program uses, the level has to be written into the request.
Why AI-text detectors are a risk for ESL writers
Teachers worried about copied AI text sometimes reach for detection software. A study led by Weixin Liang and James Zou, published in Patterns on July 10, 2023, tested seven widely used GPT detectors on 91 TOEFL essays by non-native speakers and 88 essays by US eighth graders. The detectors handled the US essays accurately but labeled more than half of the TOEFL essays as AI-generated, with an average false-positive rate of 61.3 percent. All seven agreed that 19.8 percent of the human-written TOEFL essays were AI-authored, and at least one flagged 97.8 percent of them. The authors found that the essays all seven detectors flagged had significantly lower text perplexity, a measure of how predictable the wording is.
The finding concerns the detectors tested in 2023, and tools change. It still gives a reason for caution: a detector score is weak evidence about an ESL student's work. Drafts written in class, early outlines, and a short conversation with the student about their choices give better evidence of authorship.
What the research can and cannot say
A 2024 qualitative synthesis in the journal ReCALL by Koç and Savaş reviewed 57 studies on voice-based chatbots for English learning published from 2010 to 2024. It mapped theoretical frameworks, study designs, learner reports, and teaching approaches. It was not a test of how much learners improve. Many studies in this area are short, use small groups, and rely on learner opinions, which is why this article does not promise gains. The sensible position for a classroom is a small trial: pick one activity, compare it with the usual practice for a few weeks, and keep what the teacher and learners find useful.
Privacy and fair use in class
Canada's federal privacy regulator released generative AI principles on December 7, 2023, and calls for particular consideration of vulnerable groups, including children and groups that have faced bias. Newcomer learners may be sharing language practice that touches on housing, work, or health. A safe habit is to practice with invented details and to leave out names, addresses, immigration numbers, and documents. Schools and programs also set their own rules about which tools students may use, and age limits differ by provider, so the teacher or program coordinator should confirm what is allowed before a class activity depends on a chatbot.
FAQ
Can AI replace an ESL teacher or tutor?
The research reviewed here does not support that. The activities above add practice time, but accurate correction, cultural context, and judgment about progress still come from a trained instructor.
Which activity is the easiest to start with?
The draft-feedback activity needs the least setup. The learner writes a short paragraph, asks for no more than three corrections with reasons, and rewrites it without looking, which keeps the writing with the learner.
Should an AI detector be used on ESL students' essays?
A published test found detectors wrongly flagging most TOEFL essays by non-native writers. A flag alone should not decide a case, and other evidence such as drafts and class writing is more reliable.
How can accuracy be checked when the learner cannot tell if the AI is right?
Use a second source: a dictionary, a grammar reference from the program, or the teacher. Learners can also ask the chatbot to show its reasoning and then check one claim per session.
Related reading
- AI in English language learning
- AI support for students with disabilities
- More AI activities for students
- Planning a study schedule with AI
More in AI in education.
Last verified: October 2026
Sources
- Liang et al., GPT detectors are biased against non-native English writers, Patterns, 2023
- About the CLB, CCLB (Canadian Language Benchmarks), 2026
- Koç and Savaş, The use of artificially intelligent chatbots in English language learning: a systematic meta-synthesis, ReCALL, 2024
- Generative AI principles for responsible and privacy-protective use, OPC (Canada), December 7, 2023