Using AI in the English classroom without outsourcing your judgement

There are two unhelpful conversations about AI in language teaching. One says it changes everything and teachers who resist will be replaced. The other says it is a plagiarism machine and should be banned. Neither survives contact with a real classroom of thirty-eight students.

Here is a more useful frame: AI is very good at producing plausible text quickly, and language learning is largely about the effortful production of text. That tension is the whole problem, and also where the opportunities are.

Where AI genuinely helps a teacher

These are the uses where the time saved is real and nothing is lost.

Differentiating a text you already chose. You have a reading at B1 and six students who need A2. Ask for a simplified version, then check it yourself — models routinely simplify vocabulary while leaving syntax untouched, or drop the exact detail your task depends on.

Generating volume for practice. Twenty gap-fill items on the past perfect, ten dialogue prompts about a topic your students actually care about, five versions of a rubric-aligned prompt. Volume is the thing AI produces cheaply and teachers produce expensively.

First-pass rubric feedback for you, not for them. Paste an anonymised set of student texts and ask which errors recur. You are not delegating the grading; you are finding the pattern faster so your next lesson targets it.

Reformatting. Turning your lesson plan into a worksheet, your notes into a slide outline, your rubric into student-facing language. Administrative work, not pedagogical work.

Rehearsing difficult conversations. Practising, in advance, how to explain a grade to a parent or a policy to a coordinator.

Where it quietly damages learning

Feedback that replaces the struggle. A student who gets a corrected version of their paragraph has learned very little. A student who gets three questions about their paragraph has learned something. When you use AI for feedback, ask it for questions, not corrections.

Model texts that flatten voice. AI writing is fluent and characterless. If students see only AI-produced models, they learn to imitate an average. Use real texts — including texts written by previous students, with permission.

Assessment you can no longer trust. Any homework that consists of producing text at home is now a test of whether the student chose to write it. This is not a reason to panic; it is a reason to move the assessment of production into the room.

Cultural and linguistic bias. Models are trained overwhelmingly on English from the United States and the United Kingdom. Ask for "a dialogue at a restaurant" and you get a scene that may not match your students' lives at all. In a country teaching English across very different territories, a generic model is a poor default.

Confident wrongness about your context. Ask about Colombian language policy or the ICFES and you may get an answer that sounds authoritative and is invented. Never pass on an AI claim about rules, dates or requirements without checking the source.

Redesigning assessment, concretely

You do not need to ban anything. You need tasks where using AI does not remove the learning.

• Move drafting into class. Twenty minutes of handwritten or in-class drafting, then AI or dictionary use permitted for revision at home, then a short oral defence of the changes.
• Assess the process. Collect the outline, the draft and the final. Grade the movement between them.
• Ask for the local and the personal. "Explain this concept using an example from your neighbourhood" is much harder to outsource than "write about pollution".
• Add a two-minute oral check. A student who wrote their text can talk about it. One who did not, cannot. This is the single most reliable and cheapest check available.
• Let students declare their use. A one-line note — I used a model to rephrase paragraph two — turns hiding into transparency. Grade the work, not the confession.

Talking to students about it

Say plainly what AI is: a system that predicts likely text. It does not know things, and it does not know you. Then explain the deal in your class: here is where it is allowed, here is where it is not, here is why.

The "why" matters more than the rule. Students accept restrictions they understand. "You will not build the ability to think in English if a machine does the thinking" is an argument. "It is cheating" is a label.

It is also worth being honest about the world they are entering. Many of them will use these tools professionally. Learning to use them well — checking, editing, knowing when the output is wrong — is part of what school owes them.

A word about your own workload

Colombian teachers are not short of ideas. They are short of hours. If AI gives you back two hours a week of formatting and item-writing, spend it on the two things no model can do: knowing which student is struggling, and deciding what to do about it.

That is the line worth holding. Automate the production of materials. Do not automate the judgement.

Keep going

• AI is one of the three threads of the 61st ASOCOPI Annual Conference — saberes glocales, diversities and AI.
• HOW Journal's current issue includes a systematic review of AI-powered natural language processing in language education. See HOW Journal.
• The webinar library has recorded sessions on AI, digital tools and gamification.