Or: what AI is actually good for, and where it quietly starts working against you
I use AI tools myself, and I recommend them to students, with real caveats. The honest version of this advice isn’t ‘avoid AI’ or ‘let AI do the work’, it’s understanding specifically what it’s good for and where it starts costing you more than it gives.
What it’s genuinely good for: low-stakes rehearsal. The single best use of AI for language learning is as a practice partner with zero social cost. You can ask it to simulate a job interview, a difficult negotiation, or a specific discussion topic, get an imperfect first attempt out, and try again immediately without embarrassment or a real person watching. This matters specifically because a lot of language anxiety comes from the fear of getting it wrong in front of someone, and AI removes that fear entirely for the rehearsal phase, before the stakes are real.
A genuinely useful technique: building your own practice material. Rather than relying on generic vocabulary lists, you can ask an AI tool to generate cloze-deletion exercises (fill-in-the-blank sentences) from a topic you actually care about, or pull contextual example sentences for a specific word you’re struggling with. This is a real improvement over static textbook exercises, because the material is built around what you actually need rather than a generic curriculum.
Where it quietly starts working against you. The failure mode isn’t using AI, it’s letting AI-generated text replace your own attempt entirely. If you ask for a polished paragraph and submit it as your own, you’ve skipped the actual skill, producing language under your own effort, and practised nothing. The tell is simple: can you say the sentence again, from memory, in a slightly different context, five minutes later? If not, you didn’t produce it, you copied it, and copying doesn’t build fluency no matter how good the copied text is.
Why human feedback still matters, and always will. AI feedback is genuinely useful for low-stakes rehearsal, spotting an obvious grammar slip, offering an alternative phrasing. It’s considerably weaker at the things that actually separate a good communicator from a merely correct one: register judgement in an ambiguous social situation, knowing when a technically correct sentence will land badly with a specific audience, the kind of nuanced read a human who’s seen you in real conversations builds up over time. Automated feedback can tell you a sentence is grammatical. It generally can’t tell you it’s the wrong thing to say to this particular client, in this particular meeting, and that judgement is exactly the part that separates real fluency from technically correct output.
The practical boundary. Use AI to rehearse, generate practice material, and get a fast first read on something low-stakes. Don’t use it to produce the final version of something that actually matters, an important email, an exam answer, a real negotiation, without your own effort somewhere in the process. The rehearsal can be AI. The performance has to be yours.
I’ve built a companion lesson plan for this one: rehearsal prompts to try live, a five-minute memory test that shows the difference between practising and copying, and a demonstration of building custom practice material. Download the AI as a Rehearsal Partner lesson plan (PDF).