Artificial intelligence is already changing language education. It can create exercises in seconds, generate dialogues, correct texts, simulate conversations, produce audio, adapt difficulty and even give learners access to speaking avatars at any time. A few years ago, much of this would have seemed extraordinary. Today, it is becoming normal. And because these tools are improving so quickly, one question keeps appearing: if AI can explain, correct, speak and create content, do we still need language teachers? I think the answer is yes — but not because teachers are better at producing content. AI can already do that extremely well. The real value of a good teacher is knowing what content should exist, when it should appear, how difficult it should be and what the learner actually needs at that moment.
A good prompt starts before the prompt
There is a lot of discussion about prompt engineering. The logic seems obvious: better instructions produce better results. A vague request like “create an A1 Portuguese lesson” will usually generate something generic. A more detailed request — target level, objective, vocabulary limits, activity type, duration — will produce something better.
But there is a problem hidden inside that process. Before you can write a good prompt, you already need to know what you want. That is where pedagogy begins.
A prompt engineer may know how to make an instruction technically precise. They can ask for European Portuguese, CEFR alignment, pair work and a specific format. But why pair work? Why this vocabulary? Why this grammar point? Why now? Should the learner be speaking more or listening more? Should the activity be controlled or unpredictable? Which errors matter at A1 and which ones can wait?
These are not prompt engineering questions. They are teaching questions.
A good teacher can look at a learner and realize that another grammar explanation will not help. The learner already understands the rule; what they need is repetition. Another student may build accurate sentences but fail to understand native speech. A third may know enough vocabulary but still freeze because every sentence is being constructed consciously.
AI does not make these priorities meaningful by itself. It responds to the information and instructions it receives.
This is why a good prompt is often good not because someone understands AI better, but because someone understands learning better.
Compare “create an engaging B1 roleplay” with a much more pedagogical instruction: the learner already knows the past tenses but still hesitates when describing unexpected events; the objective is fluency rather than grammatical accuracy; the situation should involve a real problem in Portugal; vocabulary should remain familiar; and the student should need to explain, clarify and react to unexpected developments.
That second instruction works because someone already made the difficult decisions.
AI can generate, but cannot decide what good teaching is
AI can produce enormous amounts of language material. Give it a topic and it can create twenty dialogues, hundreds of vocabulary items and as many exercises as you want. But quantity has never been the main problem.
The difficult part is selection. Which twenty words matter now? Which five structures should be repeated? Which grammar point is necessary and which one can wait? What should the learner be able to do after the lesson that they could not do before? Those choices require judgement.
AI can suggest answers, but it has no classroom experience in the human sense. It has not watched a student lose confidence after being corrected too often. It has not noticed that a group understands a grammar explanation perfectly but cannot use the structure ten minutes later. It has not spent years discovering which activities generate real speaking and which ones only look communicative on paper. This distinction matters because educational content is easy to confuse with education itself.
A beautiful exercise is not necessarily a useful exercise. A realistic avatar is not necessarily a good teacher. A grammatically correct dialogue is not necessarily the right dialogue for that learner. And an activity that an AI system labels “A2” is not automatically appropriate for an A2 student. Someone still needs to judge it.
That is why an AI specialist cannot automatically replace a linguistic or pedagogical specialist. A prompt engineer may understand models and constraints extremely well, but knowing how to instruct a system is not the same as knowing what the system should be instructed to teach.
The same logic applies elsewhere. Someone who knows how to prompt a medical AI is not therefore a doctor. A person who can generate sophisticated architectural concepts does not automatically understand structural engineering. The tool lowers the barrier to producing an output. It does not automatically provide the knowledge needed to evaluate that output. Language teaching is no different.
Why AI language apps often feel more impressive than useful
This becomes particularly clear in applications built around AI conversation. The technology can be genuinely impressive. A learner speaks to an avatar, the avatar understands, answers in Portuguese and maybe corrects mistakes.
Technically, that is remarkable. But once the novelty disappears, the real questions begin.
What should the avatar talk about? How long should the conversation last? How much new language should it introduce? Should it correct every mistake? Which ones should it ignore? Should it simplify its language when the learner struggles? Should it deliberately recycle previous vocabulary? When should the conversation become less controlled?
These questions determine whether the avatar becomes a learning tool or simply an entertaining chatbot. And the avatar itself cannot answer them in a pedagogically reliable way without a methodology behind it.
This is one of the weaknesses of technology-led language products. The starting point is often: “We have an AI that can talk. How can we use it to teach languages?”
I think the order should be reversed. First ask: “How should this learner develop this skill?” Then ask: “Can AI help us do that better?” That difference changes everything.
AI avatars have enormous potential because one of the greatest limitations of traditional language learning is lack of speaking time. A teacher cannot talk individually to twenty students at once. An AI system can. But unlimited conversation is not automatically useful conversation.
If learners are simply placed in front of an avatar and told to speak, they may repeat the same safe vocabulary, practise recurring mistakes or stay inside their comfort zone. The educational value comes from the design around the conversation: scenario, objectives, progression, restrictions, repetition and feedback.
In other words, the intelligence behind the system is not only artificial intelligence. It is pedagogical intelligence.
Conclusion
AI will continue to improve, and language teachers should use it. It can save time, create opportunities that were previously impossible and make learning much more flexible. But its educational value depends on the quality of the decisions surrounding it. Prompt engineering can make AI follow instructions more accurately; it cannot replace the subject knowledge required to know which instructions are worth giving.
The future of language teaching will not belong to teachers who reject AI, nor to AI specialists who assume the technology makes pedagogy unnecessary. It will belong to teachers who understand learning deeply enough to use AI deliberately. A good prompt can make AI follow a method. It cannot invent the method for you.