A common ESL pattern: a student practises with an AI tutor for six months. Their AI-conversation performance is fluent, confident, accurate. They feel ready. Then they're in a real conversation - a job interview, a hostel common room, a customer service call - and they freeze. The other person said something unexpected. The student's practised script doesn't cover it. Fluency collapses.
This is the predictability trap, and it's a structural problem with AI practice that affects every ESL student who relies on it as their main speaking input. The trap is invisible during practice and obvious in real situations.
YapYapGo is a classroom speaking practice tool for ESL teachers, built around pair conversation between actual humans. This post is about what real fluency requires that AI practice can't deliver.What real fluency actually requires
Fluency is not just the ability to produce language at speed. It's the ability to navigate the unexpected. Real conversation is full of:
- Topic shifts. The other person changes the subject without warning.
- Tangents. A throwaway comment turns into a five-minute story.
- Misunderstandings. You asked X; they answered as if you'd asked Y.
- Humour and irony. Their tone is the joke; the words don't say it.
- Discourse markers in unexpected positions. "Anyway", "actually", "I mean" rearranging the meaning.
- Unknown vocabulary. They used a word you don't know; you need to keep up.
- Cultural references. A film, a politician, a meme you've never heard of.
- Pace changes. A slow conversation suddenly accelerates.
A fluent speaker handles all of this on the fly. A practised-with-AI-only student handles the first 30 seconds and stalls when the conversation diverges from the patterns they trained on.
The deeper point: real fluency is a repair skill, not a production skill. The ability to recover when something unexpected happens is what separates conversational competence from monologue competence. AI practice trains the latter; classroom interaction trains the former.
Why AI conversations converge to predictability
LLM-powered tutors are tuned to be:
- Helpful (so they don't surprise students with weird directions).
- Polite and supportive (so they don't expose mistakes harshly).
- Coherent across turns (so the conversation doesn't feel disjointed).
- On-topic (so the lesson stays focused).
Each of these tuning choices is reasonable in isolation. Stacked together, they produce conversations that are smooth, expected, and structurally unlike real human dialogue. The student gets used to conversations that follow scripts. They never get the experience of a partner who interrupts, misunderstands, jokes, goes off-topic, or runs out of patience.
The relevant post on what AI listening misses covers the related point about how AI tutors smooth over student errors. The smoothing is the same mechanism here - it removes the friction that drives real-conversation skill development.
A specific test
A useful experiment: have a student practise with their favourite AI tutor for 10 minutes, then pair them with a classmate who has been instructed to be slightly chaotic in conversation (changing topics, asking unexpected questions, going on tangents, occasionally not answering directly).
The student's performance in the two conversations will be measurably different. The AI conversation will look more fluent on the surface (smoother turns, fewer pauses). The classmate conversation will be messier - more hesitations, more clarification requests, more repairs.
The classmate conversation is the one closer to real-world fluency. The AI conversation is the practised script. Students who only train against AI develop fluency at the AI conversation and not at the messy human one.
What real classrooms add
A class of real students provides predictability-shock at no extra cost. The pair-work conversation is going to:
- Drift off-topic regularly.
- Hit vocabulary gaps the AI would have smoothed over.
- Misunderstand each other (both students might be reaching for half-known structures).
- Develop personal jokes and references that an outside listener wouldn't follow.
- Include the random energy of two human personalities interacting.
Every one of these is a feature, not a bug. Students train the repair skill, the tangent-handling skill, the cultural-reference-bridging skill - all of which they will use in real conversations and will never use against an AI tutor.
You can engineer this productively. Use the Team Maker to vary pair compositions across rounds (different partners stress different aspects of the unpredictability skill). Use the Topic Generator to pull in topics students don't expect. Use the Classroom Timer to add the time-pressure that real conversations have.
The point is not to be cruel about it. The point is to expose students to enough variability during practice that the real-world version isn't a shock. This is the same exposure logic as the accent variety argument and the slang and idiom argument - real classrooms provide breadth that AI tutors structurally can't.
Where AI predictability is actually useful
AI's predictability isn't always bad. Some legitimate uses:
- Early-stage practice. Beginners need consistency. Predictable AI conversations are reasonable training wheels.
- Test preparation. Standardised tests (IELTS, TOEFL) have predictable formats. Practising on a predictable AI tutor matches the test conditions.
- Specific skill drilling. When the goal is to master one structure or one vocabulary set, predictable AI practice is efficient.
- Asynchronous self-study. Between class sessions, AI is a patient partner.
What AI shouldn't be is the sole conversation partner an ESL student trains against. The fluency that develops from AI-only practice will not survive contact with real-world unpredictability, and the student will need real classroom conversation to bridge the gap.
The bottom line
AI conversations are too predictable to build real fluency. Real fluency is the skill of navigating the unexpected, and unexpected things barely happen with AI tutors. Real classroom pair work delivers the unpredictability automatically because human partners are unpredictable. Use AI for narrow controlled drilling; use real conversation as your main fluency-building input.
Sources:
- Tarone, E. (2007). Sociolinguistic approaches to second language acquisition. Modern Language Journal.
- Skehan, P. (1998). A Cognitive Approach to Language Learning. Oxford University Press.
- Segalowitz, N. (2010). Cognitive Bases of Second Language Fluency. Routledge.