Why don't AI tutors teach current slang?

For two reasons. First, LLMs have training cutoffs and language evolves faster than the cutoffs catch up. Second, AI models are deliberately tuned toward neutral, professional register to avoid causing offence or appearing dated. Both factors mean the slang and informal idioms students actually hear in real conversation are systematically underrepresented in AI tutor output.

How important is slang for ESL learners?

Much more important than the textbook tradition suggests. Informal English is what students encounter in films, social media, casual conversation, and most workplace banter. A student fluent in textbook English and lost in slang will sound stilted and miss large parts of real interaction.

Where can ESL students learn current English idioms?

From people, not from books or AI. Real conversation with native or fluent speakers is the high-bandwidth channel for idiom transmission. Classroom pair work with proficient peers also works. Self-study from idiom dictionaries is the slow, lossy fallback.

Are idioms the same as slang?

Overlapping but distinct. Idioms are conventionalised figurative expressions (raining cats and dogs, kick the bucket) often with long histories. Slang is informal, often generational vocabulary (slay, bet, no cap) that changes faster. Both are underrepresented in AI tutor outputs.

← All posts

Pedagogy

Slang, Idioms, and Why Real Conversation Beats AI Practice

5 min read

The problem with AI ESL tutors and informal English is that the AI is structurally biased away from exactly the language students need. Students show up wanting to learn how their favourite Netflix show actually talks. The AI gives them a polished, professional, register-neutral English that sounds correct, sounds clear, and sounds nothing like the show.

This is not a flaw teachers can fix by prompting the AI better. It's baked into how LLMs are trained, tuned, and deployed. Real conversation between humans is the only reliable channel for the current, informal English students need.

YapYapGo is a classroom speaking practice tool for ESL teachers, designed around pair conversation between actual humans. This post is about why AI tutors lag on slang and idioms, and what to do about it.

What "informal English" actually contains

When ESL students say they want to learn "real English", they usually mean some combination of:

Current slang. Generationally fresh vocabulary that changes year by year. "Slay" / "bet" / "no cap" today; some other set in 2027.
Idioms. Conventionalised figurative expressions. "Hit the books", "ball is in your court", "under the weather".
Phrasal verbs. Verb-particle combinations that don't decompose semantically. "Put up with", "look forward to", "give in to".
Discourse markers. Connectors that real speech is full of. "Anyway", "actually", "honestly", "I mean", "like".
Register shifts. Knowing when to drop "would you mind" for "can you" or "gimme".

This bundle is what makes a fluent speaker sound fluent and an upper-intermediate textbook student sound like a textbook. The bundle is what AI tutors systematically underdeliver on. (For the broader question of what speaking English actually means, see the related post.)

Why AI tutors lag on slang

Three structural reasons:

Training cutoffs. Every LLM was trained on text up to a specific date. Slang that emerged after the cutoff doesn't exist in the model. Even after the model gets retrained, there's a lag of months between when slang enters circulation and when the model can use it confidently.

Deliberate register neutralisation. AI companies tune their models toward professional, broadly inoffensive register to reduce liability, brand risk, and offence. Slang and informal idioms are deprioritised in tuning. Even when the model knows the slang, the tuning pushes it back toward neutral.

Misalignment with classroom marketing. AI ESL products are marketed to teachers and parents who often want "correct" English. Slang is perceived as a risk to that brand promise, so products are tuned away from it. The user experience reinforces the conservative register.

The combined effect: AI tutors produce English that sounds slightly old-fashioned, slightly formal, and slightly off the mark for real conversational use. Students who train exclusively against AI absorb the same register bias.

Why classroom pair work doesn't have this problem

Real students share live language. A class of 20 international students has 20 sources of current slang from 20 backgrounds, 20 hours per week of streaming exposure to current shows, and 20 informal social channels where current English is being produced.

When these students do pair work, the slang circulates. Student A used "lowkey" in a sentence. Student B asks what it means. Student A explains. Student B adopts it. This is the uptake process that language acquisition research has been describing for 40 years - new lexis enters when the learner notices it, asks about it, and incorporates it. AI tutors don't introduce slang for students to notice; real conversation does it automatically.

Even when the class is homogeneous (all students from one country), the variety is still richer than AI output. Different students have different exposure profiles, different favourite shows, different friend groups. The cross-pollination during pair work is constant and free.

You can structure this by setting up pair rotations using the Team Maker, giving the Topic Generator prompts that pull toward informal register ("Tell your partner about a TV show you can't stop watching"), and using the Classroom Timer to keep things moving. The slang will surface; students will notice; uptake will happen.

A specific test: ask an AI tutor about "lowkey"

A worthwhile experiment: ask any AI ESL tutor to use "lowkey" naturally in three different sentences, then have a real classroom of late-teen or twenty-something students do the same.

The AI will produce sentences that look correct but read slightly stilted ("I am lowkey excited about this development"). The students will produce live, idiomatic, generationally-current usage with the discourse markers and intonation patterns that go with it.

The gap is not subtle. Anyone can run the test in 10 minutes and see it. The gap also tells you which channel is doing the real teaching work on informal English.

Where AI tutors still earn their slot

This isn't a wholesale dismissal of AI. AI tutors are useful for:

Grammatical practice. Verb tense drills, article use, conditional structures. AI is patient, consistent, and gives instant feedback.
Vocabulary review. Drilling word lists, definitions, collocations. AI handles this well.
Pronunciation feedback. AI listeners give immediate response on segmental pronunciation. (We've covered the limits of AI listening elsewhere - it's a narrow but real feedback channel.)
Solo practice between classes. Students who want extra practice have AI as an always-available partner.

The pattern is the same we've described for AI listening exposure and predictability in fluency development: AI for narrow targeted drilling; humans for the broad messy register that real fluency needs. Both, not either.

The bottom line

AI tutors lag months or years behind current spoken English, deliberately neutralise informal register, and produce students whose grammar is correct and whose slang is dated. Real classroom conversation delivers current, informal English automatically through peer exchange. Place AI in your stack for grammar and vocabulary; place real pair work in your stack for everything that makes a speaker sound human.

Sources:

Long, M. H. (1996). The role of the linguistic environment in second language acquisition. Handbook of Second Language Acquisition.
Schmidt, R. (1990). The role of consciousness in second language learning. Applied Linguistics.
McCarthy, M., & O'Dell, F. (2017). English Idioms in Use. Cambridge University Press.

Ready to try it in your classroom?

YapYapGo is free to start — no account needed. Set up your first speaking session in under a minute. New to YapYapGo? Read the overview.

Start for free →