Beyond the Chatbot: An SLA-Informed Imperative for AI in Language Learning

I recently watched a video of an intermediate Spanish learner in a 20-minute interaction with a state-of-the-art AI tutor. The bot was a marvel of engineering: instantaneous, grammatically flawless, offering perfect translations and subjunctive conjugations on demand. The learner, however, finished the session visibly deflated, their confidence more fractured than when they began.

The technology might seem exciting to someone not following closely the advancements in the filed, but the pedagogy was a travesty. This scene encapsulates the profound and dangerous disconnect at the heart of AI-driven language learning today: we are building tools that operate on linear, predictable logic, while language acquisition itself is a non-linear, complex, and deeply human process. We risk automating the very things that prevent learning.

The Great Forgetting: Ignoring a Century of MFL Pedagogy

To understand how we arrived at this juncture, we must remember where our field has been. Modern Foreign Language (MFL) pedagogy underwent a century of seismic shifts, evolving from the sterile exercises of Grammar-Translation and the behaviourist drills of Audio-Lingualism to the communicative revolution of the 1970s. This wasn't mere fashion; it was a paradigm shift, culminating in foundational principles now conspicuously absent from most AI tools.

FL teachers and experts in SLA may differ significantly in their methods or approaches. However, I am sure a significant unmber would agree with me in that we could consider these three hypothesis as common grou d. For exageration and simplification purposes (pardon me - I don’t think a full deep dive into SLA/FLT is where I’m trying to go today), we’ll call it the "holy trinity" of modern Second Language Acquisition (SLA) theory that underpins effective MFL teaching:

1. Krashen's Input Hypothesis: the bedrock concept that acquisition occurs through exposure to "comprehensible input" (CI) – language input understood by the learner, ideally at a level slightly beyond their current competence (i+1). Yet, many AI tutors bypass this crucial stage, forcing learners into immediate output or providing an "infinite sandbox with no curated sand," lacking the structured, comprehensible input necessary for acquisition.

2. Long's Interaction Hypothesis: refining Krashen, Long proposed that acquisition is facilitated when learners are actively involved in "negotiating for meaning." Communication breakdowns—clarification requests ("Sorry, what do you mean by...?"), comprehension checks, and confirmation checks—are not failures but potent catalysts for interlanguage development. Most AI chatbots, engineered for frictionless conversation, systematically eliminate these crucial interactional sequences. They comply or correct; they don't negotiate.

3. Swain's Output Hypothesis: Swain argued that producing language ("pushed output") compels learners to move from semantic to syntactic processing. It's in the struggle to be understood that learners notice gaps in their linguistic knowledge, test hypotheses about target language structures, and gain fluency. The critical issue is that AI often demands this output prematurely, before sufficient input processing or interactional competence has developed, leading to the rehearsal and potential fossilization of errors rather than genuine hypothesis testing.

These three pillars alone demonstrate that a simplistic "AI conversation partner" is a pedagogically naive construct. But the field's insights didn't conclude there.

Complexity, Systems, and the Social Turn

Some more recent theoretical developments render the current state of AI in language learning even more jarring:

• Sociocultural Theory (Vygotsky): learning is fundamentally social, occurring in the Zone of Proximal Development (ZPD) through scaffolding provided by a "More Knowledgeable Other" (MKO). AI is uniquely positioned to be an infinitely patient MKO, yet most implementations fail to provide true, dynamic scaffolding. They adjust difficulty with a blunt algorithmic slider, rather than the responsive, contingent support that defines the ZPD.

• Complexity Theory (Larsen-Freeman): a second language is not a static, accumulable body of knowledge, but as a complex, adaptive, dynamic system. An individual's linguistic system is emergent, non-linear, and exquisitely sensitive to context. Most current AI implementations, with its deterministic feedback loops and predictable responses, treats language like a closed system to be mastered, fundamentally misrepresenting its living, evolving nature. This explains why a learner can "succeed" with a bot for an hour yet feel their real-world communicative competence remains brittle.

Before proposing solutions, it's worth summarizing what learners practically need, derived from these theories: input at the right level, real-world tasks, feedback that matches the error, and manageable cognitive load.

A Principled Framework for SLA-Aligned AI

Given this wealth of knowledge, we must shift from asking "What can AI do?" to "What does SLA research compel AI to do?" This demands a framework built on core pedagogical principles:

---

Consider the practical application. A typical AI prompt: “Let’s talk about coffee. I’ll correct you,” is vague and intimidating. An SLA-aligned version: “You’re ordering a café con leche and a tostada con tomate at a busy Madrid café. Your goal is to also ask for the bill. I’ll keep my replies under 15 words, flag any major errors in {braces} for your review later, and we can practice any new phrases afterwards.” This provides context, comprehensible input expectations, a clear task, non-intrusive error flagging, and scaffolding for future learning.

AI as a "digital more knowledgeable other"

Imagine diagnostic AI that identifies not just CEFR level but also preferred learning modalities and cognitive styles. Picture AI that intelligently switches pedagogical approaches when a learner struggles, provides truly contingent scaffolding, and facilitates connection with human interlocutors. This is the potential of AI as a genuine "Digital MKO."

The risk is not merely ineffective products, but the systemic propagation of flawed pedagogical models that could set language learning back decades. The challenge is threefold:

• To & Product Managers: Look beyond your engineers. Embed learning scientists, applied linguists, and experienced MFL educators within your design and development teams from inception. Measure success not by engagement metrics, but by demonstrable gains in communicative competence and learner confidence.

• To educators: Be the guardians of our pedagogy. Critically evaluate the tools you recommend or use. Demand more than technological novelty; demand robust theoretical grounding and evidence of efficacy.

• To learners: Understand that if a tool makes you feel defeated, the failure is likely in its design, not in your ability. Seek tools that respect the complex, individual, and beautiful process of language acquisition. Ask if it knows your level, sets clear tasks, gives helpful feedback, and keeps you motivated.

We have the capacity to build AI that can finally deliver on the promise of personalized, theoretically-sound learning at scale. To do anything less is a profound failure of imagination and a disservice to our learners. The real question is whether we'll transform AI into something that actually helps humans learn languages—or just get better at talking to machines.