Speaking with Ghost

“We are no longer practicing for the conversation. The simulation is the conversation.”

For decades, the “Holy Grail” of Computer-Assisted Language Learning (CALL) was a machine that could talk back. We tried with cassette tapes in the 70s, CD-ROMs in the 90s, and early voice assistants like Siri in the 2010s. But these interactions were brittle. They were “turn-based” games: you spoke, the machine processed (spun a wheel), and then it delivered a canned response. It was speaking at a computer, not with one.

In 2026, the arrival of multimodal, “Latency Zero” models has fundamentally altered the landscape of oral proficiency. We are now “Speaking with Ghosts”—entities that possess no body, no vocal cords, and no consciousness, yet can replicate the prosody, pacing, and interruption patterns of a human interlocutor with uncanny precision.

This article explores how to harness this spectral fluency without letting our students lose touch with the messy, unpredictable reality of human speech.

Latency Zero: The Psychology of Instant Voice-to-Voice

The defining feature of human conversation is not vocabulary; it is timing.

According to the seminal work on conversational turn-taking by Sacks, Schegloff, and Jefferson (1974), the average gap between one person stopping and another starting is roughly 200 milliseconds. This is faster than the human brain can cognitively plan a full sentence. We achieve this by “predicting” the end of our partner’s turn.

Until roughly 2024, AI models had a latency of 2 to 5 seconds. This “lag” destroyed the illusion of presence. It forced the student into “dictation mode”—speaking clearly, slowly, and unnaturally to ensure the machine understood.

With the release of models like GPT-4o and Gemini Live (OpenAI, 2024; Google, 2024), latency dropped below 350ms. This is the “Latency Zero” threshold.

The Pedagogical Shift:

 * From Monologue to Interruption: In older ELT classrooms, we taught students to wait their turn. In the Latency Zero era, we must teach Back-channeling (listening noises like “uh-huh,” “really?”) and Interruptibility.

 * The “Barge-In” Protocol: Modern AI agents allow students to cut them off. Teachers should design activities where the student must interrupt the AI to correct a fact or change the topic, mimicking the chaotic flow of real argument.

As distinct from the “wait-and-listen” drills of the audio-lingual method, Latency Zero creates a Flow State (Csikszentmihalyi, 1990) where the learner forgets they are practicing grammar and focuses entirely on maintaining the rhythm of interaction.

Lowering the Affective Filter through AI Avatars

Krashen’s (1982) Affective Filter Hypothesis remains the most durable concept in Second Language Acquisition (SLA). It posits that anxiety acts as a mental block, preventing input from reaching the language acquisition device.

The classroom “speaking circle” is often a high-anxiety environment. The fear of peer judgment (“social evaluation threat”) can silence even competent learners. Here, the AI acts as a Digital Sandbox.

Why the “Ghost” is Safe:

 * No Social Cost: An AI cannot judge you. It has no memory of your embarrassing mistake yesterday. It does not roll its eyes.

 * Infinite Patience: A human teacher creates anxiety when they have to repeat a correction for the third time. An AI will repeat it for the thousandth time with the exact same emotional tone.

Recent studies on “Foreign Language Anxiety” (FLA) in human-robot interaction suggest that learners are willing to take significantly higher risks with vocabulary and syntax when speaking to an anthropomorphic agent than with a human peer (stockwell, 2022; Fryer et al., 2020).

Classroom Application: The “Pre-Game” Warmup

Before a high-stakes human debate or presentation, students spend 10 minutes debating the AI. This “burns off” the initial cortisol spike and allows them to “warm up” their articulatory muscles (phonology) without the risk of social failure.

The Human Premium: Pragmatics, Irony, and Nuance

If the AI can converse fluently, what is left for the human teacher? The answer lies in the Pragmatics of Ambiguity.

While AI has mastered syntax (the rules of structure) and semantics (the meaning of words), it still struggles with pragmatics (language in use). As Taguchi (2015) notes, pragmatic competence involves understanding not just what is said, but why it is said in a specific context.

The “It’s Cold in Here” Test:

 * The Machine: If you tell an AI, “It’s cold in here,” it might interpret this as a meteorological fact or a request for data on heating systems.

 * The Human: A human understands this as an indirect speech act—a polite request to close the window (Grice, 1975).

Teaching the Premium:

We must shift our curriculum from “Correctness” to “Appropriateness.”

 * Irony and Sarcasm: AI models are notoriously literal. Teachers must design lessons on “Reading the Room,” using video clips or role-plays where the tone contradicts the text.

 * Phatic Communication: Small talk is not about information exchange; it is about social bonding (Malinowski, 1923). AI is bad at this because it has no social needs. Students must practice “useless” talk—gossip, banter, complaining about the weather—with humans to learn the glue of social solidarity.

The “Human Premium” is the ability to navigate the unwritten rules of culture that the training data of the LLM often flattens or misses entirely.

Toolkit: Designing High-Stakes Interactional Vivas

In the age of AI, the written essay is suspect (see Chapter 1), and the multiple-choice test is obsolete. The most robust assessment in 2026 is the Interactive Viva Voce (Oral Defense).

However, we cannot simply ask “What did you learn?” We must test Strategic Competence (Canale & Swain, 1980)—the ability to repair communication breakdowns.

The “Saboteur” Protocol

In this assessment, the student interacts with an AI agent (monitored by the teacher) or the teacher themselves. The interlocutor is instructed to be difficult—not grammatically, but interactionally.

The Rubric for the Viva:

Skill Domain	Assessment Criteria (The Human Premium)
Repair Strategies	Does the student freeze when interrupted, or do they use phrases like “Let me rephrase that” or “As I was saying…”?
Clarification Requests	When the “Saboteur” uses a vague term, does the student ask: “By ‘efficient’, do you mean cost-effective or time-saving?” (AI rarely asks for clarification; humans must).
Turn-Taking Management	Does the student handle “barge-ins” gracefully? Do they know how to hold the floor using fillers (“Well…”, “You see…”)?
Pragmatic Force	Does the student adjust their register? (e.g., shifting from casual to formal when the “Saboteur” becomes aggressive or official).

Sample Prompt for the AI Agent:

> “Act as a skeptical landlord. I am a tenant trying to negotiate a lower rent. Interrupt me frequently. Misunderstand my politeness as weakness. Force me to be firm but polite.”

>

By assessing the student’s ability to navigate friction, we measure the one thing the AI cannot replicate: the messy, emotional resilience required to communicate with another human being.

References

 * Canale, M., & Swain, M. (1980). Theoretical bases of communicative approaches to second language teaching and testing. Applied Linguistics, 1(1), 1-47.

 * Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience. Harper & Row.

 * Fryer, L. K., Nakao, K., & Thompson, A. (2020). Chatbot learning partners: Connecting learning interest, engagement and anxiety in EFL course challenges. Computers & Education, 154, 103930.

 * Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41-58). Academic Press.

 * Krashen, S. (1982). Principles and Practice in Second Language Acquisition. Pergamon Press.

 * Malinowski, B. (1923). The problem of meaning in primitive languages. In C. K. Ogden & I. A. Richards (Eds.), The Meaning of Meaning. Routledge.

 * Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696-735.

 * Stockwell, G. (2022). Mobile Assisted Language Learning: Concepts, Contexts and Challenges. Cambridge University Press.

 * Taguchi, N. (2015). “Contextually” speaking: A survey of pragmatic learning abroad, in class, and online. System, 48, 3-20.