Your AI tutor is making you worse at English — and what errorless teaching does instead

Quick answer

AI English tutors are good at fixing surface grammar errors in your writing — and that is exactly the problem. They correct output you have already produced without making you produce the corrected form yourself. Over time, the patterns you need under pressure stay fossilized while the polished output keeps coming back. The alternative is errorless teaching: scaffolded production rehearsal that builds the target pattern in your own voice before you deploy it.

Most senior professionals using ChatGPT to clean up their English think they are getting sharper. They are not. They are getting more polished output while practicing a different skill than they think they are practicing. The output looks better; the writer underneath does not change.

Worse, the writer's existing patterns — the ones running on autopilot in board meetings, negotiations, and high-stakes correspondence — calcify around the errors AI keeps quietly fixing.

This post is about what that gap costs and what the pedagogy that closes it looks like. By the end you will be able to spot the AI English tutor habit in your own writing and recognize errorless teaching as the principled alternative to it.

The conventional view — and why it isn't wrong about what it measures

The argument for AI as an English tutor is not made up. It rests on findings the 2024–2025 research record genuinely supports.

Recent work on GPT-4 with chain-of-thought prompting shows the model outperforms commercial automated written corrective feedback (AWCF) tools on both accuracy and consistency of error detection (Wu et al., 2025). A twelve-week comparative study of sixty Chinese students of English as a Foreign Language (EFL), published in Innovation in Language Learning and Teaching, found that the group receiving generative AI (GenAI)-only feedback improved measurably on grammar and sentence variety over the study window (Zhang et al., 2025).

If you have been pasting paragraphs into ChatGPT and watching cleaner versions come back, the cleaner versions are not an illusion. The tool is doing what it claims to do at the level it operates on.

The empirical conversation about AI feedback in second-language writing is not a fight between believers and skeptics. The evidence for short-term, surface-level effectiveness is real. So far, so concession. The trouble is what the same studies do not measure — and what the field's own systematic reviews explicitly say has not yet been measured.

Why surface accuracy isn't acquisition — and what the field hasn't measured yet

The simplest way to see what's missing is to ask what the AI is doing for you when it corrects a sentence — and what it is asking you to do back.

The AI produces the corrected version. You read it. Sometimes you paste it. Sometimes you note the change. What you almost never do is produce the corrected form yourself, on demand, before deployment.

That asymmetry is not a small detail. It is the entire mechanism by which second-language acquisition (SLA) either happens or doesn't. Merrill Swain's output hypothesis identifies three functions of producing language that comprehensible input alone cannot fulfill (Swain, 1995):

• Noticing the gap between what you intend to say and what you can actually produce.
• Hypothesis-testing through actual use — trying a form and seeing how it lands.
• Metalinguistic reflection — thinking about the language itself while using it.

Reading a correction is comprehensible input. Producing the correction yourself is output. They are not interchangeable.

When the production gap persists, the long-arc consequence has a name. Larry Selinker introduced interlanguage — the systematic linguistic system a learner constructs en route to a target language — and named fossilization as the persistent retention of non-target features in that system, observable across learners regardless of motivation or further instruction (Selinker, 1972). ZhaoHong Han's synthesis of four decades of SLA research argues that fossilization is best understood as a stabilization phenomenon affecting specific features rather than the whole interlanguage, and that once a feature stabilizes, it is robust against further input, instruction, or motivation (Han, 2004).

The AI English tutor habit is, in this frame, the most efficient possible delivery vehicle for fresh input that the learner never has to produce — precisely the conditions under which fossilized features stay fossilized.

The asymmetry matters more for adult learners than for younger ones. Robert DeKeyser's analysis of critical-period effects shows that adult-onset L2 learners exhibit greater persistence of L1-influenced patterns and reduced sensitivity to negative evidence — meaning the same fossilization mechanism is both more probable and harder to reverse (DeKeyser, 2000). Han's work treats the adult-versus-younger asymmetry as a distinct empirical question, not a footnote (Han, 2004).

For B2+ professionals — fluent enough to operate in English daily, with patterns already automatized at the level of register — the patterns at risk from the AI-correction shortcut are precisely the ones least amenable to retroactive intervention.

Here is where the AI feedback literature deserves credit for honesty. The twelve-week comparative study cited above found GenAI-only feedback improved grammar and sentence variety — but the same study found limited impact on higher-order writing skills such as critical thinking and organization within its window (Zhang et al., 2025). A 2025 systematic review of twenty-five PRISMA-selected studies on AI-driven feedback in ESL writing puts it directly:

"Few longitudinal studies examining their long-term impact on writing development."

— Sharmithashini & Hashim, 2025

The fossilization question is not a CTL invention. It is a research-acknowledged gap in a literature that has measured short-term surface effectiveness and has not yet measured long-term retention.

The reframe — errorless teaching scaffolds production on the first attempt

The alternative to receptive correction is not more correction. It is a different pedagogy entirely, one with sixty years of theoretical foundation and a clear application to adult L2 work.

Errorless teaching is rooted in Applied Verbal Behavior (AVB), the field that derives from B.F. Skinner's analysis of language as operant behavior (Skinner, 1957). The functional categorization of verbal operants — mand, tact, echoic, intraverbal, autoclitic — underpins how errorless approaches structure learning to make correct responses highly probable from the start.

The principle was sharpened in O. Ivar Lovaas's Discrete Trial Training (DTT) research, published in the Journal of Consulting and Clinical Psychology, which demonstrated the effects of structured, high-prompt-probability trials in intensive behavioral intervention (Lovaas, 1987). Mark Sundberg formalized the modern assessment framework — the Verbal Behavior Milestones Assessment and Placement Program (VB-MAPP) — that embeds the same principle: every milestone is targeted with sufficient scaffolding that the learner produces the correct form before independent production is even assessed (Sundberg, 2008).

Across this lineage, the core move is the same: the learner is set up to succeed before being asked to perform, and the error is never encoded in the first place because it never gets produced.

CareerTalkLab (CTL) applies this lineage to adult L2 Business English through a four-phase instructional structure, described in detail in our pedagogy reference.

1. Pre-teach the target form with full prompts.
2. Demonstrate it in context, with prompts still visible.
3. Practice with diminishing prompts as the form internalizes.
4. Produce independently — no prompts, the learner doing the work.

The point is not to teach beginners. The point is that the same principle scales to fluent professionals refining specific high-stakes language patterns. A B2+ executive preparing for a board pushback does not need someone to "correct" the answer she has already drafted. She needs to produce the target opening twice, with scaffolding, before she walks in. That is errorless teaching applied to the register she already operates at.

The contrast with paste-into-AI-correct is structural, not stylistic. AI feedback addresses output that has already happened. Errorless teaching addresses output that is about to happen. The first repairs; the second produces. For the patterns that show up under pressure — the ones that have been automatized for years — only the second matters.

What this means for you in practice

The behavioral shift is not "stop using AI" — that misreads the diagnosis. The shift is recognizing what AI feedback is useful for and what it is not.

AI is useful for output you are not trying to internalize.

• Drafting a non-strategic email.
• Summarizing a long document.
• Generating a first pass on a deck where the structure matters more than the wording.

In those uses, AI correction is a productivity tool, and the fact that you read the correction instead of producing it is a feature — you were not practicing English, you were producing a deliverable. The receptive nature of the interaction is fine because the goal was never acquisition.

AI is not useful for the patterns you need to own.

• The opening of a board pushback.
• The structure of a salary-negotiation counter.
• The register of a difficult-feedback conversation with a direct report.

These are exactly the moments where the output has to come from you, accurately and at speed, with no time to paste and refine. For those, the practice has to be production-shaped, not reception-shaped — which is what the SLA literature has been saying since 1985 and what errorless teaching has operationalized for adult learners with a methodological backbone that goes back further.

The practical replacement for the paste-into-AI habit is structured production rehearsal of target language patterns before high-stakes deployment. In CTL's consulting work, this looks specifically like this: you bring the situation — the meeting on the calendar, the document you need to write, the conversation you are preparing for — and we work the target language patterns through the four-phase errorless structure.

• Pre-teach the register and the moves.
• Demonstrate the target sentences in your actual context.
• Practice with diminishing scaffolding.
• Produce independently, in front of someone whose job is to make sure the production is correct on the first attempt rather than corrected after the fact.

The artifact is not a corrected document. The artifact is a pattern you now own, in your own voice, ready to deploy.

If you have been pasting drafts into ChatGPT to "check" them, the question is not whether the tool is good. The tool is good — at what it does. The question is whether what the tool does is what your career situation actually needs.

AI correction is not the wrong tool. It is the wrong practice for the patterns that matter most. Production scaffolded into correct form on the first attempt is what does the work AI feedback cannot do — and what the patterns you operate on under pressure are made of.

If your work involves high-stakes English moments — and if you have been pasting drafts into ChatGPT to "check" them — that habit is fossilizing the patterns you will need most. Errorless scaffolding applied to your actual upcoming conversations is what working with me one-on-one is for. Specific situations, specific language. Let's talk.

Frequently asked questions

What is fossilization in second language acquisition?

Fossilization is the persistent retention of incorrect language forms in a learner's interlanguage, observable across learners regardless of motivation or further instruction. Larry Selinker introduced the concept in 1972; ZhaoHong Han's 2004 synthesis treats fossilization as the stabilization of specific features that becomes robust against further input, instruction, or motivation once established. For adult L2 learners, fossilized features are harder to reverse than in younger learners.

Does AI grammar correction actually help with English fluency?

AI grammar correction tools improve short-term writing performance on surface-level features like grammar and sentence variety, as 2024–2025 empirical studies confirm. They do not, however, replicate the production practice that second-language acquisition research identifies as necessary for actual fluency gains. A 2025 systematic review noted "few longitudinal studies examining their long-term impact on writing development" — the field has not yet measured retention effects.

What is errorless teaching?

Errorless teaching is a pedagogy rooted in Applied Verbal Behavior (AVB) that structures learning so the learner produces the correct response on the first attempt, never encoding the error in the first place. Originally developed in B.F. Skinner's verbal behavior framework, refined in Discrete Trial Training (DTT), and operationalized in modern verbal-operant assessment, it applies to adult L2 learners through scaffolded production rehearsal.

Should I stop using ChatGPT or other AI tools for English?

No. AI feedback is genuinely useful for output you are not trying to internalize — drafting non-strategic emails, summarizing documents, generating first-pass content. The shift is recognizing where AI correction fits (productivity work) and where it does not (deliberate skill-building for high-stakes patterns). For patterns you need to deploy from memory under pressure, the practice must be production-shaped, not reception-shaped.

Who benefits most from errorless teaching for Business English?

B2+ professionals operating in English daily — senior managers, founders, technical experts in international roles — whose existing patterns are automatized at the register level. These are the patterns most at risk from AI-correction shortcuts and least amenable to retroactive intervention. Errorless teaching applies the same lineage developed for early intensive intervention to adult L2 work through structured pre-teaching, demonstration, scaffolded practice, and independent production.

Sources

• Selinker, L. (1972). "Interlanguage." IRAL — International Review of Applied Linguistics in Language Teaching, 10(3), 209–231.
• Han, Z.-H. (2004). Fossilization in Adult Second Language Acquisition. Multilingual Matters.
• Swain, M. (1995). "Three functions of output in second language learning." In G. Cook & B. Seidlhofer (Eds.), Principle and Practice in Applied Linguistics: Studies in Honour of H.G. Widdowson, pp. 125–144. Oxford University Press.
• Skinner, B.F. (1957). Verbal Behavior. Appleton-Century-Crofts.
• Sundberg, M.L. (2008). VB-MAPP: Verbal Behavior Milestones Assessment and Placement Program. AVB Press.
• Lovaas, O.I. (1987). "Behavioral treatment and normal educational and intellectual functioning in young autistic children." Journal of Consulting and Clinical Psychology, 55(1), 3–9.
• DeKeyser, R. (2000). "The robustness of critical period effects in second language acquisition." Studies in Second Language Acquisition, 22(4), 499–533.
• Zhang, Z., Aubrey, S., Huang, X., & Chiu, T.K.F. (2025). "The role of generative AI and hybrid feedback in improving L2 writing skills: a comparative study." Innovation in Language Learning and Teaching. DOI: 10.1080/17501229.2025.2503890.
• Wu, J., Li, J., Ge, Z., Xu, M., Lin, L., & Zhang, R. (2025). "Effectiveness of Generative AI in Automated Written Corrective Feedback With Prompting." Journal of Educational Computing Research, 63(6), 1493–1527. DOI: 10.1177/07356331251359430.
• Sharmithashini, M. & Hashim, H. (2025). "Sustaining ESL Writing Development with AI-Driven Automated Feedback Systems: A Systematic Review (2006–2025)." International Journal of Research and Innovation in Social Science, 9(8). Open access.