Learning Transfer in AI-Supported English Language Teaching
A research brief on learning transfer in AI-supported ELT: the distinction between AI-assisted performance and durable, transferable learning in English language teaching (ELT), and what the available evidence does and does not show.
Pedagogy First. AI Enhanced. Future Ready.
Research-informed; not externally peer-reviewed. This brief summarises published sources that were verified by title, DOI, and abstract in October 2026. It is not a systematic review and it presents no NIR data. Numerical details should be checked against the full texts.
How to read the labels
Reported in a verified source, within that source’s context and limits.
NIR’s reading of how findings may bear on ELT. Not a finding.
Not established by the sources reviewed here.
1. Definition and terminology
In this brief, learning transfer in AI-supported ELT means a learner’s ability to apply what was learned in a new situation, without the support that was present during learning. The question for AI-supported ELT is whether work done with AI support leaves the learner able to perform without it.
The word “transfer” has several meanings in applied linguistics and education. They are kept apart here.
| Term | Meaning in this brief | Not to be confused with |
|---|---|---|
| Performance | Observable behaviour or output during, or immediately after, acquisition. | Learning. Performance is often an unreliable guide to it (Soderstrom & Bjork, 2015). |
| Learning | A relatively long-term change in knowledge or skill, inferred from later performance. | Anything visible during the task itself. |
| AI-assisted performance | Output produced while AI support is available. | Evidence of what the learner can do without that support. |
| Unassisted performance | Output produced without AI support, later in time, in the same or a new task. | A teaching method. It is a condition under which learning is judged. |
| Transfer of learning | Applying what was learned in a new situation. It varies by content and by context (Barnett & Ceci, 2002). | Crosslinguistic influence. |
| Crosslinguistic influence (“language transfer”) | The effect of one language on another. | Transfer of learning, despite the shared word. |
| Transfer of training | Applying professional development to workplace practice. Used in the teacher-training section only. | Learner transfer. It is a bounded special case. |
2. Performance versus learning
Soderstrom and Bjork (2015) review evidence from motor- and verbal-learning research. They report that what can be observed during instruction is performance, and that performance is often an unreliable guide to whether durable learning has occurred. Learning can occur without visible performance gains, and improvements in performance can fail to yield learning. They also review research suggesting that people tend to read their own performance during acquisition as a guide to long-term learning.
In AI-supported ELT, a fluent AI-assisted text or answer is first of all performance. It becomes evidence of learning only if later, unassisted tasks show retained or transferred ability. Teacher and learner impressions of progress during an AI-assisted task are open to the misreading the review describes.
How large the gap between AI-assisted performance and durable L2 learning is across ELT skills, proficiency levels, and contexts has not been established by the sources reviewed here.
3. Transfer theory relevant to ELT and SLA
Barnett and Ceci (2002) argue that transfer is not a single phenomenon. Their taxonomy separates the content of what is transferred from the context in which it is applied, and describes context as varying along several dimensions, including knowledge domain, physical, temporal, functional, social, and modality. Whether transfer is “near” or “far” depends on how far the new situation differs along these dimensions.
The taxonomy gives ELT a way to state a transfer target precisely. The table gives examples. It is an illustration of how the dimensions could be read in ELT, and not a NIR taxonomy.
| Context dimension | Example in ELT | Status |
|---|---|---|
| Knowledge domain | Writing unassisted on a new topic after AI-supported practice on another | Interpretation |
| Temporal | Repeating an unassisted task weeks after the AI-supported work | Interpretation |
| Modality | Moving from AI-supported writing practice to speaking | Interpretation |
| Functional and social | Moving from practice with an AI partner to interaction with other people | Interpretation |
| Physical and functional | Moving from classroom practice to an exam or workplace task | Interpretation |
SLA-specific accounts of transfer, such as processing-based and skill-based accounts, are relevant. They are not cited here because their sources were not verified in this review. They are listed under research gaps.
4. Evidence from AI-supported L2 research
The table gives the verified sources that bear most directly on learning transfer in AI-supported ELT, that is, on assisted versus unassisted outcomes, plus the teacher-training review used in Section 5. Fan et al. and Bastani et al. are outside ELT. They are included because they test transfer or measure what happens when AI support is removed.
| Source | Context | Design | Reported finding | What it does not show |
|---|---|---|---|---|
| Li, Wang & Yang (2025) | GenAI chatbots and second language acquisition | Meta-analysis of 41 experimental or quasi-experimental studies (48 effect sizes, 3,515 participants) | Moderate to large positive pooled effect (ES = 0.576, 95% CI 0.385 to 0.768). Moderator analyses, including intervention duration | Whether gains persist or hold without AI. The abstract verified for this brief reports no separate delayed or unassisted estimate |
| Fan et al. (2025) | University students on a writing task; ChatGPT, human expert, analytics tool, or no tool | Randomised experiment, N = 117 | ChatGPT group improved essay scores most. Knowledge gain and transfer did not differ significantly. Self-regulated learning processes differed between groups | Not described as an L2 study. One task, short-term, controlled setting |
| Bastani et al. (2025) | High school mathematics in Turkey | Field experiment, nearly 1,000 students; unguarded GPT-4 tutor, safeguarded tutor, or no AI | Students with the unguarded tutor later scored lower without AI than students who never had it (reported 17% lower). Safeguards largely mitigated this | Not L2. One school, one subject, one term |
| Kennedy (2016) | Teacher professional development in US K-12 core academic content | Review of 28 studies against research-design standards | Many popular design features were not associated with effectiveness. Pedagogies for helping teachers enact ideas differed in effectiveness | Not ELT and not AI. See Section 5 |
A meta-analysis of generative AI chatbots in second language acquisition reports positive pooled effects (Li et al., 2025). Two studies outside ELT report that AI support improved task performance without matching gains in knowledge or transfer (Fan et al., 2025), or that unguarded AI access was followed by lower unassisted performance, with safeguards largely mitigating the effect (Bastani et al., 2025).
These results are compatible with the view that how AI is built into learning matters, including whether the learner must still do the cognitive work. They do not show that the same effects occur in ELT, and they do not show that ELT effects are absent. The L2 meta-analysis and the two non-ELT studies measure different things.
Delayed and unassisted outcomes in ELT have not been synthesised in the sources verified here. The review encountered several recent single-context EFL studies that include delayed posttests, typically two to four weeks later. They are not summarised because their full citation details could not be verified in this review.
5. Teacher training and transfer to classroom practice
This section is bounded. It concerns only whether teacher professional development leads to changes in classroom practice.
Kennedy (2016) reviewed 28 studies of teacher professional development in US K-12 core academic content. Many popular design features were not associated with program effectiveness, and programs differed in how effectively they helped teachers enact ideas in their own practice. The review does not concern ELT or AI.
The review suggests asking how a professional learning activity helps a teacher carry an idea into their own classroom, and not only whether the activity was completed. Practice environments, including simulations, are one candidate answer.
Whether practice in an AI teaching simulator transfers to classroom practice in ELT has not been established by the sources reviewed here. The NIR page for NIR SIM™ describes it as a simulator for rehearsal and feedback and presents no transfer data.
6. Research-informed design questions for learning transfer in AI-supported ELT
These are prompts for designers, teachers, and researchers. They are derived from the sources above. They are not criteria, they are not scored, and they do not constitute a framework.
- What is the intended transfer target: which task, topic, modality, and setting?
- Is AI present or absent when the learner’s ability is later judged?
- Is there a delay between AI-supported work and later judgement, and how long is it?
- What must the learner do before, during, and after the AI output: retrieve, explain, revise, produce?
- Does the design make the learner’s own thinking visible, for example through drafts and explanations, or does it hide it?
- What evidence would count against the intended effect, and who will look for it?
- For teachers, how does the activity help them carry what they practised into their own classes?
7. What NIR can and cannot claim
This brief does not claim that any NIR tool or course demonstrably causes learning transfer in AI-supported ELT. The boundary is stated so that the pages that link here stay within it.
NIR can
- Describe what its tools do, as stated on their pages.
- Report published research with its context, design, and limits.
- State what NIR has not measured.
- Ask the design questions above of its own tools and courses.
NIR cannot
- Claim that NIR tools cause learning transfer.
- Present tool-generated scores as validated measures of durable learning.
- Generalise from non-ELT studies to ELT.
- Cite figures that cannot be traced to a source.
- Describe this brief as peer-reviewed or externally validated.
8. Research gaps and future directions
The gaps below concern learning transfer in AI-supported ELT specifically.
- Delayed, unassisted outcomes in ELT. Studies that report both assisted and unassisted results, with the delay stated, and a synthesis of them.
- Transfer across modality and setting. Evidence on whether AI-supported writing or practice carries over to speaking and to interaction with people.
- Learner differences. Proficiency level, age, and first-language background as moderators of transfer.
- Design features. Which ways of building AI into tasks preserve learner cognitive work, tested in ELT rather than inferred from other subjects.
- SLA transfer theory and AI. A verified synthesis of SLA-specific accounts of transfer with the AI-supported evidence.
- teacher training in AI contexts. Evidence on whether AI-based practice, including simulation, changes classroom practice in ELT.
- Reporting. Studies stating whether AI was present when outcomes were measured.
Frequently asked questions
What is the difference between AI-assisted performance and learning?
Does AI use harm language learning?
What does learning transfer in AI-supported ELT mean?
How could an institution look for evidence of transfer?
Does NIR claim that its tools improve learning transfer?
Is this brief peer-reviewed?
9. References
Each source was checked against its publisher, indexing, or abstract page in October 2026. Findings above are summarised from abstracts and metadata unless stated, so numerical details should be checked against the full texts.
- Barnett, S. M., & Ceci, S. J. (2002). When and where do we apply what we learn? A taxonomy for far transfer. Psychological Bulletin, 128(4), 612–637. https://doi.org/10.1037/0033-2909.128.4.612 (Peer-reviewed journal article.)
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122 (Peer-reviewed journal article. An earlier working-paper version exists.)
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544 (Peer-reviewed journal article, first published online December 2024. A preprint version also exists.)
- Kennedy, M. M. (2016). How does professional development improve teaching? Review of Educational Research, 86(4), 945–980. https://doi.org/10.3102/0034654315626800 (Peer-reviewed journal article.)
- Li, M., Wang, Y., & Yang, X. (2025). Can generative AI chatbots promote second language acquisition? A meta-analysis. Journal of Computer Assisted Learning, 41(4), e70060. https://doi.org/10.1111/jcal.70060 (Peer-reviewed journal article.)
- Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000 (Peer-reviewed journal article.)
Research-informed; not externally peer-reviewed. This research brief is published by NIR Academy. It presents no NIR data, introduces no new framework or score, and does not claim that NIR tools cause learning transfer in AI-supported ELT. It does not provide legal or professional advice. Last reviewed: October 2026.

