Research & Learning Insights · Evidence review

AI in Second-Language Learning: What Can It Actually Help With?

A systematic evidence synthesis separating the functions with credible learning support from the functions that remain overclaimed, under-tested or risky.

63 scholarly sources reviewedMeta-analyses · systematic reviews · controlled studiesUpdated August 2026

01 · Executive summary

AI is not one teaching method

The phrase “AI for language learning” combines technologies that do very different things. Automatic speech recognition listens. Automated writing evaluation detects and classifies text features. Machine translation transforms one language into another. Chatbots generate dialogue. Large language models generate explanations, examples, feedback and text. Their evidence therefore has to be judged function by function.

The broadest recent GenAI meta-analysis synthesized 51 studies and 175 independent effect sizes and reported large pooled effects for language achievement and affective-cognitive outcomes.1 That is encouraging, but not a license to conclude that any AI feature causes large gains. The same literature reports large heterogeneity, highly variable controls, a fast-moving technology base and a research corpus concentrated in higher education and English.11314

0.69

Meta-analytic effect of ASR on ESL/EFL pronunciation.7

0.61

Overall chatbot effect on second/foreign-language learning.3

0.39

AI-enabled assessment effect on language learning.11

The useful framing

Ask “What learning action is the AI increasing or improving?” rather than “Does AI work?” The best-supported answers are: more feedback, more speaking turns, more pronunciation attempts, more revision cycles, targeted vocabulary practice and faster formative information.

02 · Review method

Evidence was weighted by what the study actually measured

This review prioritizes meta-analyses and systematic reviews, then controlled and quasi-experimental L2 studies, and finally established SLA theory needed to explain mechanisms. Searches covered generative AI, ChatGPT/LLMs, conversational agents, ASR, automated writing evaluation, AI-enabled assessment and machine translation. General educational AI studies were not treated as language-learning evidence unless they directly informed a mechanism.

Interpretive safeguards

  • Performance is not automatically learning. A better essay produced with AI may reflect assistance during the task rather than a durable increase in unaided writing ability.
  • Tools change faster than trials. Findings from GPT-3.5, older ASR engines or rule-based AWE cannot be transferred perfectly to current multimodal models.
  • English dominates. Evidence for French, Mandarin, Spanish and less commonly taught languages is far thinner.
  • Very large effects deserve scrutiny. Early-stage literatures often contain small samples, weak controls and strong novelty effects.59

03 · Capability map

What can AI actually help with?

Strong evidence

Writing feedback & revision

AWE meta-analyses consistently show positive writing effects, particularly for accuracy and surface-level revision. Human or teacher-mediated feedback remains important for rhetoric, argument and prioritization.8947

Strong evidence

Pronunciation & speech practice

ASR improves pronunciation overall, especially segmental targets when feedback is explicit and practice is sustained.719

Moderate evidence

Conversational speaking

Chatbots increase low-stakes interaction; controlled studies show gains in oral proficiency and reductions in anxiety, but transfer to human conversation is less established.322

Promising / heterogeneous

Vocabulary

AI vocabulary interventions show large pooled gains, but heterogeneity is extreme and enriched non-AI multimedia controls reduce the apparent advantage.6

Moderate evidence

Formative assessment

AI assessment can produce useful, rapid diagnostic information and improve learning when it is iterative and instructionally connected.11

Under-researched

Reading & listening

These skills appear frequently in AI claims but much less often in rigorous comparative research; systematic reviews explicitly identify them as gaps.14

Figure 1 · Author synthesis

The most defensible AI learning loop

Learner attemptsSpeak, write, retrieve, choose or interpret before seeing the answer.
AI respondsFeedback, transcription, hint, comparison, follow-up question or targeted example.
Learner revisesExplain the error, repair output, try again and make a new decision.
Delayed transferUse the form again without the AI in a new task, person or context.
Author synthesis based on corrective-feedback, interaction, output, noticing and sociocultural accounts.5354555657

04 · Writing & feedback

The clearest case is not “AI writes better”—it is “AI can create more revision cycles”

Writing is the most heavily researched area of GenAI language learning, but it is also where performance and learning are easiest to confuse.

Pre-GenAI automated writing evaluation already had a substantial evidence base. Zhai and Ma found a large overall effect on writing quality, and later syntheses similarly report benefits for EFL/ESL writing.89 New LLMs broaden feedback from grammar and mechanics toward explanations and rhetorical comments, but higher-order feedback is still less dependable than surface correction.47

Teacher-in-the-loop designs are particularly defensible. Han and Li used ChatGPT to help teachers generate detailed corrective and rhetorical feedback for large EFL classes rather than delivering unreviewed AI comments directly to students.29 This uses AI to increase feedback capacity while retaining human judgment.

Important distinction

Assisted text quality ≠ unaided writing development. Woo et al. found that AI-generated text could contribute to higher-quality student products, but learner competence strongly shaped how well students used it.30 The educational question is whether students can later plan, formulate and revise without the model.

01

Error detection

AI can flag grammar, spelling, lexical and some discourse problems immediately, increasing the number of feedback events per learner.8

02

Explanation on demand

LLMs can rephrase a correction, generate contrasts and answer follow-up questions; quality depends on the model and prompt, as shown in L2 Chinese feedback research.31

03

Revision support

The useful outcome is a better learner revision, not merely a better model rewrite. AWE research repeatedly finds that uptake and revision behaviour mediate value.3537

05 · Speaking & pronunciation

AI can manufacture practice opportunities that many learners otherwise do not get

For pronunciation, the evidence is mature enough to be specific. ASR has a medium overall effect, with stronger results when systems provide explicit corrective information, target segmental pronunciation and are used for more than a very short exposure.7 Dynamic-assessment designs can further reduce anxiety and cognitive load by scaffolding feedback instead of simply marking an attempt wrong.23

Conversational GenAI adds another affordance: a learner can create dozens of speaking turns without waiting for a teacher or partner. In one randomized four-week trial, LLM dialogue improved oral proficiency, willingness to communicate and self-efficacy while lowering foreign-language anxiety.22 Meta-analysis of informal GenAI learning likewise found its strongest skill-specific effect for speaking.4

What it does not prove

A conversation with a patient, always-available bot is not psychologically identical to a conversation with a human interlocutor. Evidence for far transfer to spontaneous, socially consequential human interaction remains much thinner than evidence for practice performance and anxiety reduction.

06 · Vocabulary & language knowledge

Personalization can help vocabulary—but the huge effects should be read with caution

The 2026 AI-vocabulary meta-analysis reported a very large pooled effect, g = 1.309, across 18 studies. Yet the same analysis reported I² = 93.868%, meaning results varied enormously across studies.6 The relative advantage also shrank when AI was compared with multimedia-supported learning rather than traditional instruction.

Published research figure · CC BY

AI-supported vocabulary interventions vary dramatically across studies

Forest plot reproduced from Şimşek et al. (2026), Frontiers in Artificial Intelligence, CC BY. The pooled effect is large, but the wide dispersion is the important design message: AI is not a uniform treatment.6

Pedagogically, the strongest vocabulary use is likely to combine AI’s generative flexibility with established learning mechanisms: tailored examples, retrieval questions, contrastive examples, spaced re-encounters and required learner production. Open-ended chat without retrieval or delayed testing may feel rich while producing less durable memory.

07 · Assessment & personalization

Fast formative information is useful; high-stakes judgment is a different problem

AI-enabled language assessment has a moderate positive learning effect in meta-analysis, but the educational gains come primarily from using assessment formatively—to trigger another learning action.11 A score is useful when it tells the learner what to try next; a score that is opaque or inaccurate can simply automate confusion.

“Personalization” should also be defined precisely. AI can change difficulty, examples, pacing, feedback language and practice topics quickly. What remains much less established is whether a generic LLM accurately models an individual learner’s evolving interlanguage over months. Persistent learner models, validated diagnostics and controlled progression are stronger forms of personalization than remembering a few conversational preferences.

08 · Reading & listening

Useful support tools, but surprisingly little direct causal evidence

AI can simplify passages, define words, generate comprehension questions, produce transcripts, create listening scripts and answer questions about a text. These are plausible supports. Yet the main systematic reviews of post-ChatGPT language education repeatedly report that reading and listening are under-studied compared with writing.1314

That gap matters. A chatbot that explains a reading passage may improve task completion while reducing the amount of inferencing the learner performs. A transcript may help listening comprehension while decreasing attention to the acoustic signal. Good design therefore uses assistance contingently—after an attempt, at the point of breakdown, and with later removal of support.

09 · Why AI can help

The useful mechanisms are older than ChatGPT

01

More interaction

Interactionist theory predicts benefits when learners negotiate meaning and modify output. AI can dramatically increase access to such exchanges if tasks require genuine learner responses.54

02

More output

Producing language forces learners to test what they can say. AI helps when it elicits output; it can hurt when it supplies the output before the learner tries.55

03

More noticing

Highlighting contrasts, errors and alternatives can focus attention on form. But automatic rewriting can hide the exact gap the learner needed to notice.56

04

More individualized feedback

Feedback is powerful when it is timely, specific and actionable. AI lowers the marginal cost of delivering another feedback cycle, although correctness still has to be monitored.51

Published research figure · CC BY 4.0

The early GenAI evidence is positive—and highly dispersed

Language-achievement forest plot reproduced from Saarela, Gunasekara & Kumarage (2026), Discover Computing, CC BY 4.0. The plot visually demonstrates why pooled headline effects should not be treated as a single expected result for every learner or product.1

10 · Limits, failure modes & safeguards

The same flexibility that makes GenAI useful also creates its main educational risks

Six recurring problems in the literature

  • Incorrect or invented language information. Fluent explanations can be wrong, incomplete or overconfident; systematic reviews identify content quality as a core challenge.1314
  • Over-reliance. If the learner asks AI to produce the answer, they may complete the task with less retrieval, formulation and monitoring of their own language.
  • Model drift. Feedback quality can change with model version; L2 Chinese research found meaningful version effects.31
  • Uneven benefit. More proficient learners may be better positioned to judge, prompt and integrate AI output, potentially widening rather than closing gaps.30
  • Privacy and governance. Language practice can contain personal, educational or assessment data; review literature repeatedly flags privacy leakage and unclear data handling.14
  • Construct substitution. AI can make a product look better without improving the underlying skill—for example, rewriting an essay rather than teaching the learner to write one.
A practical test

After using the AI, remove it. Can the learner now perform better on a new item, in a delayed session, with a different prompt or with a human partner? If not, the system may have optimized task completion rather than learning.

11 · Design implications

Use AI to increase productive struggle, not erase it

Evidence-aligned design rules

  1. Attempt before assistance. Require the learner to answer, speak, translate, predict or draft before the model reveals a solution.
  2. Prefer feedback over replacement. “Here is what to fix and why” is usually more educational than “Here is a better paragraph.”
  3. Make feedback actionable. Ask for a revised attempt immediately and revisit the same issue later.
  4. Use constrained roles. A pronunciation coach, examiner, Socratic tutor or error explainer is easier to evaluate than a generic “AI teacher.”
  5. Keep a human escalation path. Teachers should handle ambiguity, higher-order discourse, persistent misconceptions and high-stakes decisions.
  6. Measure delayed unaided transfer. Platform scores during AI assistance are insufficient evidence of learning.
  7. Log model/version and prompts. Replicability requires knowing which system produced the feedback.
  8. Do not generalize English findings automatically. French and other target languages require direct validation of feedback, speech recognition and cultural/pragmatic quality.

12 · Evidence matrix

All 63 sources, with the conclusion used in this review

The table makes the review auditable. Each “main conclusion” is a concise paraphrase of the finding or synthesis relevant to AI-supported second-language learning; it is not a quotation from the paper.

63-source evidence matrix
#StudyPublicationMain conclusion used hereLink
1Saarela, Gunasekara & Kumarage (2026)GenAI meta-analysisA meta-analysis of generative AI effects on language proficiency and affective–cognitive outcomes in language learning. Discover Computing, 29, 116.Across 51 studies and 175 effect sizes, GenAI interventions produced large pooled effects for language achievement (g = 1.111) and affective–cognitive outcomes (g = 1.029), but the authors stress substantial heterogeneity and moderator dependence.Source ↗
2Wang (2024)AI meta-analysisArtificial Intelligence in L2 learning: A meta-analysis of contextual, instructional, and social-emotional moderators. System, 126, 103498.A broad synthesis of AI-supported L2 learning reported an overall moderate-to-large benefit (SMD about .98) and found that environment, duration, educational level, interactivity and feedback source help explain wide variation in outcomes.Source ↗
3Lyu, Lai & Guo (2025)Chatbot meta-analysisEffectiveness of Chatbots in Improving Language Learning: A Meta-Analysis of Comparative Studies. International Journal of Applied Linguistics, 35, 834–851.Thirty-one studies (41 effects; 2,943 learners) yielded a medium chatbot advantage (g = .608); mobility, modality, GenAI involvement and the comparison condition moderated effectiveness.Source ↗
4Guan, Li & Gu (2024)GenAI meta-analysisAI in informal digital English learning: A meta-analysis of its effectiveness on proficiency, motivation, and self-regulation. Computers & Education: Artificial Intelligence, 7, 100323.In 15 informal-learning studies, GenAI improved overall English proficiency (g = .447) and self-regulation (g = .508); speaking showed the clearest skill-specific benefit (g = .794), while motivation was not significantly improved.Source ↗
5Li, Wang & Yang (2025)GenAI chatbot meta-analysisCan Generative AI Chatbots Promote Second Language Acquisition? A Meta-Analysis. Journal of Computer Assisted Learning, 41, e70060.Forty-one experimental/quasi-experimental studies with 3,515 participants found a moderate-to-large overall GenAI-chatbot effect (ES = .576), with outcomes varying across target language, tool, objective and intervention length.Source ↗
6Şimşek et al. (2026)Vocabulary meta-analysisArtificial intelligence-assisted vocabulary teaching: a meta-analysis study. Frontiers in Artificial Intelligence, 9, 1806700.Eighteen studies (23 effects; 1,823 participants) produced a large vocabulary effect (g = 1.309), but heterogeneity was extreme (I² = 93.868%), making context and comparator crucial to interpretation.Source ↗
7Ngo, Chen & Lai (2024)ASR meta-analysisThe effectiveness of automatic speech recognition in ESL/EFL pronunciation: A meta-analysis. ReCALL, 36(1), 4–21.Across 15 studies and 38 effects, ASR produced a medium pronunciation benefit (g = .69). Explicit corrective feedback, segmental targets, longer treatments and peer-supported practice were associated with stronger outcomes.Source ↗
8Zhai & Ma (2023)AWE meta-analysisThe Effectiveness of Automated Writing Evaluation on Writing Quality: A Meta-Analysis. Journal of Educational Computing Research, 61(4), 875–900.Twenty-six studies with 2,468 participants showed a large AWE effect on writing quality (g = .861), with stronger benefits for post-secondary and EFL/ESL learners.Source ↗
9Ngo, Chen & Lai (2024)AWE meta-analysisThe effectiveness of automated writing evaluation in EFL/ESL writing: a three-level meta-analysis. Interactive Learning Environments, 32(2), 727–744.A three-level synthesis found automated writing evaluation beneficial for EFL/ESL writing, while emphasizing that effects depend on learner, feedback and implementation characteristics rather than the software alone.Source ↗
10Huang, Xu, Li & Yu (2024)AWE meta-analysisA Meta-analysis of Effects of Automated Writing Evaluation on Anxiety, Motivation, and Second Language Writing Skills. The Asia-Pacific Education Researcher.The synthesis concludes that AWE can improve L2 writing while also influencing affective outcomes; it recommends combining automated and human rating/feedback rather than treating AWE as a complete substitute.Source ↗
11Chen et al. (2025)Assessment meta-analysisA systematic review and meta-analysis of AI-enabled assessment in language learning: Design, implementation, and effectiveness. Journal of Computer Assisted Learning, 41, e13064.Twenty-five studies yielded a medium learning effect (g = .390). Formative/iterative assessment design, duration, educational level and learner context moderated the value of AI-enabled assessment.Source ↗
12Lee (2023)Machine translation reviewThe effectiveness of machine translation in foreign language education: a systematic review and meta-analysis. Computer Assisted Language Learning, 36(1–2), 103–125.Machine translation can support foreign-language writing and vocabulary when learners critically compare, edit and reflect on output; uncritical substitution risks bypassing language processing.Source ↗
13Lee, Choe, Zou & Jeon (2025)Systematic reviewGenerative AI (GenAI) in the language classroom: A systematic review. Interactive Learning Environments, 34, 335–359.Reviewing 49 classroom studies, the authors found writing and learner perceptions heavily overrepresented; GenAI served as feedback provider, tutor, cognitive stimulator and conversation partner, but evidence was dominated by higher education and self-report.Source ↗
14Lo et al. (2024)Systematic reviewExploring the application of ChatGPT in ESL/EFL education and related research issues: a systematic review of empirical studies. Smart Learning Environments, 11, 50.Across 70 empirical studies, ChatGPT expanded practice and personalization but research was heavily concentrated on writing; objective evidence for reading, listening and speaking remained sparse, alongside concerns about inaccuracy, privacy and academic integrity.Source ↗
15Law (2024)Scoping reviewApplication of generative artificial intelligence (GenAI) in language teaching and learning: A scoping literature review. Computers and Education Open, 6, 100174.The scoping review mapped 41 publications and found considerable promise alongside a shortage of rigorous empirical work, explicit ethical treatment and skill-specific intervention research.Source ↗
16Bao et al. (2025)Systematic reviewA systematic review of AI in second language acquisition using the expanded SAMR model (2015–2024). Discover Computing, 28, 292.The review argues that AI use in SLA ranges from simple substitution to task transformation; pedagogical value increases when AI changes the quality of interaction or feedback rather than merely automating an old worksheet.Source ↗
17Li, Shadiev & Chiu (2026)Systematic review / meta-analysisApplying Generative Artificial Intelligence to Task-based Language Teaching and Learning: A Systematic Review and Meta-analysis. TechTrends, 70, 150–163.GenAI can support task preparation, interaction and feedback in task-based language learning, but effectiveness depends on task design, learner agency and how much cognitive work remains with the learner.Source ↗
18Huang, Hew & Fryer (2022)Systematic reviewChatbots for language learning—Are they really useful? A systematic review of chatbot-supported language learning. Journal of Computer Assisted Learning, 38(1), 237–257.Twenty-five empirical studies showed technological, pedagogical and social affordances for language practice, but also highlighted shallow dialogue, design constraints and the need for theory-informed chatbot implementation.Source ↗
19Liu, Ab Rahman & Mohamad Zain (2025)Systematic reviewA systematic literature review of research on automatic speech recognition in EFL pronunciation. Cogent Education, 12, 2466288.Twenty-four empirical studies largely supported ASR for pronunciation accuracy, especially segmentals, while revealing weak theoretical grounding, limited teacher perspectives and benefits from combining ASR with peer or other feedback.Source ↗
20AI-driven chatbot review (2025)Systematic reviewAI-driven chatbots in second language education: A systematic review of their efficacy and pedagogical implications. Ampersand, 14, 100224.Thirty empirical studies suggested the clearest chatbot benefits in productive skills, particularly speaking and writing, through practice volume, rapid feedback and lower-stakes interaction.Source ↗
21Wang, Zou, Du & Wang (2024)Speaking experimentThe impact of different conversational generative AI chatbots on EFL learners: willingness to communicate, speaking anxiety, self-perceived communicative competence, and performance. System, 127, 103533.Conversational GenAI conditions improved speaking-related outcomes relative to a conventional class comparison, illustrating that voice-capable chatbots can create additional low-stakes speaking turns rather than only text practice.Source ↗
22Zheng, Zhou, Chen & Ye (2025)Speaking RCTThe influence of large language models as collaborative dialogue partners on EFL English oral proficiency and foreign language anxiety. Computer Assisted Language Learning.In a four-week randomized study with 83 learners, GPT-4 dialogue practice improved oral proficiency, willingness to communicate and self-efficacy while reducing foreign-language anxiety; technical delays and topic drift remained practical limits.Source ↗
23Chen, Koong & Liao (2022)ASR experimentInfluences of Integrating Dynamic Assessment into a Speech Recognition Learning Design to Support Students’ English Speaking Skills, Learning Anxiety and Cognitive Load. Educational Technology & Society, 25(1), 1–14.Both speech-recognition conditions improved speaking and lowered anxiety, while dynamic-assessment scaffolding reduced performance anxiety and extraneous cognitive load more than simple corrective-feedback ASR.Source ↗
24Liakin, Cardoso & Liakina (2017)French pronunciationLearning L2 pronunciation with a mobile speech recognizer: French /y/. CALICO Journal.Mobile ASR-based practice helped learners notice and practice a difficult French vowel; the study illustrates how speech technology can target specific segmental contrasts rather than provide a generic ‘accent score.’Source ↗
25McCrocklin (2016)ASR / autonomyPronunciation learner autonomy: The potential of automatic speech recognition. System, 57, 25–42.ASR can increase opportunities for autonomous pronunciation practice and self-monitoring, but learners need training to interpret recognition errors and convert them into useful pronunciation decisions.Source ↗
26Chen (2024)Vocabulary / semanticsThe AI chatbot interaction for semantic learning: A collaborative note-taking approach with EFL students. Language Learning & Technology, 28(1), 1–25.Combining AI-chatbot interaction with collaborative note-taking supported semantic learning and learner-related outcomes, suggesting that the pedagogical wrapper around the chatbot matters as much as conversational access itself.Source ↗
27Chen (2025)Vocabulary / semanticsImproving English semantic learning outcomes through AI chatbot-based ARCS approach. Interactive Learning Environments, 33(6), 3909–3924.An ARCS-motivational design around an AI chatbot improved semantic learning by structuring attention, relevance, confidence and satisfaction rather than relying on open-ended chatting alone.Source ↗
28Fryer & Carpenter (2006)Foundational chatbotBots as language learning tools. Language Learning & Technology, 10(3), 8–14.Early chatbot work showed why bots are attractive for low-pressure language interaction, while also exposing the conversational shallowness and breakdowns of pre-LLM systems.Source ↗
29Han & Li (2024)Writing / feedbackExploring ChatGPT-supported teacher feedback in the EFL context. System, 126, 103502.A teacher-in-the-loop design used ChatGPT to help instructors produce corrective and rhetorical feedback at scale, illustrating a stronger use case than giving unfiltered model feedback directly to learners.Source ↗
30Woo et al. (2024)Writing / GenAIExploring AI-generated text in student writing: How does AI help? Language Learning & Technology, 28(2), 183–209.Among 23 secondary EFL writers, both human- and AI-generated words contributed to writing scores, but learner competence shaped how well AI text was integrated; AI did not erase proficiency differences.Source ↗
31Yang & Chen (2026)Writing / corrective feedbackChatGPT and L2 Chinese writing: evaluating the impact of model version and prompt language on automated corrective feedback. Computer Assisted Language Learning, 39(4), 1119–1147.ChatGPT-4 generated more accurate, detailed and understandable corrective feedback than 3.5 on L2 Chinese errors; model version mattered more than whether prompts were written in Chinese or English.Source ↗
32Shi, Chai, Zhou & Aubrey (2025)Writing comparisonComparing the effects of ChatGPT and automated writing evaluation on students’ writing and ideal L2 writing self. Computer Assisted Language Learning.Comparative work suggests generative feedback and conventional AWE do not function identically: the nature of feedback, learner uptake and motivational effects should be assessed separately rather than treating all AI writing tools as one category.Source ↗
33Barrot (2021)AWE experimentUsing automated written corrective feedback in the writing classrooms: Effects on L2 writing accuracy. Computer Assisted Language Learning.Automated corrective feedback can improve L2 writing accuracy when learners revise from it, but benefits depend on feedback accuracy, learner interpretation and repeated revision opportunities.Source ↗
34Dikli & Bleyle (2014)AWE / human comparisonAutomated essay scoring feedback for second language writers: How does it compare to instructor feedback? Assessing Writing, 22, 1–17.Automated and instructor feedback differ in focus and usefulness; machine feedback can handle recurring surface issues efficiently but does not fully reproduce human rhetorical judgment.Source ↗
35Stevenson & Phakiti (2014)AWE experimentThe effects of computer-generated feedback on the quality of writing. Assessing Writing, 19, 51–65.Computer-generated feedback supported revision, but writing improvement depended on whether learners engaged with and understood the feedback rather than on feedback availability alone.Source ↗
36Wang, Shang & Briody (2013)AWE experimentExploring the impact of using automated writing evaluation in English as a foreign language university students’ writing. Computer Assisted Language Learning, 26(3), 234–257.AWE increased revision opportunities and could support EFL university writing, while learners still valued human guidance for issues that automated systems handled poorly.Source ↗
37Li, Link & Hegelheimer (2015)AWE pedagogyRethinking the role of automated writing evaluation (AWE) feedback in ESL writing instruction. Journal of Second Language Writing, 27, 1–18.AWE is most useful as part of a feedback ecology: students selectively use machine feedback and benefit when teachers help them interpret, prioritize and revise from it.Source ↗
38Link, Mehrzad & Rahimi (2020)AWE pedagogyImpact of automated writing evaluation on teacher feedback, student revision, and writing improvement. Computer Assisted Language Learning.AWE can redistribute feedback work and affect revision patterns; its instructional value emerges from interaction with teacher feedback rather than simple automation of correction.Source ↗
39Jiang, Yu & Wang (2020)Teacher / AWESecond Language writing instructors’ feedback practice in response to automated writing evaluation: A sociocultural perspective. System, 93, 102302.Teachers adapt their feedback practices around AWE, often compensating for what automated systems miss; effective deployment therefore changes the teacher role instead of removing it.Source ↗
40Chang, Li, Huang & Whitfield (2021)Writing / acceptanceExploring EFL students’ writing performance and their acceptance of AI-based automated writing feedback. Proceedings of the International Conference on Education Development and Studies.Learner acceptance and perceived usefulness are important implementation variables: technically available feedback has little value if students do not trust, understand or act on it.Source ↗
41Hassanzadeh & Fotoohnejad (2021)AWE experimentImplementing an automated feedback program for a foreign language writing course: A learner-centric study. Journal of Computer Assisted Learning, 37(5), 1494–1507.Learner-centred AWE use can support writing development, but students vary in the strategies they use to interpret and incorporate automated feedback.Source ↗
42Sun & Fan (2022)AWE / anxietyThe effects of an AWE-aided assessment approach on business English writing performance and writing anxiety. Studies in Educational Evaluation, 72, 101123.AWE-assisted assessment improved business-English writing and influenced writing anxiety, showing that feedback technology can affect both performance and emotion.Source ↗
43Tan, Cho & Xu (2022)AWE + peer feedbackExploring the effects of automated written corrective feedback, computer-mediated peer feedback and their combination on EFL writing performance. Interactive Learning Environments.Automated and peer feedback offer different information; combined modes can broaden revision support compared with relying on either source alone.Source ↗
44Wilson & Roscoe (2020)AWE evaluationAutomated writing evaluation and feedback: Multiple metrics of efficacy. Journal of Educational Computing Research, 58(1), 87–125.AWE effectiveness varies depending on which outcome is measured—quality, revision, motivation or engagement—so a single score is an inadequate evaluation of instructional value.Source ↗
45Xu & Zhang (2021)AWE / proficiencyUnderstanding AWE feedback and English writing of learners with different proficiency levels in an EFL classroom: A sociocultural perspective. The Asia-Pacific Education Researcher.Learner proficiency shapes how automated feedback is interpreted and used, supporting differentiated scaffolding rather than one feedback stream for all students.Source ↗
46Cheng (2017)AWE experimentThe impact of online automated feedback on students’ reflective journal writing in an EFL course. The Internet and Higher Education, 34, 18–27.Online automated feedback can stimulate revision and writing improvement, but its effectiveness depends on sustained learner engagement with iterative feedback.Source ↗
47Fu, Zou, Xie & Cheng (2022)AWE reviewA review of AWE feedback: Types, learning outcomes, and implications. Computer Assisted Language Learning.The review concludes that automated feedback is strongest for timely, repeated support and surface-level issues; feedback design and pedagogical integration determine whether it develops broader writing competence.Source ↗
48Nunes, Cordeiro, Limpo & Castro (2022)AWE systematic reviewEffectiveness of automated writing evaluation systems in school settings: A systematic review of studies from 2000 to 2020. Journal of Computer Assisted Learning, 38(2), 599–620.School-based AWE can improve writing under some conditions, but effects are heterogeneous and stronger designs are needed to determine which features and instructional supports cause gains.Source ↗
49Lv, Ren & Xie (2021)Feedback meta-analysisThe effects of online feedback on ESL/EFL writing: A meta-analysis. The Asia-Pacific Education Researcher, 30(6).Online feedback has a positive overall effect on L2 writing, reinforcing the broader principle that immediacy and revisability matter, regardless of whether the feedback agent is fully automated.Source ↗
50Mohsen (2022)Feedback meta-analysisComputer-mediated corrective feedback to improve L2 writing skills: A meta-analysis. Journal of Educational Computing Research, 60(5).Computer-mediated corrective feedback improves L2 writing, but moderator patterns show that technology is an interaction channel, not a guarantee of useful feedback.Source ↗
51Hattie & Timperley (2007)Feedback theoryThe power of feedback. Review of Educational Research, 77(1), 81–112.Feedback is most effective when it reduces the gap between current and desired performance and addresses task/process/self-regulation; vague praise or poorly targeted feedback is much weaker.Source ↗
52Kang & Han (2015)Corrective feedback meta-analysisThe efficacy of written corrective feedback in improving L2 written accuracy: A meta-analysis. Modern Language Journal, 99(1), 1–18.Written corrective feedback improves L2 written accuracy overall, providing a mechanism that AI can scale—but only if the generated feedback is accurate and learners process it.Source ↗
53Li (2010)Corrective feedback meta-analysisThe effectiveness of corrective feedback in SLA: A meta-analysis. Language Learning, 60(2), 309–365.Corrective feedback has durable positive effects in SLA, supporting AI feedback systems in principle while leaving feedback timing, explicitness and learner response as critical design variables.Source ↗
54Long (1996)Interaction theoryThe role of the linguistic environment in second language acquisition. In Handbook of Second Language Acquisition.Interaction promotes acquisition when learners receive comprehensible input, negotiate meaning and modify output; conversational AI can provide these opportunities but cannot guarantee they actually occur.Source ↗
55Swain (1985)Output hypothesisCommunicative competence: Some roles of comprehensible input and comprehensible output in its development. In Input in Second Language Acquisition.Producing language can push learners to notice gaps and test hypotheses. AI is most educationally useful when it elicits learner output rather than simply generating fluent text for them.Source ↗
56Schmidt (1995)Noticing theoryConsciousness and foreign language learning: A tutorial on the role of attention and awareness in learning. Attention and Awareness in Foreign Language Learning.Attention and noticing are central to converting input into learning; AI explanations and error highlighting help when they direct attention, but answer-generation can do the opposite by removing the need to notice.Source ↗
57Lantolf, Thorne & Poehner (2020)Sociocultural theorySociocultural theory and second language development. In Theories in Second Language Acquisition (3rd ed.).Learning is mediated through tools and scaffolded interaction. AI can serve as a mediational tool, but the quality of scaffolding and gradual transfer of responsibility to the learner are the key educational questions.Source ↗
58Boulton & Cobb (2017)Technology meta-analysisCorpus use in language learning: A meta-analysis. Language Learning, 67(2), 348–393.Corpus-based data-driven learning has a substantial positive effect, reminding AI designers that access to authentic, inspectable language evidence can be more pedagogically useful than opaque fluent answers.Source ↗
59Plonsky & Oswald (2014)MethodsHow big is ‘big’? Interpreting effect sizes in L2 research. Language Learning, 64(4), 878–912.L2 effect sizes should be interpreted using field-sensitive benchmarks; very large early AI effects deserve caution when samples are small, heterogeneous or compared with weak controls.Source ↗
60Pack & Maloney (2023)GenAI perspectiveUsing generative artificial intelligence for language education research: Insights from using OpenAI’s ChatGPT. TESOL Quarterly, 57(4), 1571–1582.GenAI can assist language-education work, but reliability, transparency, prompt dependence and verification remain fundamental methodological concerns.Source ↗
61Chapelle (2025)GenAI perspectiveGenerative AI as game changer: Implications for language education. System, 132, 103672.GenAI changes access to interaction, language generation and feedback, but language-education value still depends on established SLA mechanisms and evidence rather than novelty.Source ↗
62Stockwell (2024)GenAI perspectiveChatGPT in language teaching and learning: Exploring the road we’re travelling. Technology in Language Teaching & Learning, 6(1).The paper urges language educators to treat ChatGPT as an evolving tool requiring AI literacy, explicit task design and critical evaluation rather than as an autonomous teacher.Source ↗
63Alharbi (2023)AI writing reviewAI in the foreign language classroom: A pedagogical overview of automated writing assistance tools. Education Research International, 2023, 4253331.Automated writing assistants can provide scalable language feedback and revision support, while limitations in accuracy, depth and learner dependence justify human oversight.Source ↗

Bottom line

AI is most credible as a multiplier of practice, feedback and interaction

Second-language learning still depends on the learner doing language-processing work. AI can make that work more available, individualized, immediate and repeatable. It can give a learner another pronunciation attempt, another conversation turn, another explanation, another revision cycle and another tailored example. Those are meaningful advantages. The evidence is much weaker for the idea that a fluent general-purpose model can replace curriculum, teachers, assessment standards, deliberate practice or real human communication. The design goal should therefore be AI that makes the learner do more useful language work—not AI that does the language work for the learner.

References

Sources

Links point to DOI records, publisher pages or stable article records where available. Published figures reproduced above are limited to sources whose stated licences permit reuse with attribution.

  1. Saarela, Gunasekara & Kumarage (2026). A meta-analysis of generative AI effects on language proficiency and affective–cognitive outcomes in language learning. Discover Computing, 29, 116. Source ↗
  2. Wang (2024). Artificial Intelligence in L2 learning: A meta-analysis of contextual, instructional, and social-emotional moderators. System, 126, 103498. Source ↗
  3. Lyu, Lai & Guo (2025). Effectiveness of Chatbots in Improving Language Learning: A Meta-Analysis of Comparative Studies. International Journal of Applied Linguistics, 35, 834–851. Source ↗
  4. Guan, Li & Gu (2024). AI in informal digital English learning: A meta-analysis of its effectiveness on proficiency, motivation, and self-regulation. Computers & Education: Artificial Intelligence, 7, 100323. Source ↗
  5. Li, Wang & Yang (2025). Can Generative AI Chatbots Promote Second Language Acquisition? A Meta-Analysis. Journal of Computer Assisted Learning, 41, e70060. Source ↗
  6. Şimşek et al. (2026). Artificial intelligence-assisted vocabulary teaching: a meta-analysis study. Frontiers in Artificial Intelligence, 9, 1806700. Source ↗
  7. Ngo, Chen & Lai (2024). The effectiveness of automatic speech recognition in ESL/EFL pronunciation: A meta-analysis. ReCALL, 36(1), 4–21. Source ↗
  8. Zhai & Ma (2023). The Effectiveness of Automated Writing Evaluation on Writing Quality: A Meta-Analysis. Journal of Educational Computing Research, 61(4), 875–900. Source ↗
  9. Ngo, Chen & Lai (2024). The effectiveness of automated writing evaluation in EFL/ESL writing: a three-level meta-analysis. Interactive Learning Environments, 32(2), 727–744. Source ↗
  10. Huang, Xu, Li & Yu (2024). A Meta-analysis of Effects of Automated Writing Evaluation on Anxiety, Motivation, and Second Language Writing Skills. The Asia-Pacific Education Researcher. Source ↗
  11. Chen et al. (2025). A systematic review and meta-analysis of AI-enabled assessment in language learning: Design, implementation, and effectiveness. Journal of Computer Assisted Learning, 41, e13064. Source ↗
  12. Lee (2023). The effectiveness of machine translation in foreign language education: a systematic review and meta-analysis. Computer Assisted Language Learning, 36(1–2), 103–125. Source ↗
  13. Lee, Choe, Zou & Jeon (2025). Generative AI (GenAI) in the language classroom: A systematic review. Interactive Learning Environments, 34, 335–359. Source ↗
  14. Lo et al. (2024). Exploring the application of ChatGPT in ESL/EFL education and related research issues: a systematic review of empirical studies. Smart Learning Environments, 11, 50. Source ↗
  15. Law (2024). Application of generative artificial intelligence (GenAI) in language teaching and learning: A scoping literature review. Computers and Education Open, 6, 100174. Source ↗
  16. Bao et al. (2025). A systematic review of AI in second language acquisition using the expanded SAMR model (2015–2024). Discover Computing, 28, 292. Source ↗
  17. Li, Shadiev & Chiu (2026). Applying Generative Artificial Intelligence to Task-based Language Teaching and Learning: A Systematic Review and Meta-analysis. TechTrends, 70, 150–163. Source ↗
  18. Huang, Hew & Fryer (2022). Chatbots for language learning—Are they really useful? A systematic review of chatbot-supported language learning. Journal of Computer Assisted Learning, 38(1), 237–257. Source ↗
  19. Liu, Ab Rahman & Mohamad Zain (2025). A systematic literature review of research on automatic speech recognition in EFL pronunciation. Cogent Education, 12, 2466288. Source ↗
  20. AI-driven chatbot review (2025). AI-driven chatbots in second language education: A systematic review of their efficacy and pedagogical implications. Ampersand, 14, 100224. Source ↗
  21. Wang, Zou, Du & Wang (2024). The impact of different conversational generative AI chatbots on EFL learners: willingness to communicate, speaking anxiety, self-perceived communicative competence, and performance. System, 127, 103533. Source ↗
  22. Zheng, Zhou, Chen & Ye (2025). The influence of large language models as collaborative dialogue partners on EFL English oral proficiency and foreign language anxiety. Computer Assisted Language Learning. Source ↗
  23. Chen, Koong & Liao (2022). Influences of Integrating Dynamic Assessment into a Speech Recognition Learning Design to Support Students’ English Speaking Skills, Learning Anxiety and Cognitive Load. Educational Technology & Society, 25(1), 1–14. Source ↗
  24. Liakin, Cardoso & Liakina (2017). Learning L2 pronunciation with a mobile speech recognizer: French /y/. CALICO Journal. Source ↗
  25. McCrocklin (2016). Pronunciation learner autonomy: The potential of automatic speech recognition. System, 57, 25–42. Source ↗
  26. Chen (2024). The AI chatbot interaction for semantic learning: A collaborative note-taking approach with EFL students. Language Learning & Technology, 28(1), 1–25. Source ↗
  27. Chen (2025). Improving English semantic learning outcomes through AI chatbot-based ARCS approach. Interactive Learning Environments, 33(6), 3909–3924. Source ↗
  28. Fryer & Carpenter (2006). Bots as language learning tools. Language Learning & Technology, 10(3), 8–14. Source ↗
  29. Han & Li (2024). Exploring ChatGPT-supported teacher feedback in the EFL context. System, 126, 103502. Source ↗
  30. Woo et al. (2024). Exploring AI-generated text in student writing: How does AI help? Language Learning & Technology, 28(2), 183–209. Source ↗
  31. Yang & Chen (2026). ChatGPT and L2 Chinese writing: evaluating the impact of model version and prompt language on automated corrective feedback. Computer Assisted Language Learning, 39(4), 1119–1147. Source ↗
  32. Shi, Chai, Zhou & Aubrey (2025). Comparing the effects of ChatGPT and automated writing evaluation on students’ writing and ideal L2 writing self. Computer Assisted Language Learning. Source ↗
  33. Barrot (2021). Using automated written corrective feedback in the writing classrooms: Effects on L2 writing accuracy. Computer Assisted Language Learning. Source ↗
  34. Dikli & Bleyle (2014). Automated essay scoring feedback for second language writers: How does it compare to instructor feedback? Assessing Writing, 22, 1–17. Source ↗
  35. Stevenson & Phakiti (2014). The effects of computer-generated feedback on the quality of writing. Assessing Writing, 19, 51–65. Source ↗
  36. Wang, Shang & Briody (2013). Exploring the impact of using automated writing evaluation in English as a foreign language university students’ writing. Computer Assisted Language Learning, 26(3), 234–257. Source ↗
  37. Li, Link & Hegelheimer (2015). Rethinking the role of automated writing evaluation (AWE) feedback in ESL writing instruction. Journal of Second Language Writing, 27, 1–18. Source ↗
  38. Link, Mehrzad & Rahimi (2020). Impact of automated writing evaluation on teacher feedback, student revision, and writing improvement. Computer Assisted Language Learning. Source ↗
  39. Jiang, Yu & Wang (2020). Second Language writing instructors’ feedback practice in response to automated writing evaluation: A sociocultural perspective. System, 93, 102302. Source ↗
  40. Chang, Li, Huang & Whitfield (2021). Exploring EFL students’ writing performance and their acceptance of AI-based automated writing feedback. Proceedings of the International Conference on Education Development and Studies. Source ↗
  41. Hassanzadeh & Fotoohnejad (2021). Implementing an automated feedback program for a foreign language writing course: A learner-centric study. Journal of Computer Assisted Learning, 37(5), 1494–1507. Source ↗
  42. Sun & Fan (2022). The effects of an AWE-aided assessment approach on business English writing performance and writing anxiety. Studies in Educational Evaluation, 72, 101123. Source ↗
  43. Tan, Cho & Xu (2022). Exploring the effects of automated written corrective feedback, computer-mediated peer feedback and their combination on EFL writing performance. Interactive Learning Environments. Source ↗
  44. Wilson & Roscoe (2020). Automated writing evaluation and feedback: Multiple metrics of efficacy. Journal of Educational Computing Research, 58(1), 87–125. Source ↗
  45. Xu & Zhang (2021). Understanding AWE feedback and English writing of learners with different proficiency levels in an EFL classroom: A sociocultural perspective. The Asia-Pacific Education Researcher. Source ↗
  46. Cheng (2017). The impact of online automated feedback on students’ reflective journal writing in an EFL course. The Internet and Higher Education, 34, 18–27. Source ↗
  47. Fu, Zou, Xie & Cheng (2022). A review of AWE feedback: Types, learning outcomes, and implications. Computer Assisted Language Learning. Source ↗
  48. Nunes, Cordeiro, Limpo & Castro (2022). Effectiveness of automated writing evaluation systems in school settings: A systematic review of studies from 2000 to 2020. Journal of Computer Assisted Learning, 38(2), 599–620. Source ↗
  49. Lv, Ren & Xie (2021). The effects of online feedback on ESL/EFL writing: A meta-analysis. The Asia-Pacific Education Researcher, 30(6). Source ↗
  50. Mohsen (2022). Computer-mediated corrective feedback to improve L2 writing skills: A meta-analysis. Journal of Educational Computing Research, 60(5). Source ↗
  51. Hattie & Timperley (2007). The power of feedback. Review of Educational Research, 77(1), 81–112. Source ↗
  52. Kang & Han (2015). The efficacy of written corrective feedback in improving L2 written accuracy: A meta-analysis. Modern Language Journal, 99(1), 1–18. Source ↗
  53. Li (2010). The effectiveness of corrective feedback in SLA: A meta-analysis. Language Learning, 60(2), 309–365. Source ↗
  54. Long (1996). The role of the linguistic environment in second language acquisition. In Handbook of Second Language Acquisition. Source ↗
  55. Swain (1985). Communicative competence: Some roles of comprehensible input and comprehensible output in its development. In Input in Second Language Acquisition. Source ↗
  56. Schmidt (1995). Consciousness and foreign language learning: A tutorial on the role of attention and awareness in learning. Attention and Awareness in Foreign Language Learning. Source ↗
  57. Lantolf, Thorne & Poehner (2020). Sociocultural theory and second language development. In Theories in Second Language Acquisition (3rd ed.). Source ↗
  58. Boulton & Cobb (2017). Corpus use in language learning: A meta-analysis. Language Learning, 67(2), 348–393. Source ↗
  59. Plonsky & Oswald (2014). How big is ‘big’? Interpreting effect sizes in L2 research. Language Learning, 64(4), 878–912. Source ↗
  60. Pack & Maloney (2023). Using generative artificial intelligence for language education research: Insights from using OpenAI’s ChatGPT. TESOL Quarterly, 57(4), 1571–1582. Source ↗
  61. Chapelle (2025). Generative AI as game changer: Implications for language education. System, 132, 103672. Source ↗
  62. Stockwell (2024). ChatGPT in language teaching and learning: Exploring the road we’re travelling. Technology in Language Teaching & Learning, 6(1). Source ↗
  63. Alharbi (2023). AI in the foreign language classroom: A pedagogical overview of automated writing assistance tools. Education Research International, 2023, 4253331. Source ↗