What actually helps you learn a foreign language — and what works only in the imagination of marketers? This article examines only what is backed by meta-analyses and randomised controlled studies: spaced repetition, corrective feedback, task-based learning, explicit grammar, pronunciation training, and the role of AI. Along the way, we debunk the most stubborn myths — from "learning styles" to "just listen and everything will come naturally."
Why Do Experts Disagree?
The answer lies in the hierarchy of evidence. A meta-analysis combines results from many primary studies, averages them through a standardised effect size, and identifies which findings are reproducible and which are chance. Plonsky & Oswald (2014) analysed 346 primary studies and 91 meta-analyses in SLA (N > 604,000), demonstrating that typical effects in this field are systematically higher than Cohen's thresholds. For SLA: d ≈ 0.4 = "small," d ≈ 0.7 = "medium," d ≈ 1.0 = "large." DOI: 10.1111/lang.12079.
It is also important to distinguish between between-group (experimental vs. control) and within-group (pre–post) effect sizes. Pre–post effects are always larger because they include natural development. Lab studies show larger effects than classroom studies (Li, 2010) — apply lab numbers to practice with caution.
Memory: Retrieval Practice and Spaced Repetition
Rowland (2014) — 47 studies (N = 6,801): d = 0.87 in favour of retrieval practice over restudy. Adesope et al. (2017) in Review of Educational Research: testing consistently outperforms re-reading. DOI: 10.3102/0034654316689306. Instead of re-reading your word list, close it and try to recall the translation.
Kim & Webb (2022) — 98 effects from 48 experiments (N = 3,411): distributed practice yields a medium-to-large effect vs. massed; short and long intervals equal on immediate tests but short inferior on delayed; equal and expanding intervals statistically equivalent. DOI: 10.1111/lang.12479. Kim & Webb (2023) with 150 Korean students, 48 words: on a 2-week delayed test, spaced practice significantly better. Cepeda et al. (2008): optimal interval is 10–20% of the target retention interval. DOI: 10.1111/j.1467-9280.2008.02209.x.
Schimke et al. (2021) — 25 studies (N = 1,396): sleep after learning yields g = 0.50 vs. wakefulness (between-group); recall g = 0.57, recognition g = 0.52, lexical integration g = 0.39. DOI: 10.3758/s13423-021-01980-3. Practical recommendation: schedule vocabulary sessions before sleep.
Speaking: Task-Based Learning and Feedback
Bryfonski & McKay (2017) — 52 studies: d = 0.93; effect stronger with needs analysis. DOI: 10.1177/1362168817744389. Language is better acquired through real communicative tasks than isolated drills.
Lyster & Saito (2010) — 15 studies (N = 827): prompts more effective than recasts; effect stronger in younger learners. DOI: 10.1017/S0272263109990520. Li (2010) — 33 studies: explicit feedback more effective than implicit. Hattie & Timperley (2007): d = 0.48 for feedback overall. Swart et al. (2019): post-reading feedback more effective than during-reading; showing the correct answer with explanation beats simple correct/wrong: g+ = 0.35. DOI: 10.1016/j.edurev.2019.100296.
Lexy note: In lessons 1–15, Lexy intentionally does not correct errors, to maintain conversational flow and reduce the affective barrier.
Tabari et al. (2025) — 31 studies: task repetition in written production d = 1.19; lexical complexity d = 0.75; fluency d = 0.50. Each repetition reduces cognitive load.
Grammar: Explicit vs. Implicit
Norris & Ortega (2000) — 49 studies: explicit instruction outperforms implicit; effect stable over time. DOI: 10.1111/0023-8333.00136. Spada & Tomita (2010) — 41 studies: confirmed the advantage. Shintani et al. (2013): input instruction better for receptive knowledge (immediate); production instruction better for productive knowledge (delayed). DOI: 10.1111/lang.12001. Honest position: Lexy focuses on oral practice; learners seeking systematic grammar should supplement with explicit explanations.
Input: Reading, Listening, Subtitles
Nakanishi (2015) — 34 extensive reading studies (N = 3,942): d = 0.46 vs. control (between-group), d = 0.71 pre–post. DOI: 10.1002/tesq.157. Montero Perez et al. (2013) — subtitled video: large effect on listening and vocabulary. DOI: 10.1016/j.system.2013.07.013. Uchihara et al. (2019) — 26 studies (N = 1,918): r = 0.34 between word encounters and incidental acquisition. DOI: 10.1111/lang.12343. Kim et al. (2020): L1 glosses more effective than L2 glosses (g = 0.33), especially for beginners. DOI: 10.1177/1362168820981394.
Pronunciation
Lee, Jang & Plonsky (2014) — 86 studies: pronunciation instruction d = 0.89; stronger with feedback and longer interventions. DOI: 10.1093/applin/amu040. Sakai & Moorman (2017): perceptual training d = 0.54. DOI: 10.1017/s0142716417000418. Uchihara et al. (2024) — 31 studies: gain 10.50% trained, 4.50% untrained. DOI: 10.1017/s0142716424000195. Ngo et al. (2023) — 15 studies: speech recognition g = 0.69; large for individual sounds, small for intonation; partner practice more effective. DOI: 10.1017/s0958344023000113.
Technology and AI
Wang et al. (2024) — 28 studies, 70 effects: g = 0.484 (moderate); depends on education level, proficiency, interface design, interactivity. DOI: 10.3102/003465432412556. VanLehn (2011): human tutors g = 0.79, step-based ITS g = 0.76, problem ITS g = 0.40. DOI: 10.1037/a0037123. Lyu et al. (2024) — 51 studies: large effect for generative AI; stronger in informal learning, productive skills, less common languages. DOI: 10.1007/s10791-026-10015-1. De Simone et al. (2025, World Bank) — RCT Nigeria, 6 weeks Copilot/GPT-4: +0.31 SD overall, +0.23 SD English (between-group). hdl.handle.net/10986/43212.
Wu & Yu (2023) — 24 RCTs: short interventions show larger effects — novelty effect. DOI: 10.1111/bjet.13334. Bastani et al. (2025): GPT-4 access during learning improved scores (48% better); but access during tests led to dependency and worse results. Critical warning: AI that does the task for the student undermines learning. Gollwitzer & Sheeran (2006) — 94 studies: "if-then" planning d = 0.65 for intention implementation. DOI: 10.1016/S0065-2601(06)38002-1.
Evidence Table
| Method | Level | Key Result | Source |
|---|---|---|---|
| Retrieval practice | A | d = 0.87 (47 studies, N = 6,801) | Rowland (2014) |
| Spaced repetition | A | Medium–large effect (48 exp., N = 3,411) | Kim & Webb (2022) |
| Sleep after learning | A | g = 0.50 (25 studies, N = 1,396) | Schimke et al. (2021) |
| TBLT | A | d = 0.93 (52 studies) | Bryfonski & McKay (2017) |
| Corrective feedback (prompts) | A | Prompts > recasts (15 studies, N = 827) | Lyster & Saito (2010) |
| Explicit grammar | A | Explicit > implicit (49 studies) | Norris & Ortega (2000) |
| Pronunciation instruction | A | d = 0.89 (86 studies) | Lee et al. (2014) |
| Extensive reading | A | d = 0.46 vs. control (34 studies, N = 3,942) | Nakanishi (2015) |
| Task repetition | A | d = 1.19 (31 studies) | Tabari et al. (2025) |
| AI chatbots | B | g = 0.484 (28 studies, 70 effects) | Wang et al. (2024) |
| Generative AI | B | Large effect; novelty noted (51 studies) | Lyu et al. (2024) |
| Speech recognition | B | g = 0.69 (15 studies) | Ngo et al. (2023) |
| Learning styles | C | Hypothesis not supported | Pashler et al. (2008) |
| Growth mindset | C | Weak overall effects | Sisk et al. (2018) |
| Highlighting / re-reading | C | Low utility | Dunlosky et al. (2013) |
Popular Myths
Learning styles: Pashler et al. (2008) — no evidence for the "meshing" hypothesis. DOI: 10.1111/j.1539-6053.2009.01038.x. Growth mindset: Sisk et al. (2018) — weak overall effects; at-risk groups benefit modestly. DOI: 10.1177/0956797617739704. Highlighting and re-reading: Dunlosky et al. (2013) — low utility. Interleaving: Brunmair & Richter (2019) — helpful for similar categories, caution for words and descriptive texts. DOI: 10.1037/bul0000209. App research: Loewen et al. (2019) — 9 participants; Smith et al. (2024) — 48 participants, single group — insufficient for reliable conclusions.
How Lexy Aligns with Research
Lexy is a conversational AI trainer for oral practice across 49 languages. Each 25-minute lesson consists of seven connected stages: Vocabulary, Warm-up, Lesson, Exercises, Dialogue, Story, and Memory Flashcards.
- Stage 1 (Vocabulary): Flashcards and quiz: pure retrieval practice (Rowland, 2014; Adesope et al., 2017, d = 0.87).
- Stage 2 (Warm-up): Written chat with Lexy: prepares articulation and motor typing before voice dialogue.
- Stage 3 (Lesson): Lexy asks → student answers aloud → student asks → Lexy answers without typing. Combines retrieval practice with TBLT (Bryfonski & McKay, 2017, d = 0.93).
- Stage 4 (Exercises): Keyboard input, tests, and listening: supports receptive and productive knowledge while relieving vocal fatigue (Shintani et al., 2013).
- Stage 5 (Dialogue): Two-voice authentic audio: managed input (Montero Perez et al., 2013; Uchihara et al., 2019).
- Stage 6 (Story): Narrated story with translation: extensive reading in miniature (Nakanishi, 2015; L1 glosses: Kim et al., 2020).
- Stage 7 (Memory Flashcards): Spaced retrieval of previous lesson phrases from lesson 2 onward: long-term consolidation (Kim & Webb, 2022).
- AI voice trainer: Speech recognition + generative AI → g = 0.484 (Wang et al., 2024) with novelty-effect caveat (Wu & Yu, 2023).
- No corrections in lessons 1–15: Deliberate to maintain flow and reduce affective barrier.
- Roleplay scenarios (A2+): 24 real-world scenarios with Bingo goals: advanced TBLT.
- 16 lessons: starter course. Sustained proficiency requires months of regular practice.
FAQ
Do I need explicit grammar study if I practise speaking?
Meta-analyses (Norris & Ortega, 2000; Spada & Tomita, 2010) show explicit grammar is more effective than passive immersion for most learners. Supplement speaking practice with structured explanations, especially for productive skills on delayed tests (Shintani et al., 2013).
How many word encounters does it take to learn a word?
Uchihara et al. (2019) found r = 0.34 between encounters and incidental acquisition (26 studies, N = 1,918). The exact number depends on word difficulty, context, and mode of encounter. Active retrieval is far more efficient than passive encounters.
Can I learn a language with an AI app alone?
Current data show moderate-to-large effects (Wang et al., 2024; Lyu et al., 2024), but long-term studies are rare and the novelty effect is real (Wu & Yu, 2023). Combine AI practice with live conversation for optimal results.
Does watching TV shows in the target language help?
Subtitled video yields a large effect on listening and vocabulary (Montero Perez et al., 2013). Word encounter frequency correlates with acquisition (r = 0.34). But passive watching is less effective than active retrieval. Best: pause, repeat, note new words, use them immediately.
Research Limitations
- Publication bias: Positive results are over-represented. Real effects may be somewhat lower.
- Sample heterogeneity: Average effects may not apply to your specific context.
- Short study duration: Long-term data (months, years) are much thinner.
- AI novelty: Most AI studies date from 2020–2024; long-term effectiveness needs further study.
- Retracted research: Wang & Fan (2025) DOI: 10.1057/s41599-025-04787-y — retracted. Not used in this article.
Seven Key Conclusions
- Active retrieval (d = 0.87) outperforms passive re-reading — test yourself.
- Spaced practice significantly outperforms massed study on delayed tests; optimal interval is 10–20% of the retention target.
- Sleep after learning improves word retention (g = 0.50) — schedule vocabulary before bed.
- Task-based learning (d = 0.93) — speak from the very start.
- Explicit grammar generally outperforms implicit immersion — supplement speaking with structured explanations.
- AI tools show moderate-to-large effect (g = 0.484) as a trainer, not a substitute; dependency undermines learning.
- Learning styles, highlighting, re-reading are myths. Invest time in high-evidence methods.
Try Lexy free at app.lexylearn.com.
Key Takeaways
- Retrieval practice (d = 0.87) beats re-reading — test yourself.
- Spaced repetition outperforms massed study, especially on delayed tests.
- Sleep after learning improves word retention (g = 0.50).
- Task-based learning (d = 0.93) — speak from minute one.
- Explicit grammar outperforms implicit immersion per meta-analyses.
- AI chatbots yield moderate effects (g = 0.484); watch for novelty and dependency.
- Learning styles and highlighting are myths without evidentiary support.
Test Your Knowledge: Article Quiz
90-second quiz: check whether you understood the key themes of this article. 5 quick questions. If you make a mistake, we explain it right away.
Try Evidence-Based Methods
Lexy combines retrieval practice, task-based learning, and an AI voice trainer — all in one 25-minute lesson.
Start Free Lesson →