AI for Learning English: What Actually Works (Science)

"AI-powered" has become a marketing label. That's why it's worth separating types of tools and looking at the evidence for each. At HelloVozi we use AI to support practice consistent with the strategies described in this guide: English speaking practice with feedback, at your level and often.

If you already use ChatGPT, we share ready-to-use prompts in ChatGPT English speaking practice. And if your challenge is moving from understanding to speaking, read understand English but can't speak it?

Does AI work for learning English?

It works as support, with effects ranging from small to large depending on the type of tool. The strongest evidence is for systems that give step-by-step practice with immediate feedback. Chatbots show large effects in short studies that shrink in longer interventions. AI doesn't replace real conversation or the teacher: the best results appear when it's used to practice more, not less.

What the evidence says, tool by tool

Data from each study's published abstract (accessed Sep 29, 2026). Effect sizes aren't directly comparable across studies.

Intelligent tutors and adaptive learning

Adaptive learning adjusts difficulty and the next step based on your answers. Kurt VanLehn (2011) reviewed experiments comparing human tutoring, computer tutoring and instruction without tutoring. Contrary to the belief that human tutors are far more effective, he found an effect of 0.79 for human tutoring and 0.76 for step-based intelligent tutoring systems. Kulik and Fletcher (2016) analyzed 50 controlled evaluations: the median gain was 0.66 standard deviations, though much smaller when measured with standardized tests.

The key point: these reviews study tutoring systems in general, not language apps. They show that immediate feedback and adjusted difficulty work, not that any AI language app has that effect.

Speech recognition and pronunciation

The meta-analysis by Ngo, Chen and Lai (2024) pooled 15 studies on automatic speech recognition with English learners. The overall effect was medium (g = 0.69). It was large when the tool gave explicit corrective feedback and for individual sounds, and small for intonation and rhythm. Short treatments showed no advantage, and practicing with peers worked better than practicing alone.

Chatbots and ChatGPT

Wu and Yu (2024) analyzed 24 randomized studies of AI chatbots and found a large effect on learning, larger in higher education. But there's a catch: short interventions had a bigger effect than long ones, which the authors partly attribute to novelty wearing off over time. For speaking, the study by Karami, Feizarbabi and Xodabande (2026) with 65 learners found more fluency and complexity with guided ChatGPT practice, but no extra gain in accuracy compared with teacher-led class.

Language models also have limits. In their analysis of ChatGPT in education, Kasneci and colleagues (2023) point to the "unexpected brittleness" of these systems, possible bias in their output and the need for human oversight and fact-checking.

Gamification: streaks, points and badges

Gamification uses game elements (points, levels, streaks, challenges) to get you to practice more. Sailer and Homner's (2020) meta-analysis found significant but small effects: g = .49 for cognitive outcomes, .36 for motivation and .25 for study behavior. Only the cognitive effect held up in the most methodologically rigorous studies. It worked best when it included a story (game fiction) and when it combined competition with collaboration.

In practice: a daily streak helps you come back tomorrow, but what you learn depends on what you do in those minutes. A streak of flashcard review won't make you speak if you never produce sentences.

Immersive learning and simulations

Immersive learning simulates real situations (ordering at a restaurant, a job interview, a meeting) so you can practice them before living them. It can be done with virtual reality or, more simply, with role-plays and conversation simulations. Parmaxi's (2023) review of 26 papers from 2015 to 2018 concludes that virtual reality is a valuable tool in the language classroom, but with technical and pedagogical challenges.

You don't need a headset to benefit from the idea: what works is practicing a specific situation with a goal, in a setting where mistakes have no consequences.

How to choose an AI tool for learning English

1. Does it make you produce? Look for tools that make you speak or write, not just pick answers.
2. Does it give specific feedback? The strongest evidence (intelligent tutors, speech recognition) is for immediate, explicit feedback.
3. Does it adapt to your level? If everything is too easy or too hard, you don't progress.
4. Would you use it for months? Novelty wears off. Choose something that fits your daily routine.
5. Is it honest about what it doesn't do? Be wary of promises like "fluent in 30 days".

What AI still doesn't do well

• It doesn't replace talking with people. Real interactions have accents, interruptions and cultural nuance.
• It can be wrong. Double-check doubtful grammar explanations: fact-checking is one of the recommendations of Kasneci et al. (2023).
• It doesn't guarantee accuracy. In Karami et al. (2026), accuracy didn't improve more than with traditional classes.
• It won't be consistent for you. Ten minutes a day with any tool beats an occasional hour with the best one.

Next step: if you don't know your level, take the English level test (about 15 minutes, with instant results). To keep working on this topic, see how to practice English speaking with AI in Vozi Coach.