AI-Assisted Language Learning Evidence Map, 2023–2026

Executive finding

Recent studies suggest that AI chatbots and speech-feedback tools can increase practice volume and improve selected speaking outcomes in structured short-term interventions. The evidence does not support the broader claim that an AI tutor is proven to create general language fluency or replace a qualified teacher and real human interaction. Much of the available research uses small or institution-specific samples, short interventions, bundled treatments, and outcomes that do not test long-term transfer.

The defensible product position is therefore: AI can be a scalable practice layer whose outcomes depend on task design, feedback quality, learner use, and connection to instruction—not an autonomous proof of learning.

Scope and method

This is a rapid evidence map, not a systematic review or meta-analysis. It was prepared on 20 July 2026 from peer-reviewed studies and reviews discoverable in English, with emphasis on 2023–2026 research about chatbots, generative AI, speech practice, and oral outcomes. Studies were mapped by intervention, comparison, outcome, and transfer evidence. Marketing reports and preprints may inform the questions but do not receive the same weight as peer-reviewed controlled research.

The map should be refreshed before publication and annually thereafter because models, product behaviour, and the study base are changing quickly.

What the recent study base contains

Evidence type	What it can tell us	Recurring limitation
Controlled classroom studies	Difference between an AI-supported condition and a comparison during a course	Often one institution, one teacher context, and a short intervention
Pre/post single-group studies	Whether participants changed during use	Cannot isolate AI from time, instruction, motivation, or test familiarity
Mixed-method studies	Test change plus learner experience	Self-report may reflect novelty or satisfaction rather than skill transfer
Systematic/scoping reviews	Shape and gaps of the accumulated field	Conclusions inherit the quality and heterogeneity of included studies
Product efficacy reports	Detailed data from a specific course or user sample	Selection, funding, author affiliation, and outcome scope require scrutiny

The category “AI chatbot” also hides major differences. One intervention may be text-only, another voice-based, another a scripted conversational agent, and another a large language model combined with teacher-created tasks. They should not be treated as one reproducible treatment.

Selected recent evidence

Chatbot-supported speaking practice

A 2025 study in Social Sciences & Humanities Open reports 120 undergraduate participants, with 60 assigned to an experimental chatbot-practice group. It examined fluency, pronunciation, grammar, vocabulary, and confidence. The design is more informative than a satisfaction survey, but a reader still needs the allocation procedure, intervention fidelity, rater blinding, effect sizes, and delayed outcomes before generalising. Source: https://www.sciencedirect.com/science/article/pii/S2590291125006618

A 2025 TESOL Quarterly study examined ChatGPT-mediated oral summary tasks against a control condition. Its task-specific design is valuable because it tests a defined speaking activity rather than an undefined promise of fluency. The result should be interpreted for oral summarisation and the studied learner population, not automatically for spontaneous social conversation. Source: https://doi.org/10.1002/tesq.70001

A 2025 Humanities and Social Sciences Communications mixed-method study examined an AI-powered conversation application with 60 learners in an IELTS preparation setting. It reports speaking and anxiety outcomes, but the app, course context, participant population, and test-preparation goal are part of the treatment. Replication across languages, providers, and delayed real-world tasks remains necessary. Source: https://www.nature.com/articles/s41599-025-05550-z

Reviews of AI-supported language learning

A 2024 systematic review of AI tools and learner self-regulation included 18 peer-reviewed articles from 2009–2024. It maps tools such as chatbots and automated feedback across language skills. The small, diverse corpus illustrates both promise and the difficulty of making one pooled efficacy claim. Source: https://doi.org/10.1080/2331186X.2024.2433814

A 2025 systematic review focused on chatbot effects on English learners' speaking proficiency. It is useful for locating studies, but review conclusions still depend on whether the included papers used comparable speaking constructs, credible controls, and independent ratings. Source: https://riset.unisma.ac.id/index.php/JREALL/article/view/23866

What appears promising

More available speaking turns

On-demand tools remove scheduling and social-cost barriers. This mechanism is credible even before making an efficacy claim: a learner can repeat a role-play or request another question without consuming a partner's time. Research should measure whether this access produces more spoken minutes and whether those minutes predict later task performance.

Controlled variation

An AI system can keep the function constant while changing the setting, vocabulary, or listener stance. That supports retrieval and transfer better than one memorised script if the variations remain accurate and level-appropriate.

Lower-stakes rehearsal

Several studies examine confidence, willingness to communicate, or anxiety. Private rehearsal may help some learners approach human interaction, but lower self-reported anxiety inside the tool is not proof of reduced anxiety with an examiner, employer, or unfamiliar community member.

Structured feedback at scale

Transcripts, recurring-error summaries, and delayed feedback can focus review. The unresolved question is validity: a system's confidence, pronunciation score, or correction is only useful if it measures the intended construct reliably across accents, languages, devices, and learner groups.

What remains unproven

• General long-term fluency gains across unrelated speaking tasks.
• Equivalent outcomes across languages and writing systems.
• Durable transfer after the tool or course ends.
• Reliable cultural and pragmatic advice without human verification.
• Fair automated speech scoring across accent, disability, device, and acoustic conditions.
• Replacement of trained teachers for diagnosis, safeguarding, motivation, and high-stakes assessment.
• Safe and eligible use for every age group, institution, and jurisdiction.

Minimum standard for future LinguaLive claims

Any efficacy page should name:

1. the exact learner population and language;
2. the intervention version and duration;
3. what the comparison group actually received;
4. the outcome task and scoring process;
5. assessor independence or blinding;
6. attrition and actual usage;
7. effect sizes with uncertainty, not percentages alone;
8. delayed and unpractised transfer tasks;
9. adverse events, privacy, and accessibility findings;
10. funding, author affiliation, preregistration, and data availability.

Practical conclusion

The current evidence supports testing AI as a way to deliver more deliberate, structured speaking practice. It does not justify “proven fluency,” “teacher replacement,” or universal learning claims. A strong pilot should measure actual speaking time, task completion, intelligibility, pauses and repairs, human-transfer performance, and sustained use—then publish null findings and limitations alongside gains.

Editorial disclosure

LinguaLive produces an AI speaking product and therefore has a commercial interest in this topic. A research-methods reviewer should verify the search, study descriptions, and interpretations before canonical publication. This brief must remain clearly labelled as a rapid evidence map, not original experimental research.