GenAI chatbots in language learning: A systematic review of emerging affordances

Synthesis: Lopez and Arabadzhy (2026) review 24 empirical studies of GenAI chatbots in language classrooms published between 2022 and 2025, coding them against the usefulness framework from computer-supported collaborative learning. Their central move is taxonomic rather than experimental: they show that all 11 sub-affordances catalogued for pre-GenAI chatbots still appear, add three new ones (sequential workflow, exploration, translanguaging), and argue for a fourth category they call emotional affordances — engagement, safety, and self-efficacy — nested under utility. The reviewed studies mostly report learners' self-reported experience of chatbot conversation, so the framework describes what teachers and students say the tools enable, not what they have been shown to improve. The result is a useful map for designing tasks around Scaffolding and Self-Efficacy, and a clear statement that the emotional layer is a hypothesis the field has not yet tested.

Key Findings

1. The evidence base is 24 empirical studies from 2022 to 2025. Searches of four EBSCOhost databases returned 541 records; after removing 64 duplicates and 339 off-criteria titles, 68 went to full-text screening and 24 met all criteria.
2. All 11 earlier sub-affordances recur, even when authors never cite the framework. Timeliness, ease of use, personalization, interlocutor, simulation, transmission, helpline, recommendation, and the three communication sub-affordances all reappear in post-2022 classrooms.
3. Three new sub-affordances appear within the existing categories. Sequential workflow (a roadmap of steps, or scaffolded guidance), exploration (unstructured inquiry beyond the assigned task), and translanguaging (asking for contextualized phrases rather than word-by-word translation).
4. A fourth category is proposed: emotional affordances. Engagement, safety, and self-efficacy do not fit technological, pedagogical, or social affordances, so the authors nest them under utility and ground them in Bandura's reciprocal determinism.
5. The corpus is narrow in skill and grouping. Speaking accounts for 11 studies (45.8%) and writing for 9 (37.5%), reading for 1 (4.2%); in 22 studies (91.7%) learners worked individually all the time.
6. Coding was collaborative and checked. Both authors coded 5 articles (21%) independently before discussing categories, then split the rest, reporting inter-coder agreement above 90%.

From usability to utility, then a fourth branch

The review inherits its structure from Kirschner et al.'s (2004) usefulness framework, which separates usability — whether a system lets a user accomplish tasks efficiently — from utility, which in collaborative environments is set by educational and social functionality. That split yields the familiar three categories: technological, pedagogical, and social affordances. Huang et al.'s (2022) review of chatbots then populated them with 11 sub-affordances, from timeliness to cohesive communication. Lopez and Arabadzhy keep the vocabulary but tighten it, reserving "affordance" for the three broad categories and "sub-affordance" for the specific capabilities — a terminological choice they flag as a deliberate departure from Huang et al. Their contribution is the fourth branch. Because engagement, safety, and self-efficacy concern neither tool characteristics, nor instructional design, nor communication, the authors place them under utility as emotional affordances and claim, following Bandura, that the three reinforce one another: confidence makes practice feel safe, and safety sustains engagement. That reciprocity is offered as theoretical grounding, not as a tested mechanism.

What the 24 studies actually document

The corpus is geographically concentrated — China (7), Iran (4), Taiwan (4), Türkiye (4) — and almost entirely inside formal institutions, with only one study set outside one. English is the target language throughout, and ChatGPT in some version is the tool in most rows of the study table, alongside named products such as Elsa Speak and Speeko. The design of the reviewed work matters for how much weight the framework can bear: the authors excluded studies of perceptions, designs, and developments, keeping only empirical work on usability or utility with language learners as participants. What remains is largely self-report inside a single course or intervention, which is enough to establish that a capability was used and noticed — a "roadmap to improving their papers," a non-judging interlocutor — but not enough to show that it changed proficiency. The review is candid that the emotional findings come from learner commentary and that the reciprocal model is an extension of Bandura rather than a result.

Where the review says the field is thin

Three thin patches are explicit. First, grouping: 22 of 24 studies had learners working alone, so the social affordances — the category the framework was built to explain — are observed almost entirely in student-to-chatbot talk rather than student-to-student talk, and the authors call for translanguaging to be examined as a social affordance too. Second, language coverage: bias toward dominant narratives and underperformance in less-resourced languages mean exploratory chat will not be a uniform experience across classrooms, and the review asks instructors to check a tool's capability in the language they teach. Third, the search itself: four databases, one search string in the title field, English-only publications, and peer review's slow cycle all bound what could be found. The emotional category is the largest open question — the authors call for it to be investigated as an outcome, not just described as a perception, since foreign language anxiety and willingness to speak are measurable and were not measured against it.

What this means for practice

• Instructors. Use chatbot practice for the low-stakes rehearsal that learners describe as less threatening than teacher correction, and keep the relational work — the Feedback that carries empathy — with humans, since students in these studies said the tools did not replace it.
• Curriculum designers. Design the workflow rather than leaving it to the chat: studies reporting sequential workflow and exploration show the difference between a bot that scaffolds steps and one that answers whatever is asked, so build the steps into the task.
• Researchers. Treat emotional affordances as constructs to measure, with pre/post anxiety and willingness-to-communicate instruments and grouped rather than solo tasks, instead of coding them from satisfaction comments.

Limitations

• The review covers 24 studies, all published in English and drawn from four EBSCOhost databases with a single title-field search string, so non-English and differently indexed work is absent.
• It excluded perception, design, and development studies, which removes the largest body of GenAI classroom research but also most of the evidence about how tools were built.
• No effect sizes are pooled; the review maps affordances and does not estimate learning gains.
• The emotional category rests on learner self-report and an analogical extension of Bandura's reciprocal determinism, which the reviewed studies did not test.
• Only one reviewed study examined reading, so claims about literacy skill beyond speaking and writing are unsupported.

Citation

Lopez, L., & G. Arabadzhy (2026). GenAI chatbots in language learning: A systematic review of emerging affordances. Language Learning & Technology, 30(1), 1-25.