Beyond the interface: Supporting beginner-level CEFR mediation and learner autonomy with ChatGPT

1 CEFR mediation and the companion volume

The CEFR Companion Volume (Council of Europe, 2020) defines mediation as the process of constructing meaning by relaying, simplifying, or adapting content for others. Mediation is organized into three categories: mediating texts, mediating concepts, and facilitating communication. These categories reflect integrated communicative skills, such as summarizing, clarifying, paraphrasing, and adapting messages for specific audiences. While originally developed for human-to-human communication, CEFR descriptors provide a robust framework for analysing human–AI interactions.

2 Sociocultural theory and AI as a mediating tool

Rooted in Vygotsky’s (1978) concept of the Zone of Proximal Development, sociocultural theory emphasizes that learning occurs through mediated interaction. While ChatGPT is technically a tool, in the context of this study it functions as a dialogic mediator, an agent-like interface that interacts responsively, shaping learner output and co-constructing meaning. Following Lantolf (2000), we conceptualize ChatGPT as a semi-agentive scaffold: a system that offers dialogic affordances similar to peer interaction, even though it lacks human intentionality or consciousness.

Although dialogic mediation and AI-assisted scaffolding both involve supportive processes during learning, they differ significantly in their dynamics and implications for learner agency. AI-assisted scaffolding typically refers to instructional support that enables learners to complete tasks beyond their current capabilities. This support often includes vocabulary assistance, grammar correction, or content explanations and tends to be system-initiated and unidirectional in nature. In contrast, dialogic mediation involves reciprocal interaction in which meaning is actively co-constructed through negotiation, clarification, and adaptation.

In the context of this study, ChatGPT is understood not merely as a scaffold that delivers linguistic assistance but as a dialogic mediator that participates in responsive exchanges. These exchanges enable learners to formulate questions, refine their output, and critically reflect on the interaction. While ChatGPT does not possess intentionality or consciousness, its dialogic functionality creates opportunities for learners to engage in meaning-making processes and develop metacognitive strategies. This interpretation aligns with the CEFR’s conceptualization of mediation as a collaborative, audience-aware activity that integrates comprehension, production, and reflection.

3 Learner autonomy and self-regulated learning

Learner autonomy, defined as the ability to take charge of one’s own learning (Holec, 1981), is closely linked to self-regulated learning processes such as planning, monitoring, and evaluating (Little, 1991; Zimmerman, 2002). In AI-mediated environments, this autonomy includes initiating interactions, managing digital tools, and critically evaluating feedback. Previous research indicates that autonomous learners tend to perform better in computer-assisted language learning (CALL) when they are given opportunities for reflection and control (Reinders & Benson, 2017).

In this study, learner autonomy refers broadly to the learner’s capacity to direct their own language learning, both during tasks and in independent study. It involves behaviours such as initiating interactions, selecting appropriate strategies, and responding adaptively to feedback. Within this broader framework, metacognitive autonomy is defined as the ability to reflect on, evaluate, and regulate one’s strategic use of language tools and resources. It includes actions such as setting learning goals, critically assessing AI-generated feedback, comparing digital tools, and engaging in post-task reflection. Whereas learner autonomy is often observable during task performance, metacognitive autonomy tends to emerge more clearly through reflective analysis after the learning activity.

4 Critical digital and AI literacy

As learners increasingly engage with AI systems, developing critical digital literacy is essential, not only for navigating tools effectively but also for evaluating the reliability, accuracy, and socio-technical implications of their outputs. Pangrazio and Selwyn (2019) argue that digital literacy must extend beyond functional skills to encompass reflective and ethical dimensions of technology use. Similarly, Jones and Hafner (2012) emphasize that language learners must critically examine how digital tools shape communication, identity, and knowledge production. Buckingham (2010) further highlights that critical digital literacy includes awareness of the power dynamics embedded in digital environments and algorithmic outputs.

AI literacy frameworks (Luckin et al., 2016) build on these perspectives, urging learners and educators to understand not only how AI systems work but also how to question their assumptions, biases, and pedagogical roles. In the context of AI-mediated language learning, this form of literacy intersects with learner autonomy and reflective practice, fostering a capacity to use AI tools like ChatGPT not just as linguistic aids, but as objects of inquiry and critique.

5 Mediation in multilingual contexts

This study also addresses concerns about linguistic bias in LLMs, which are often trained primarily on English data (Brown et al., 2020; Seghier, 2023). To explore mediation in non-English environments, this research was conducted in Spanish, a major global language that is underrepresented in LLM training data. Participants were multilingual students whose native languages were neither English nor Spanish. This multilingual setting allowed for the analysis of cross-linguistic mediation, language alternation, and meaning-making using learners’ complete linguistic repertoires.

6 AI in language learning: Efficacy, autonomy, and multilingual affordances

A wave of 2025 research has expanded our understanding of how conversational and generative AI can be harnessed in language education, offering robust empirical evidence and nuanced insights into learner experience. A meta-analysis by Gu (2025) synthesising 55 studies reported consistent positive effects of conversational AI on language proficiency, motivation, and learner engagement, underscoring its pedagogical value across diverse contexts. These benefits are often linked to adaptive learning pathways, with Nwanakwaugwu et al. (2025) demonstrating that AI-driven personalization promotes learner autonomy by tailoring instruction to individual needs and proficiencies. In the affective domain, Kittredge et al. (2025) found that generative AI integrated into mobile language applications significantly enhanced learners’ self-efficacy and communicative confidence, while Kruk and Kałużna (2024) identified its capacity to foster motivation, positive emotional engagement, and translation skills development. Beyond discrete skills, Lin et al. (2025) highlighted that AI-supported multimodal composition tasks can strengthen critical thinking, digital literacy, and multilingual competence, particularly in multicultural classrooms. Furthermore, Akhter et al. (2025) showed that embedding AI activities within a CEFR-informed framework measurably improves EFL learners’ communication skills, reinforcing the relevance of CEFR descriptors in structuring AI-mediated pedagogy. Collectively, these 2025 studies establish a contemporary evidence base for examining how AI tools such as ChatGPT can facilitate CEFR-based mediation and learner autonomy in multilingual beginner-level contexts.

1 Participants

The study involved 24 university students from the Czech Technical University, all enrolled in the Computer Science and Open Informatics program. Participants were purposefully selected based on their prior exposure to AI tools, ensuring that the study’s focus remained on language learning and mediation, rather than on digital literacy or tool familiarization. Given their routine engagement with advanced technologies, their digital competencies were presumed to be high, and no formal assessment of digital skills was conducted. Although no pre-task questionnaire was administered, all participants informally confirmed prior experience with ChatGPT or comparable large language models and reported confidence using conversational AI interfaces. This was further supported by the program’s academic curriculum, which includes a detailed study of large language models (LLMs) within artificial intelligence and natural language processing courses.

A CEFR-aligned diagnostic test was administered to assess Spanish language proficiency. Based on the results, students were divided into two groups: Group 1 (A1, n = 12) and Group 2 (A2, n = 12). Participants ranged in age from 19 to 26 years and included both bachelor’s and master’s students (16 males and 8 females). Their linguistic backgrounds were diverse. In addition to Spanish, participants used Czech, Slovak, Ukrainian, French, German, English, Italian, and Serbian during the tasks, languages that supported contrastive mediation and metalinguistic reflection.

2 Task design and procedure

The activity was divided into two parts to address the CEFR categories of mediation: Mediating Texts, Mediating Concepts, and Facilitating Communication. The tasks also promoted learner autonomy and digital critical engagement.

The study was conducted over a one-month period, from March 1 to March 31, 2024. During this time, students completed a series of structured language learning tasks in individual sessions with ChatGPT. Each session was designed to be completed within a flexible timeframe, allowing students to engage with the tasks at their own pace outside of classroom settings. All tasks were completed independently to ensure that learners interacted with the AI autonomously and without influence from peers or instructors. Students received activities with model prompts, described in detail in Appendix A, which they could use as given or modify to suit their preferences. They were encouraged to adapt the tasks to their needs, seek clarification on relevant points, and explore topics of personal interest in greater depth. The aim was to maximize opportunities for individualized task adaptation.

c Duration and scope

The activity was designed for short, focused blocks to keep cognitive load manageable for beginner learners working alone. The recommended schedule is presented in Table 1, which summarizes the timing and sequencing of the activity steps for both parts of the task.

OPEN IN VIEWER

Table 1.

Recommended time allocation per activity step.

Step	Activity description	Estimated time
Comprehension:
1	Summary read/compare	10–15 min
2	Vocabulary translate/explain + quick re-read	15–25 min (A1 ≈ 15–20 min; A2 ≈ 20–25 min)
3	Answer five comprehension questions	10–15 min
4	Review corrections + revise answers	10–15 min
5	Read explanation, find evidence, adjust	10–15 min
Discussion:
1	Expressions list + quick rehearsal	8–12 min
2	Role-play discussion with corrections	20–30 min (A1 ≈ 20–25 min; A2 ≈ 25–30 min)

This structure was presented as a guideline, but actual completion times varied considerably among students. Some reported esta conversación ha durado aproximadamente entre 20 y 30 minutos (‘this conversation lasted approximately between 20 and 30 minutes’), while others indicated aproximadamente entre 60 y 90 minutos (‘approximately between 60 and 90 minutes’). Given that ChatGPT does not have an internal timer to record interaction length, these self-reported durations cannot be verified with complete accuracy. The average self-reported completion time across all participants was 63 minutes. There was no statistically significant difference between the A1 and A2 groups: while A2 students tended to produce longer responses, A1 students reported needing more time to think through their replies. This variation in timing does not undermine comparability, as both groups completed the full set of steps in the prescribed order, with differences reflecting individual pacing preferences rather than unequal task exposure.

3 Data collection

Primary data included anonymized transcripts from student–ChatGPT interactions. Students also submitted written reflections in their native language to elicit more detailed insights into their experiences, perceived benefits, challenges, and suggestions.

4 Data analysis

b Qualitative analysis of student feedback

Reflections were analysed using a hybrid approach combining thematic analysis with grounded theory techniques. Drawing on autonomy and digital literacy frameworks (Holec, 1981; Little, 1991; Luckin et al., 2016; Pangrazio & Selwyn, 2019; Zimmerman, 2002), the analysis proceeded in two stages:

Stage 1: Thematic Analysis: Five overarching themes were identified:

Planning and strategy selection

Monitoring and reflection

Autonomous interaction

Critical evaluation of AI

Tool comparison and digital navigation

Stage 2: Grounded Analysis: Open coding was employed to label specific learner behaviours (e.g. self-correction, prompting AI, rejecting suggestions). These were grouped into axial categories such as metacognitive engagement, digital tool use, and language control. A collaborative coding process ensured reliability, with recurring themes compared among participants to ensure consistency and thematic saturation. Tables 2 and 3 present the coding schemes used to analyse learner autonomy and critical engagement, respectively.

OPEN IN VIEWER

Table 2.

Learner autonomy coding scheme.

Code	Category	Sub-category	Definition	Example
LA-1	Planning and strategy	Explicit goal-setting	Student defines learning goals during interaction.	My goal is to practice vocabulary related to festivals.
LA-2	Monitoring and reflection	Self-correction	Student identifies errors or evaluates their progress.	Oops, I used the wrong tense here.
LA-3	Independent action	Autonomous initiation	Student initiates tasks or prompts without instruction.	Student spontaneously asks ChatGPT for examples.

OPEN IN VIEWER

Table 3.

Critical engagement coding scheme.

Code	Category	Sub-category	Definition	Example
CE-1	Evaluating AI responses	Explicit scepticism	Student questions accuracy or usefulness of AI output.	Is that really correct? It doesn’t match what I remember.
CE-2	Identifying errors	Error identification	Student spots factual or linguistic errors in AI responses.	I think ChatGPT got the tense wrong.
CE-3	Tool comparison	Cross-referencing sources	Student compares AI output with other tools or resources.	I checked this sentence in DeepL and got something different.

1 Learner use of ChatGPT for CEFR-based mediation tasks (research question 1)

Spontaneous prompts across groups

Structured prompts (Appendix A) were excluded from the quantitative analysis, as they were identical for all participants. Only four students from the A1 group and four from the A2 group adapted the model prompts, while the remaining students used them in their original form. Instead, the analysis focused on spontaneous prompts, that is, learner-generated requests and questions that emerged naturally during AI interaction. A visual representation of these proportions is shown in Figure 1.

A1 learners: 92.6% of spontaneous prompts fell into Mediating Concepts, with only 3.7% each for Mediating Texts and Facilitating Communication.

A2 learners: 77.8% were Mediating Concepts, 18.9% Mediating Texts, and 3.3% Facilitating Communication.

This distribution suggests both groups relied primarily on conceptual mediation beyond the given tasks. However, A2 learners more often initiated additional text-based mediation (e.g. clarifications, summaries, vocabulary checks), while A1 learners focused on idea clarification and maintaining basic understanding. In both groups, spontaneous Facilitating Communication prompts were rare, reflecting a focus on comprehension over discourse management.

OPEN IN VIEWER

Figure 1.

Distribution of spontaneous prompt types by proficiency level.

2 Observable behaviours of learner autonomy in AI-mediated tasks (research question 2)

Interpretive contrast

To be considered as having achieved full autonomy, students were required to demonstrate all three aspects of learner autonomy: initiative, strategic behaviour, and self-regulation. Only one student in the A1 group met this criterion. Although A1 learners had lower proficiency, they exhibited slightly higher levels of initiative and self-regulation, which may suggest a greater willingness to manage the flow of interaction. In contrast, A2 learners produced more advanced linguistic structures but relied more heavily on AI confirmation.

Figure 2 illustrates students’ autonomy behaviours across the A1 and A2 groups. Initiative was the most frequently observed behaviour in both groups (83% in A1 and 75% in A2). However, A1 students demonstrated higher levels of self-regulation (42%) and strategic behaviour (25%) than A2 students (17% each), suggesting a stronger tendency toward self-directed learning among A1 participants.

OPEN IN VIEWER

Figure 2.

Autonomy behaviours by proficiency level.

3 Critical engagement with ChatGPT (research question 3)

Interpretive contrast

Critical engagement behaviours were more frequent in A2 (especially error detection), but still heavily skewed toward technical rather than substantive challenges. Across both groups, the absence of real-time verification suggests over-reliance on ChatGPT output during interaction.

Figure 3 presents students’ critical engagement with ChatGPT during the task. A2 students showed slightly higher levels of questioning AI output (42%) compared to A1 (33%) and identified errors more frequently (50% vs. 17%). No instances of real-time verification were observed in either group, indicating a limited tendency to cross-check ChatGPT’s responses using external tools.

OPEN IN VIEWER

Figure 3.

Critical thinking behaviours by proficiency level.

4 Linguistic complexity (cross-cutting observation)

This section synthesizes linguistic complexity findings across both learner groups (A1 and A2), highlighting syntactic variety, lexical richness, discourse management, and common developmental patterns.

A2 learners

Syntactic variety:

Frequent coordination and subordination, including concessives (aunque), contrastives (sin embargo), and relative clauses (que tiene más transporte).

Multi-part arguments with introductions, justifications, and conclusions.

Present indicative dominant, with some future/conditional for hypotheticals (podría, sería mejor).Example: Aunque la ciudad tiene más oportunidades, el pueblo es más tranquilo y seguro para los niños.

Lexical richness:

Topic-specific lexis from the pueblo vs. ciudad task (infraestructura, coste de vida, congestión).

More varied evaluatives (ventajoso, económico), intensifiers (mucho más, bastante), exemplifiers (por ejemplo).

Broader synonym range than A1.Example: Vivir en un pueblo es más económico y ofrece una mejor calidad de vida, por ejemplo, menos estrés y menos contaminación.

Pragmatic markers and discourse management:

Regular connectors for addition (además), contrast (sin embargo), concession (aunque), exemplification (por ejemplo).

Balanced argumentation: acknowledgment of counterarguments before preference.

Clear paragraph-style structuring in oral turns.Example: Entiendo que en la ciudad hay más transporte, sin embargo, en el pueblo hay más tranquilidad y seguridad.

Accuracy patterns:

Occasional ser/estar/hay confusion, resolved after modelling.

Sporadic article omission (less frequent than A1).

Agreement errors mostly in early turns, corrected after feedback.

In-task development:

Increase in syntactic complexity during exchanges (from single-clause to multi-clause).

Adoption of new vocabulary and connectors mid-task.

Greater control of agreement and complex connectors over time. Before: En ciudad es mejor porque hay trabajo. → After: En la ciudad es mejor vivir porque hay más oportunidades de trabajo y estudio, y además el transporte público es más rápido.

In summary, the A1 group produced functional, topic-relevant output but relied heavily on memorized frames, simple coordination, and high-frequency vocabulary, with persistent morphosyntactic errors. Development during the task was visible mainly in accuracy and slightly longer turns. The A2 group showed greater syntactic flexibility, richer and more precise vocabulary, and more coherent discourse organization. Development was evident in form and content, with learners integrating advanced connectors, concessives, and topic-specific vocabulary into increasingly complex, multi-clause arguments. A comparative overview is presented in Table 4.

OPEN IN VIEWER

Table 4.

Comparative overview: A1 vs. A2 linguistic complexity.

Feature	A1 group (fiestas españolas)	A2 group (pueblo o ciudad)
Syntactic variety	Simple declaratives with basic coordination; rare concessives; reliance on fixed opinion frames.	Coordination + subordination, including concessives, contrastive markers, and relative clauses; multi-part arguments.
Verb use	Present indicative; occasional gustar + infinitive; hay. Few tense variations.	Present indicative with some future/conditional in hypotheticals; better control of hay/estar/ser.
Lexical range	Topic vocabulary and target adjectives; limited synonyms.	Broader topic-specific lexis; varied evaluatives, intensifiers, exemplifiers, and synonyms.
Discourse organization	Short turns, one idea per sentence; limited counterargument use.	Longer turns with structured arguments, counterarguments, and conclusions.
Pragmatic markers	Opinion openers, minimal hedging; few contrastive signals.	Variety of markers for addition, contrast, concession, exemplification.
Accuracy patterns	Frequent agreement and article errors; L1 transfer in adverbs/pronouns.	Fewer agreement errors; occasional ser/estar/hay confusion; less frequent article errors.
In-task development	Gains in accuracy and turn length after feedback; minor structural expansion.	Gains in both form and discourse structure; adoption of advanced connectors and topic-specific vocabulary.

1 Qualitative analysis of student feedback

Building on the observable interaction patterns reported in Section V.1, this section examines learners’ post-task written reflections to explore how they perceived and managed their learning processes. These qualitative data reveal dimensions of strategic thinking, affective response, and critical engagement with AI that were not always visible in real-time transcripts. The thematic and grounded analyses link directly to the three research questions:

a Thematic analysis

To provide an overview of recurring patterns, Table 5 summarizes five themes emerging from student reflections, illustrating how each theme connects to the relevant research question and theoretical constructs. As shown in Table 5, learners demonstrated a combination of planning, monitoring, and evaluative behaviours that reflect increasing levels of autonomy and critical engagement.

OPEN IN VIEWER

Table 5.

Thematic analysis results.

Theme	Research question link	Representative evidence (student quotes)	Interpretation /conceptual link
1. Planning and strategy selection	Research question 2: Learner autonomy	‘I had as much time as I needed . . . I could go through the text in detail.’ ‘I like making written notes . . . I filled about two-thirds of an A4 with phrases and vocabulary.’ ‘I use DeepL . . . switch the languages to get the correct form.’	Strategic planning and personalized resource selection; aligns with Holec’s learner autonomy and Zimmerman’s planning phase of self-regulated learning.
2. Monitoring, reflection, and evaluation	Research question 2: Learner autonomy	‘I asked it why it used that verb form, and it replied it was a mistake.’ ‘Sometimes I wasn’t sure if ChatGPT’s corrections were actually accurate.’ ‘I think I benefited from trying to form sentences in a more complex way and checking words I didn’t know.’	Reflective monitoring and evaluation of learning processes; evidence of metacognitive autonomy.
3. Tool use and autonomous interaction	Research question 1, research question 2	‘It was useful that I could ask for a translation directly in the chat.’ ‘It forced me to think in Spanish again instead of defaulting to English.’ ‘It helped me create sentences, which I think is the most important part of language.’	Independent and intentional use of ChatGPT for learning management; digital learner autonomy.
4. Critical evaluation of AI	Research question 3: Critical engagement	‘Sometimes it just repeated arguments from the original text instead of coming up with new ones.’ ‘I had to instruct it how to argue, which was annoying.’ ‘It corrected words that were actually fine, just not copied from the original text.’	Awareness of AI limitations; aligns with critical digital literacy frameworks.
5. Technical navigation and tool comparison	Research question 3: Critical engagement	‘Phind . . . lets you set a profile, use search mode, and shows where it found information.’ ‘I use DeepL for rewriting sentences, more of a habit, but it works.’ ‘It’s not bad, but in school, face-to-face is better.’	Evidence of AI literacy and multimodal tool integration; critical evaluation of technological affordances.

b Grounded analysis: Open and axial coding

The grounded analysis complemented the thematic coding by identifying specific learner behaviours and grouping them into broader conceptual categories. These categories (Table 6) reflected dimensions of autonomy, critical thinking, and digital literacy, grounded in the theoretical framework. This analysis highlights how learners combined reflective thinking, technical skills, and strategic behaviours to enhance their interaction with ChatGPT. Although not always visible in real-time transcripts, these behaviours indicate a developing autonomy and critical engagement with AI that is consistent with the CEFR and digital competence frameworks.

OPEN IN VIEWER

Table 6.

Grounded analysis results.

Open code	Axial category	Theoretical connection
Asked ChatGPT to explain grammar choices	Metacognitive monitoring	Self-regulated learning (Zimmerman); reflection on input	Research question 2
Took notes during ChatGPT interaction	Strategy use and personalization	Learner autonomy (Holec); strategic note-taking as planning	Research question 2
Attempted more complex sentence structures	Productive risk-taking	Self-regulated strategy; active cognitive engagement	Research question 2
Critiqued ChatGPT’s repetition	Evaluation of AI output	Critical digital literacy (Pangrazio & Selwyn); AI limitations awareness	Research question 3
Compared ChatGPT to other tools	Tool selection and triangulation	AI literacy (Luckin et al.); multimodal critical use	Research question 3
Controlled interaction pace	Temporal autonomy	Learner regulation of task engagement (Little; Zimmerman)	Research question 2
Questioned ChatGPT corrections	Critical stance	Metacognitive reflection and digital agency	Research question 3
Combined tools strategically	Multimodal integration	Technical fluency and strategic learner autonomy	Research question 2, Research question 3
Used ChatGPT for contextual translation	Functional mediation strategy	Language problem-solving and initiative	Research question 1
Expressed positive emotion about task	Affective engagement	Motivation and sustained autonomy (Little)	Research question 2

2 Divergence between observed behaviour and self-perceived autonomy

The transcript analysis captures what students did in the moment, showing autonomy and critical thinking as enacted behaviours. The feedback analysis reveals what students thought, planned, and felt, including strategies and evaluations that were invisible in real-time interactions. Combining the two provides a fuller picture. A comparison of observed and reported behaviours is presented in Table 7.

OPEN IN VIEWER

Table 7.

Comparison of observed vs. reported behaviours.

Dimension	Observed in conversations	Reported in feedback
Initiative	Frequent in both groups, seen in prompt adaptation and topic extension.	Often linked to purposeful use of AI for targeted learning aims.
Self-regulation	More common in A1; visible in vocabulary requests, clarification-seeking.	Expanded to include strategic planning, note-taking, productive risk-taking, and pacing control.
Strategic behaviour	Relatively infrequent in transcripts, mostly in A1.	Stronger presence in reflections, with detailed strategies and tool integration.
Questioning AI output	Mostly focused on language form; limited factual checking.	Included awareness of AI’s argumentative limits, unnecessary corrections, and repetition.
Identifying errors	More common in A2; mostly technical corrections, rare content checks.	Broader critical stance including evaluation of logic and alternative tool use.
Digital literacy	Seen only indirectly in transcripts (e.g. directing AI format/style).	Strongly present: students explicitly compared tools and integrated multiple resources.
Affective engagement	Not observable. Although certain student responses exhibit indications of frustration or irony. For instance, following an unsatisfactory reply from ChatGPT to a prompt, one student commented: Well, you really did a great job there.	Frequently reported: motivation, satisfaction, and increased willingness to engage in Spanish.

A central finding of this study is the gap between students’ observed behaviour during ChatGPT interactions and their self-reported perceptions of autonomy and critical engagement. While transcript data revealed moderate autonomy, with student-initiated interactions and requests for clarification, students also relied heavily on ChatGPT for linguistic decisions. Few questioned or challenged AI outputs.

By contrast, student reflections conveyed more advanced strategic thinking. Learners described goal-setting, tool comparison, metacognitive insight, and awareness of AI limitations. These reflective accounts suggest developing digital agency and self-regulated learning beyond what is evident in real-time behaviour.

This discrepancy suggests that performance-based autonomy (visible in task execution) may differ from metacognitive autonomy (revealed in reflection). ChatGPT’s dialogic and semi-agentive interaction may scaffold learners so effectively that outward signs of autonomy are masked, even as cognitive engagement occurs internally. Moreover, task design itself, particularly when scaffolded, can shape how autonomy is expressed externally.

These findings suggest that autonomy is not always directly observable during interactions, especially in AI-mediated settings where the tool’s responsiveness can obscure learners’ internal strategies. Post-task reflections are essential for capturing the full scope of learner development.

1 Discussion

This study investigated how beginner-level multilingual learners engaged with ChatGPT to perform CEFR-based mediation tasks and develop learner autonomy. Regarding research question 1, the findings indicate that learners at both A1 and A2 proficiency levels participated in meaningful communicative activities, including summarizing texts, clarifying vocabulary, and negotiating meaning. These behaviours align with the CEFR Companion Volume, particularly within the categories of Mediating Texts and Facilitating Communication, and suggest that mediation skills can be developed even at lower proficiency levels when learners receive appropriate digital and pedagogical support.

More complex strategies associated with Mediating Concepts, such as paraphrasing and synthesizing information, were less frequently observed. Students predominantly used prompts to clarify or compare. This tendency may be attributed both to linguistic limitations and to the nature of ChatGPT’s output. Because the tool produces fluent and grammatically accurate responses, learners may be less inclined to reformulate ideas in their own words. This finding is particularly relevant to research question 1, as it underscores the importance of designing tasks that explicitly require learners to adapt or explain content rather than passively accepting AI-generated formulations.

With respect to research question 2, a central finding is the contrast between learner behaviours during interactions and the strategic thinking revealed in post-task reflections. While students initiated prompts and responded to feedback, they often relied on ChatGPT to validate their choices or resolve uncertainties. However, their written reflections revealed deeper engagement, including goal-setting, tool comparison, and critical evaluation of AI-generated responses. This supports the conclusion that metacognitive autonomy is expressed more clearly in reflective writing than during real-time interaction.

These observations imply that outward signs of autonomy may be less visible in AI-mediated settings, particularly when the tool responds efficiently to learner input. ChatGPT’s supportive output can obscure the cognitive effort involved in decision-making, making post-task reflection essential for capturing the full range of learner strategies and mental processes.

Another notable finding is that some A1 learners exhibited greater interactional autonomy than their A2 counterparts. This observation is grounded in the higher percentage of A1 participants demonstrating initiative (83% vs. 75%) and self-regulation (42% vs. 17%) during the tasks, rather than a subjective judgment of overall quality. While higher proficiency generally provides greater linguistic flexibility, these results suggest that autonomy may be influenced more by learner disposition than by language level alone. Factors such as motivation, self-confidence, and willingness to experiment appear to play a significant role. Some lower-proficiency learners took more linguistic risks and persisted in exploring meaning, whereas some more advanced learners occasionally simplified their output to avoid errors.

In relation to research question 3, the study identified early signs of critical digital and AI literacy. Although learners rarely challenged ChatGPT during the tasks, many demonstrated reflective thinking in their feedback. Several participants noted inaccuracies or repetitive reasoning in the AI’s responses, while others compared ChatGPT’s output with that of alternative tools. These behaviours suggest that while critical engagement may be minimal in real time, it emerges more clearly in post-task reflection, supporting an expanded definition of critical digital literacy. Furthermore, when students recognized factual errors in ChatGPT’s output, they often lost trust in the tool as a reliable source of information and saw little value in pursuing further clarification during the task.

The data also indirectly address research question 3 by raising questions about the emotional and psychological dimensions of AI-mediated learning. For some learners, interacting with a non-human partner reduced performance anxiety and encouraged experimentation. For others, however, the absence of human feedback appeared to limit motivation. Further research is needed to investigate how affective factors influence learner behaviour in AI-supported contexts.

Finally, although the CEFR mediation descriptors provided a useful analytical framework, they were not designed with AI interaction in mind. This has implications for both research question 1 and research question 3. Their application to AI-mediated tasks requires adaptation to account for emerging dynamics such as dialogic co-construction, intrapersonal meaning-making, and digital agency. These issues are examined in greater detail in the following section.

2 Rethinking CEFR mediation and plurilingualism in AI-enhanced contexts

While the CEFR offers a solid foundation for describing mediation competence, its descriptors were developed for interpersonal communication and do not fully account for the dynamics of learner–AI interaction. In this study, learners used ChatGPT to develop mediation strategies aligned with CEFR categories, including Mediating Texts, Mediating Concepts, and Facilitating Communication. However, AI tools such as ChatGPT pose theoretical and pedagogical challenges that warrant closer examination.

c Linear dialogue and artificial conversational dynamics

AI conversations are linear and explicit, lacking the inferential flow of human dialogue. To co-construct meaning, human speakers draw on shared knowledge, incomplete utterances, and emotional states. ChatGPT, in contrast, requires full verbalization of assumptions, repetition of context, and high degrees of precision to maintain coherence.

This changes the communicative conditions under which mediation occurs. Learners must take increased responsibility for structuring discourse and managing turn-taking. These demands may develop helpful metacognitive strategies, but they also highlight the artificiality of interaction. Unlike a human interlocutor, ChatGPT does not evolve during the conversation or incorporate affective context. To account for these differences, CEFR descriptors should be contextualized and expanded for AI-mediated settings. Key areas for adaptation are presented in Table 8.

OPEN IN VIEWER

Table 8.

Adapting CEFR mediation descriptors for AI contexts.

Standard CEFR element	Proposed AI contextualization
Mediating a text	Retain with scaffolding; AI supports summarizing, rephrasing, and simplifying with learner prompts.
Mediating concepts	Include intrapersonal meaning construction via AI before interpersonal use.
Facilitating communication in plurilingual settings	Differentiate between simulated cultural explanation with AI and experiential negotiation with people.
Managing interaction	Emphasize explicit turn-taking and clarification strategies specific to AI interaction.
Pluricultural awareness	Encourage learners to critically reflect on AI’s cultural neutrality, gaps, or inaccuracies.

In addition, CEFR-based mediation could recognize AI use as a transitional space between reception and production. When learners ask ChatGPT to clarify unknown words or rephrase texts, they engage in intrapersonal mediation – constructing understanding with digital support. These processes can prepare learners for interpersonal mediation, where they apply that understanding in social or academic contexts.

In summary, CEFR remains highly relevant in AI-supported learning, but must be rethought in terms of the nonlinear, text-dependent, and cognitively distinct features of AI conversation. Recognizing the difference between intrapersonal and interpersonal mediation, and designing tasks that foster both, will be essential to future pedagogy.