Abstract
This study examines how educators scaffold students with disabilities to interact meaningfully with generative artificial intelligence (GenAI) tools, specifically ChatGPT, in an inclusive science classroom. While GenAI is gaining traction in education, research on its use among students and educators in inclusive education remains underexplored. Addressing this gap is critical as effective GenAI use among students with disabilities demands cognitive and linguistic engagement that are challenging for them without appropriate scaffolding. This study adopted a collaborative action research approach to explore how educators supported students with disabilities in engaging with ChatGPT in the classroom. Drawing on video-based analysis of classroom interactions, the analysis resulted in an emergent taxonomy of educator support across four domains: instructional, literacy, emotional-behavioural, and technological. The findings illustrate the utility of the taxonomy that frames how established domains of support for students with disabilities are enacted, combined, and intensified in the context of student-GenAI interaction. They also highlight how real-time and adaptive scaffolding enabled students with disabilities to access, interpret, and respond to GenAI. This study reframes GenAI not merely as a tool, but as a mediated dialogic partner whose educational value depends on how it is supported by educators. Implications are discussed for inclusive pedagogy, GenAI design, and teacher professional learning.
Introduction
The emergence of generative artificial intelligence (GenAI) has sparked widespread usage in classrooms around the world. With their capacity to generate coherent and contextually relevant content in response to user prompts, GenAI tools such as ChatGPT are increasingly used to support various literacy tasks, such as writing, explanation, translation, and summarization (Angelone, 2025; Tang, 2024). These tools hold promise for learners who require individualized support, especially for students with disabilities where GenAI can function as a personalized learning assistant that adapts to their pacing, preferences, and communication styles (Farhah et al., 2025; Vorobyeva et al., 2025; Waterfield et al., 2024). Given this affordance, it is imperative to explore how GenAI can be leveraged to offer students with disabilities equitable opportunities to learn, participate, and succeed in the classroom (Marino et al., 2023). Furthermore, as GenAI technologies become increasingly embedded in everyday life and the workplace, equipping students with disabilities with AI knowledge and skills are essential beyond academic success to ensure their participation in a future AI-mediated society (Mitre & Zeneli, 2024).
Despite these opportunities, research on the use of GenAI in inclusive education remains scarce. Existing studies on GenAI have largely focused on mainstream learners and higher education contexts (Almasri, 2024; Tang, 2025; Zhai & Krajcik, 2024), with little attention to the unique affordances and challenges of GenAI for students who require additional cognitive, communication, or behavioural support. At the same time, research emphasizing special education has examined AI in general (e.g., Panjwani-Charania & Zhai, 2024; Rice & Dunn, 2023), but not GenAI as a specific type of AI tool. Addressing this gap is significant as effective use of GenAI demands a level of cognitive and linguistic engagement, which learners with disabilities may find particularly challenging without appropriate scaffolding. Understanding how educators can support and mediate these interactions is therefore vital. Without such support, students with disabilities risk being further marginalized in digitally mediated learning environments, thus reinforcing existing educational disparities rather than addressing them.
This study addresses this critical gap by examining how a group of educators provide scaffolding for students with disabilities to interact with GenAI in an inclusive science classroom in Australia. Drawing on video observations of classroom interactions involving the use of ChatGPT, the study identifies patterns of educator support that are organized into a taxonomy illustrating the types of support and the ways students respond to them in a classroom environment. The research questions that guided the study were: (1) What types of support do educators provide to enable students with disabilities to interact with ChatGPT in the science classroom? (2) How do students with disabilities access, interact with, and respond to ChatGPT during science lessons in relation to the support provided by educators?
Theoretical Framing & Literature Review
AI in Inclusive Education
In the Australian educational context, the term disability refers to a broad range of physical, cognitive, sensory, emotional, and behavioural conditions that impact a student’s ability to access and participate fully in learning without additional support. In 2022, the Australian Bureau of Statistics (ABS, 2022) reported that 13.5% of school-aged children from age 5–14 in Australia have some forms of disability, with many requiring additional supports in the classroom. Importantly, disability is not limited to visible disabilities and includes a range of conditions such as autism spectrum disorder (ASD), attention deficit hyperactivity disorder (ADHD), intellectual disability, and specific learning difficulties such as dyslexia, dysgraphia, and dyscalculia. According to the Australian Government’s (2005) Disability Standards for Education, educational institutions are required to make reasonable adjustments to promote inclusive and equitable learning outcomes to all students, including those with disabilities (Australian Government, 2005).
Students with disabilities often experience persistent challenges in basic academic skills, including reading, writing, and mathematics, as well as higher-order skills such as reasoning, problem-solving, and metacognition. These difficulties are not attributable to external factors such as socioeconomic disadvantage or language background but are instead intrinsic to the student’s cognitive and neurological profile (Büttner & Hasselhorn, 2011). In many cases, students with ASD or ADHD may also present with overlapping learning difficulties, requiring multifaceted and individualized support strategies. These challenges can affect students’ social and emotional development, contributing to lower self-esteem, anxiety, and difficulties with peer relationships. Therefore, effective support for students with disabilities involves intensive academic scaffolding with particular attention to their cognitive and communication needs. Such support must be targeted, flexible, and individualized so that students can access the curriculum while developing confidence and autonomy in their learning.
Inclusive education in Australia is guided by the principle that all students, regardless of ability or disability, have the right to learn together in mainstream classrooms (Ministerial Council on Education et al., 2008). In practice, inclusion requires both structural adjustments—such as curriculum differentiation and physical access—as well as pedagogical strategies that cater to neurodiverse learners. Educators are expected to adopt differentiated instruction, collaborative teaching, and, increasingly, technology-enhanced learning tools to personalize support for them. With the rise of GenAI tools like ChatGPT, there is growing potential to provide just-in-time assistance and adaptive feedback that cater to various disability profiles. However, realising this potential depends heavily on how such tools are integrated and scaffolded by educators within inclusive classroom environments (Rappa & Nonis, 2026).
While the use of AI to support students with disabilities is well-established over the last decade (Marino et al., 2023; Panjwani-Charania & Zhai, 2024), research focusing on generative AI remains limited (Rappa et al., 2026). Existing studies have primarily examined non-generative AI applications such as communication assistants, adaptive learning platforms, facial expression recognition, diagnostic tools, and intelligent tutoring systems—technologies designed for specific tasks rather than generating original content (e.g., Hopcan et al., 2023; Rice & Dunn, 2023). In contrast, the use of GenAI tools like ChatGPT to produce novel and context-sensitive responses may offer new possibilities for supporting learners with disabilities in more open-ended and general literacy-based tasks. However, this flexibility also introduces unique challenges, especially for students who struggle with language processing or prompt formulation. Investigating how educators scaffold these students’ interactions with GenAI is therefore crucial to ensuring that these emerging technologies advance, rather than hinder, inclusive education.
Scaffolding & Support for Students With Disabilities
Vygotsky’s (1986) sociocultural theory offers a useful framework for understanding how students with disabilities can be supported in inclusive learning environments. Rather than viewing cognitive impairments as fixed traits, Vygotsky emphasized the socially mediated nature of learning and development. He proposed that all learners, including those with neurodevelopmental differences, can develop higher mental functions beyond their current capabilities, through instructions and interactions with more capable humans and tools (Daniels, 2020). These abilities emerge first in social interaction (inter-psychological plane) before being internalized by the learner as personal competencies (intra-psychological plane). From this perspective, a learner’s potential is not determined solely by neurological constraints but is shaped through appropriate social, cultural, and technological support.
Central to this theoretical framework is Vygotsky’s concept of the zone of proximal development (ZPD), defined as the distance between a learner’s current independent functioning and their potential performance when guided by a knowledgeable other. The ZPD foregrounds the role of scaffolding, which we define in this study as the intentional, adaptive, and observable support educators provide to help students, including those with disabilities, accomplish tasks they could not complete on their own (Taber, 2025; Wood et al., 1976). It is important that scaffolding is not static or prescriptive for students with disabilities; rather, it must be adapted in real time to meet each neurodiverse student’s changing needs (Reid, 1998). In practice, scaffolding for students with disabilities may include rephrasing a task using simplified or multimodal language when a student experiences processing difficulty, breaking instructions into shorter steps to support attention or working memory, or providing brief physical assistance (e.g., guiding typing or navigation). Such practices are distinguishable from one-off prompts or instructions because they are responsive, temporary, and oriented toward supporting students’ progression within their ZPD (Puntambekar, 2022). By identifying and supporting the ZPD of students with disabilities, educators can target emerging capabilities rather than simply accommodating existing limitations. This approach reframes cognitive diversity not as a deficit, but as a condition that requires adaptive, collaborative, and socially responsive teaching practices (Puntambekar, 2022). In doing so, Vygotsky’s ZPD advocates for an inclusive pedagogy that centres on potential, growth, and participation in meaningful learning. Examples of such adaptive scaffolding in the context of GenAI are provided in the findings.
Existing research (prior to the recent release of ChatGPT-3 in 2022) has identified multiple domains of support for students with disabilities. Instructional support often involves simplifying or explaining task instructions, segmenting and sequencing activities, and skill modelling strategies (Swanson & Deshler, 2003). Literacy support includes assisting with prompt construction, supporting reading and writing through visual cues, and using augmentative and alternative communication (AAC) (Sturm & Clendon, 2004). Emotional and behavioural support are equally vital and include the use of consistent classroom routines, positive reinforcement, co-regulation strategies, and affective cues to build motivation and emotional security (Rebanal et al., 2025). Finally, technological support encompasses the use of assistive tools such as speech-to-text applications and visual schedules to enhance access and engagement with learning tasks (Chandra et al., 2025; Howorth et al., 2025). This study aims to contribute new insight to this body of research by expanding the established forms of support into the context of using GenAI tools in real classroom settings.
GenAI as Dialogic Partner
In addition to the theoretical influence from Vygotsky, this study is also shaped by our previous work in conceptualizing GenAI as a dialogic partner in education (Tang, 2024; Tang & Putra, 2025). This theorization builds on Bakhtin’s (1981) concept of heteroglossia that reflects the coexistence of multiple voices or discourses within a single text. As every text is always a response to previous texts and anticipates future responses, Bakhtin argues that no text is ever completely original as its meaning always emerges from the dialogic interaction between different voices within the text. Kristeva (1980, p. 37) further expanded on Bakhtin’s idea by proposing that “any text is constructed as a mosaic of quotations” appropriated from other authors. She coined the term intertextuality to denote that the meaning of a text is derived through its relationship with other texts, both explicitly (e.g., quotation) and implicitly (e.g., shared genres or perspectives).
GenAI tools such as ChatGPT generate text by statistically predicting word sequences based on vast amounts of human-authored training data and the user’s inputs. Drawing on the metaphor of intertextuality, we can understand that GenAI does not possess an autonomous voice; instead it produces texts that remix and recontextualize a multiplicity of human voices. Tang (2025) introduces the term AI-textuality to describe this intertextual and socially situated nature of texts generated through human–GenAI interaction. When students interact with GenAI, they are not simply retrieving static information from a machine, but they are co-constructing meaning through a dynamic process in which both the user’s input and the AI’s probabilistic response contribute to the ongoing discourse.
This dialogic framing is particularly pertinent for students with disabilities who require additional support in navigating language-rich interactions with GenAI, more so than just understanding the content generated by GenAI. It is also relevant for science education, where scientific understanding develops through engagement with competing explanations and evidence-based reasoning (Tang, 2024). From this theoretical framing, this study emphasizes GenAI’s role in stimulating engagement and dialogic inquiry with the users. In an inclusive classroom, it is critical to position GenAI as a potential dialogic partner that can respond to student interests, encourage follow-up questioning, and maintain conversational continuity in a neutral and non-evaluative manner. These dialogic functions will often require educator scaffolding in inclusive classrooms. When mediated effectively by educators, such AI interactions can create dialogic spaces that support students with disabilities to articulate ideas and construct meaning over time. Dialogic interactions also invite these students into scientific discourse not as passive recipients of AI-generated knowledge, but as active participants in personal meaning-making. Investigating how educators mediate this interaction is key to understanding how inclusive pedagogies can harness GenAI’s dialogic potential to support learning in diverse classrooms.
Methods
Research Context and Participants
This study utilized an action research approach (Ferrance, 2000) to explore the use of GenAI in inclusive classroom settings. Collaborative action research is characterized by iterative cycles of planning, acting, observing, and reflecting, with teachers and educators actively engaged as co-researchers (McAteer, 2013). This participatory methodology allows for both contextual relevance and responsiveness to the unique needs of students with disabilities. In this action research, the teachers and researchers jointly designed the learning activities, implemented GenAI use in the classroom, and jointly reflected on emerging practices and student responses (Tang et al., 2024). This iterative process ensured that the intervention evolved in response to real-time classroom dynamics, while also supporting teacher learning and ownership of the intervention. The action research followed ethical guidelines stated in the Australian Code for the Responsible Conduct of Research and was approved by the Human Research Ethics Committee of Curtin University (Project number: HREC2023-0413).
Situated in a secondary school classroom in Australia, two special education teachers (T1 and T2) worked together with the researchers to co-design and investigate pedagogical interventions involving the use of OpenAI’s (2023) ChatGPT in learning science. Besides these two teachers, there were also four education assistants (EA1, EA2, EA3, EA4) who provided individualized support to some of the students. The ongoing collaboration not only ensured the generation of knowledge in a real-world context, but also supported localized professional learning and inclusive and innovative teaching in practice. 14 Australian students were enrolled in the class, with six students and their parents giving consent to participate in the research. These students required varying levels of academic and social assistance to navigate their learning experiences effectively. The diverse profiles of these participating students are as follows: • S1: 13-year-old male student diagnosed with ASD, mild intellectual disability (ID), and impaired hearing. Despite these challenges, he demonstrated strong technological skills and engaged confidently with digital tools during lessons. • S2: 11-year-old male student transitioning from primary to secondary school, diagnosed with Fragile X syndrome. Required additional behavioural observations from a therapist to support his integration into the new learning environment. • S3: 14-year-old female student diagnosed with ADHD, learning disability, and anxiety. Although she had low literacy skills, she showed confidence and enthusiasm when using technology. • S4: 15-year-old male student diagnosed with ASD and complex mental health challenges. He received individualized support from an Education Assistant (EA) throughout the sessions to manage emotional regulation and task engagement. • S5: 13-year-old male student who is an older sibling of S2, diagnosed with Fragile X syndrome, ID, ASD, and anxiety. Required consistent one-on-one support from an EA to participate meaningfully in class activities. • S6: 12-year-old female student diagnosed with mild ID. She demonstrated strong technological aptitude and was able to engage with digital simulations and interactive science tasks with minimal assistance.
The learning activities covered six science lessons, each lasting 1 hour and 30 minutes, for students to explore various physics concepts such as forces (push & pull), motion, inertia, gravity, and friction. All the six lessons followed a similar structure. The first half of a lesson usually started with a whole-class discussion to introduce the key physics concepts, followed by a hands-on demonstration to reinforce their understanding. Students then worked in groups to carry out an experiment or investigation themselves. In the second half of the lesson, the class transitioned to individual tasks involving the use of ChatGPT 3.5 to further explore the science concepts based on what the students had experienced from the previous hands-on activities.
All the lessons occurred in a science laboratory and the seating arrangements accommodated for diverse and specific student needs. S1 sat at the front due to hearing difficulties, with S2 beside him as part of his transition to the class. S4 sat next to S2, supported closely by an EA. Behind S1, S6 worked independently, occasionally seeking assistance from the teacher or EA, while S3 preferred sitting alone due to sensitivity to noise. S5 alternated between working with S4 and his EA or focusing solely with EA support. These six students were also seated in a way that facilitated the video recording and avoid capturing other students who did not provide consent to participate in the research. Many students brought their personal laptops, while others used devices provided by the school. Thus, all students had individualized access to ChatGPT to support their inquiry.
Prior to the study, ChatGPT and other GenAI tools had not been used in the classroom. Students also had limited direct experience with GenAI, although many were generally aware of artificial intelligence and its prominence in public discourse in 2023. The introduction of ChatGPT was therefore a novel element of the learning environment and was initiated by the research team in consultation with the teachers. Apart from the introduction of ChatGPT, the classroom routines, instructional practices, and technological supports remained consistent with the regular learning environment described above. Students were also accustomed to using one-to-one personal digital devices to support the learning activities facilitated by the teachers and EAs. This design allowed the study to examine how educators scaffolded students’ initial interactions with GenAI within an otherwise familiar and stable learning environment.
Data Sources and Analysis
This study employed ethnographic and video-based methods (Garcez, 2017) to capture and analyze the complexity of scaffolded interactions involving GenAI in an inclusive science classroom. The primary data source consisted of observation notes and video recordings of six core lessons in which students with disabilities interacted with ChatGPT. Three researchers (R1, R2, and R3) were present as participant-observers during the lessons. Two high-definition video cameras were used to record the classroom interactions. One camera was mounted on a tripod with a wide-angle setting to capture the teachers’ whole-class discussion. The second camera was handheld by R3 to capture student and educator interaction as well as their interaction with ChatGPT. Secondary data sources included student chat logs with ChatGPT and semi-structured interviews with students and educators. These additional data were used to triangulate and support the interpretations drawn from the video analysis.
To address the research questions in this study, the recorded classroom videos were systematically viewed with micro-interaction as the primary unit of analysis (Garcez, 2017). Following Erickson’s (1992) method of ethnographic microanalysis, the video recordings were segmented into episodes—bounded sequences demarcated by changes in interactional focus, task structure, or participant roles. Each episode was coded for its dominant type of classroom activity (e.g., whole-class discussion, group work, lecture, demonstration, individual work) and whether GenAI was used. An event log (Jordan & Henderson, 1995) was then created to describe the key events, activities, and video timestamps of each episode, providing a contextual overview to support further close analysis. There was a total of 260 episodes over six lessons.
Episodes involving the use of ChatGPT were then examined in detail to identify scaffolding practices used by the teachers, EAs, and researchers. The initial coding was guided by a literature-derived framework (as described earlier) outlining four major categories of support for students with disabilities: instructional, literacy, emotional-behavioural, and technological (e.g., Howorth et al., 2025; Sturm & Clendon, 2004). However, the initial coding scheme did not fully capture the nuances of how educators scaffolded students’ use of GenAI. A subsequent phase of inductive coding was then conducted to identify, based on close analysis of the video data, the specific types of support related to GenAI use within each scaffolding category.
Example of an Event Log With Associated Codes.
Key:
T1, T2 – Teachers.
EA1 to EA4 – Education assistants.
R1 to R3 – Researchers.
S1 to S6 – Participating students (with consent).
Researcher Positioning and Reflexivity
The research team comprised four university-based researchers with professional backgrounds in education and experiences as former secondary school teachers. Three members are teaching-research academics with expertise spanning science education, educational technology, and special education, and one member is a research assistant and doctoral student. One researcher had an established collaborative relationship with the participating school, which facilitated the selection of the site and access to the classroom for this study. As mentioned, three researchers were present in the classroom as participant-observers, and they were perceived by students as co-teachers to support their learning activities and use of GenAI. The close and prolonged engagement in the classroom enhanced contextual understanding and credibility consistent with an action research approach. At the same time, the presence of researchers as participant-observers and co-scaffolders may have shaped the nature of classroom interactions observed. The scaffolding practices documented in this study should therefore be understood as co-constructed within this specific research context. To support principled adaptations of this study’s findings, we propose the subsequent taxonomy as an analytic framework with the goal to inform, rather than prescribe, educator practice in other inclusive settings.
Findings
Taxonomy of Support for Students With Disabilities to Interact With ChatGPT
Taxonomy of Support for Students With Disabilities to Interact With ChatGPT.
Instructional Support Type
Instructional type of support comprises scaffolding practices that guide students on how and when to use GenAI for learning, and why. This category encompasses the instructional and cognitive aspects of helping students navigate the purpose and process of engaging with GenAI during the lessons.
Explaining GenAI
The educators introduced ChatGPT explicitly by demonstrating how it could be used as a learning tool. For instance, T1 often modelled to the whole class the use of ChatGPT by showing how to pose questions and exploring different ways of interacting with the interface of the chatbot. This scaffolding practice also includes instances when the educators (including researchers and EAs) provided individual explanation on the general use of ChatGPT, which enhanced students' familiarity with the tool and gaining confidence to use it independently.
Framing Purpose
Framing purpose involved orienting students to the broader learning goal and the rationale for using ChatGPT in science learning. While the previous support (explaining GenAI) focused on how to use ChatGPT, this scaffolding practice emphasized why they could use ChatGPT. The educators explained the purpose of using ChatGPT within a clear instructional context and how it related to the students’ inquiry. This form of support was more prominent in the first two lessons when ChatGPT was relatively new to the students.
Guiding Questioning
Educators supported the students in generating prompts to ChatGPT by guiding their question formulation, including helping them identify appropriate topics to inquire and translating their thoughts and interests into specific questions. For example, when S2 wanted to ask a question about sports, EA3 suggested he could ask about his favourite sports. This scaffolding practice was often provided in the moment as students initiated the conversation with ChatGPT with the educators observing and providing support.
Recalling and Connecting
The educators supported conceptual continuity by directing students to connect their ChatGPT interactions with prior hands-on activities or discussions. For instance, when T2 asked the class whether ChatGPT could help explain the outcome of a science experiment, it encouraged the students to see GenAI not just as a separate activity, but as an integrated extension of the entire lesson. This scaffolding practice supported students in activating prior knowledge and integrating their learning from the previous embodied experience to the present dialogic inquiry with GenAI.
Literacy Support Type
Literacy support includes scaffolding practices that assist students in reading, writing, or interpreting the language required to use GenAI effectively. While instructional support involves the cognitive aspect of engaging with the AI, this support emphasizes the use of language to interact with GenAI. Given the language-rich nature of GenAI interaction, these supports were critical for students with language or communication difficulties.
Assisting With Prompt
Many students with disabilities in the study required close support to formulate a prompt for ChatGPT. The scaffolding practice in this area was realized through verbal rehearsal, dictation, or modelling sentence structure. In one instance, EA3 dictated a question to S2 to type, while in another case, EA1 helped the student to repeat each word in the question as it was read aloud. Such scaffolding helped to reduce the linguistic barriers and ensure those students could engage with ChatGPT in the interaction, regardless of their language difficulties.
Assisting With Reading
As ChatGPT’s response was provided in written texts, some students required additional support to decode the generated text. The educators supported their comprehension by reading aloud ChatGPT’s responses, occasionally pausing to explain complex words or provide a commentary. This scaffolding enabled those students with reading difficulties could still engage with ChatGPT’s written output.
Assisting With Spelling
The educators were observed to often assist students with spelling when typing their queries into ChatGPT. This included sounding out words, spelling aloud, or writing the key vocabulary or sentence on the whiteboard. For example, when S6 struggled to spell the word “break,” EA1 provided the correct spelling, enabling the student to proceed with her question. Such scaffolding ensured that technical or unfamiliar vocabulary did not become a barrier to accessing ChatGPT.
Emotional-Behavioural Support Type
This category encompasses scaffolding practices that nurture motivation, emotional security, positive behaviour, and confidence in using ChatGPT. These factors influence students’ ability to persist and participate in learning activities involving new technologies.
Encouraging and Building Confidence
Educators frequently offered verbal encouragement and positive reinforcement to support students’ self-efficacy with ChatGPT. In one instance, EA1 encouraged S4 to continue interacting with ChatGPT, providing emotional affirmation and typing on the student’s behalf when needed. Such emotional support was critical for many students in this study who often hesitated in unfamiliar learning situations.
Directing Attention
Educators redirected students’ focus when their attention wandered or when they deviated from the learning task. For instance, when R1 noticed that S2 was asking questions about his pet, which were unrelated to the physics task, she asked whether S2 had brought the dog in a car ride. R1’s question was a strategy to gently direct S2 to revisit the lesson topic on motion. These cues were often delivered to help students stay on task without feeling reprimanded and was especially useful for those students with ADHD or anxiety.
Initiating Self-Regulation
In situations where students displayed signs of frustration, fatigue, or overstimulation, educators encouraged them to reflect on their needs and offered choices (e.g., “Do you want a break or to keep going?”). For several students who were observed to have behavioural or emotional regulation challenges, this type of support modelled healthy coping strategies and supported long-term self-management skills.
Technological Support Type
Technological support refers to practical assistance that facilitates students’ physical access and operation of ChatGPT as a learning technology. These scaffolding practices addressed computer and Internet access, navigation skills, and just-in-time use of assistive technologies.
Accessing ChatGPT
As a relatively new web-based tool, many students in the study required assistance in accessing ChatGPT, troubleshooting login issues, and exporting their chat sessions for analysis. Some students, particularly those with executive functioning difficulties, needed repeated support over multiple sessions, highlighting an additional need for sustained access facilitation in inclusive settings.
Assisting With Typing
For some students with motor challenges who found it hard to type independently, educators often typed their questions and prompts on their behalf. This support is different from the “spelling words” support that was categorized as a literacy support. Here, the students may know the spelling, but they had difficulties in physically typing on the keyboard. As such, educators often asked for their input before typing and reading the output together.
Using Assistive Technology
ChatGPT 3.5 includes some features that allow users to use voice input or hear an audio playback of ChatGPT’s response. However, the students were not familiar with these technological affordances. In supporting the students’ interactions, educators often enabled these features or guided students to activate them. For instance, as S6 expressed difficulty in reading written text, R2 assisted her to play an audio response from ChatGPT.
Impact of Support for Students With Disabilities to Engage With ChatGPT
Close video analysis of individual students’ interactions with ChatGPT during the science lessons was carried out to address this research question. In this section, we show the findings from a particular student (S2) to highlight how he accessed and responded to ChatGPT in relation to the scaffolding provided by the educators. This case serves to illustrate how a student’s engagement with ChatGPT was shaped by a combination of educator supports across the four categories as described in the above taxonomy, and is representative of the six students observed in this study.
Episode 1: Initiating Interaction With ChatGPT
This episode took place during the third lesson in the six-week sequence. The lesson focused on developing students' familiarity with ChatGPT following a prior classroom discussion and 
hands-on activity about friction and road surfaces. S2 had expressed interest in asking ChatGPT about cars and sports. The episode began when S2 initiated the question, “Do you have a car?”, and ChatGPT responded that: “I don’t have a car, but I know a lot about them!”. At this point, EA3 joined in to scaffold the following interaction:
When S2 commented to EA3 that ChatGPT didn't have a car, EA3 explained that “ChatGPT is a computer” (line 3). This clarification served not only to provide a gentle correction to S2’s personification of AI, but also to set expectations about what ChatGPT could and could not do. Rather than closing down the interaction, EA3 used this moment to shift the focus of the conversation by suggesting that S2 could ask about “science or sports” (line 5). This strategic move balanced the science lesson focus with S2’s personal interests.
As S2 expressed his interest in sports (line 8), EA3 supported his next attempt to generate a new question. She scaffolded the phrasing with a question starter in line 9, “Do you…”, then extending it by asking S2: “Or what is your most favourite type of sport?” (line 10). When S2 responded with “basketball”, EA3 built on his response by suggesting they could ask ChatGPT on this topic. This exchange shows how EA3 adapted her language support responsively, providing both verbal cues and social encouragement to help S2 formulate a meaningful question.
From line 15 onwards, as S2 articulated his question hesitantly, EA3 assisted him to type his question: “Do you think basketball is a good type of sport?” She repeated and read S2’s question aloud as she typed it, reinforcing comprehension and maintaining S2’s cognitive involvement in the interaction. This moment illustrates the shift from linguistic to motor scaffolding—EA3 recognized that S2’s difficulty lay in execution, not in content generation, and adapted her support accordingly.
This episode illustrates how EA3 provided multiple dimensions of support in guiding S2 to use ChatGPT more effectively. Her language modelling and verbal cues exemplified
Episode 2: Exploring Scientific Concepts With Audio Support
This episode occurred during the fourth lesson focusing on the role of friction and brakes in stopping a moving car. The lesson started with a hands-on demonstration of how road surfaces affect car motion, followed by small group investigations. After these activities, students used ChatGPT to further 
explore what they had observed. This episode captures how S2, supported by T1 and R1, formulated a science-based question and interacted with ChatGPT’s response using assistive audio features:
In the beginning, T1 reminded S2 of the inquiry question she had modelled earlier: “How does friction work in a car?” She showed S2 the question that she typed into ChatGPT on her laptop. When it became apparent that S2 could not see the question clearly while trying to type it out (line 2), T1 wrote the full question on the whiteboard to provide a visual reference (line 3). At the same time, R1, who was observing nearby, reinforced the instructional cues by drawing S2’s attention to the board (line 4) and affirming his input after he had completed typing (line 6). This interaction illustrates how T1 and R1 worked in tandem to support S2’s formulation of a science-based prompt using a combination of verbal, visual, and affective scaffolding practices.
After entering the question into ChatGPT, S2 silently read the AI-generated response (line 7). He then hovered over icons such as “Share” and clicked on “Edit message” and “Cancel,” but did not type or respond further. His body language and some inaudible mumbling (line 9) suggested uncertainty or confusion. Sensing this hesitation, R1 stepped in to offer support. She asked, “Do you want to hear it? Do you want me to help you with the volume?” (line 10), interpreting S2’s actions as a possible search for the audio playback function. S2 agreed, and R1 proceeded to guide him through the steps to activate the read-aloud feature (lines 11–13). Before S2 played the audio, he repeated his question loudly, “How does friction help in everyday life?” (line 14).
This episode demonstrates how S2’s interaction with ChatGPT was supported by various scaffolding that addressed both linguistic formulation and sensory accessibility. T1’s initial support with constructing and writing the prompt exemplified
Episode 3: Redirecting Interaction to Science Conversation
This episode took place later in the same lesson as Episode 2. After formulating and asking a question about friction, S2 continued using ChatGPT independently. However, his attention began to shift toward more personal interests as he asked questions unrelated to physics, as shown in the following exchange taken from his chatlog: S2: Do you have any pets ChatGPT: I don’t have pets, but I think pets are wonderful! Do you have any pets yourself? I’d be curious to hear about them if you do. S2: yes ChatGPT: That’s lovely! What kinds of pets do you have? S2: 1 dog ChatGPT: A dog is such a wonderful companion! What breed is your dog, or what’s their personality like?
When R1 observed that S2 has shifted his attention from the physics task, she used the conversation to redirect him back to the topic instead of reprimanding him. The critical intervention was raised in line 3 when R1 drew a connection between the dog and the physics concept of motion by asking, “Do you bring your dog in your car ride? So do you drive slowly or fast with your dog there?” This question subtly reintroduces the theme of vehicle motion and inertia, thus connecting S2’s lived experience with the scientific concepts that were discussed earlier in the lesson.
After this exchange, S2 resumed asking questions related to friction (line 4). A few minutes later, he excitedly turned his laptop to show T1 what he had asked ChatGPT (line 6). He said, “Look! I asked him this,” and then elaborated, “I asked him ‘How does… brakes in a car work?’” (line 8). T1 responded with a warm, affirming “Ohh” (line 9), signalling her recognition of his effort and engagement. S2 continued sharing and mentioning about his pets in line 10, blending a personal context with scientific inquiry in a way that appears meaningful to him.
This episode illustrates how educator support can involve not only instruction and technical assistance but also subtle redirection strategies that sustain students’ attention and engagement. R1’s redirection was not abrupt or disciplinary; instead, it acknowledged S2’s interest while steering it back toward the scientific topic. This move reflects
Summary
The case of S2 across the three episodes provides insight into how a student with disabilities accessed, interacted with, and responded to ChatGPT when supported by responsive and context-sensitive scaffolding. The specific examples from S2’s interactions reveal that accessing ChatGPT was not merely a matter of logging in or typing prompts—it required educators to mediate linguistic, behavioural, and technological barriers in real time. Such layered support illustrates how students with disabilities require scaffolding that blends instructional and emotional-behavioural types of support to sustain interaction with GenAI tools.
S2’s continued and positive engagement with ChatGPT illustrates the impact of the educators’ support. It also shows that students’ meaningful interaction with ChatGPT emerged not from one-time access, but from the ongoing mediation of educator support. The examples across the three episodes illustrate how students with disabilities in this study were able to interact productively with GenAI when such interaction was contingent on the presence of timely scaffolding that was adapted to their specific needs across instructional, linguistic, emotional, and technological domains. While S2’s experience was distinctive, similar patterns were observed among other students, who also benefited from distributed educator support to maintain meaningful engagement with ChatGPT.
Discussion
The findings presented above offer detailed insights into how educators supported students with disabilities to access and engage with ChatGPT during science lessons. In this discussion, we reflect on the significance of these findings in relation to broader theoretical, pedagogical, and design implications for the use of GenAI in inclusive classrooms.
Reframing Scaffolding for GenAI in Inclusive Education
The findings from this study provide a new lens and taxonomy for understanding how educators support students with disabilities in using GenAI tools for learning. Although the categories of support (instructional, emotional-behavioural, technological, and literacy-based) were established and informed by prior research, this study updates and consolidates these domains into a single taxonomy that is specific to supporting students with disabilities in using GenAI in a classroom context. The contribution of this taxonomy lies not in introducing new categories, but in showing how these domains are operationalized in real-time during student–GenAI interaction. As shown in the findings, the various support categories do not operate in isolation, but are dynamically enacted in real-time in complex ways. For example, EA3’s support for S2 in Episode 1 involved modelling sentence structure (literacy), encouraging participation (emotional-behavioural), and taking over typing when needed (technological), all in the space of a few minutes.
The taxonomy that emerged from this study also captures the situated and adaptive nature of scaffolding required when GenAI is used. It goes beyond a checklist of strategies and instead surfaces the real-time responsiveness that characterizes effective scaffolding to sustain student-GenAI interaction. Such responsiveness is critical because GenAI tools like ChatGPT operate through text-based exchanges that are cognitively and linguistically demanding. For many students with disabilities, this mode of interaction introduces new forms of inaccessibility. Educators play a pivotal role not only in making the AI tool available, but in translating its affordances into meaningful and equitable learning opportunities for all students. The taxonomy therefore makes visible what is specific to GenAI-mediated interaction by showing how educator support must respond to the text-based, dialogic, and interpretive demands of AI-generated responses.
As a qualitative study situated in a single classroom context, the taxonomy does not claim to be exhaustive or universally applicable, but offers an empirically grounded way of describing educator scaffolding practices observed in this study. Nevertheless, the taxonomy offers a flexible framework that can be adapted across diverse inclusive contexts. Rather than prescribing fixed techniques, the four categories—instructional, literacy, emotional-behavioural, and technological—highlight key areas where support is often needed when students with disabilities engage with GenAI. These categories serve as a reflective tool for educators to anticipate moments that require intervention and tailor support according to their learners’ profiles and classroom conditions. In this way, the taxonomy enables principled adaptation to other specific educational contexts.
Scope and Specificity of the Scaffolding Taxonomy With GenAI
A key question arising from the findings is the extent to which the scaffolding taxonomy developed in this study reflects practices that are specific to students with disabilities when engaging with GenAI. On a general level, the four categories of support—instructional, literacy, emotional-behavioural, and technological—capture forms of educator support that are commonly observed in any classroom practice. Clear modelling of tasks, encouragement to persist, support with reading or composing digital text, and access to functional technology were all evident in the interactions analysed and formed an important foundation for students’ engagement with GenAI in this study.
What distinguishes the taxonomy in this context, however, is the specificity and intensity with which scaffolding was enacted in response to the students’ special needs. While many learners may require occasional assistance in formulating questions or interpreting digital text, the students observed in this study often required more sustained and carefully tailored scaffolding. This included repeated verbal modelling, step-by-step guidance, and ongoing emotional-behavioural support to sustain engagement with GenAI-mediated tasks. These practices reflect the forms of individualised support required by students with disabilities in this inclusive classroom and are closely tied to the interactional demands introduced by GenAI use. As such, while the taxonomy organises forms of support observed in this study, it should be understood as a situated analytic account rather than a generalisable model applicable across all learners or contexts.
Agency and Accessibility Within the Zone of Proximal Development
The findings of this study align closely with Vygotsky’s concept of ZPD, which emphasizes the role of more capable others in supporting learners to achieve tasks beyond their current independent abilities. In each of the episodes examined, educator scaffolding allowed students to operate within their ZPD in relation to GenAI. For instance, S2’s initial prompt about cars, his redirection toward a science topic, and his later use of audio playback to comprehend an AI-generated response, all illustrate how educator mediation bridged the gap between potential and actual performance.
Importantly, the scaffolding did not merely enable functional access to ChatGPT, but it also fostered student agency. In Episode 3, S2 chose to share his ChatGPT question with pride, turning his laptop to T1 and announcing what he had asked. This act signals a sense of ownership over the learning process. It is also evident in this study that when students are adequately scaffolded within their ZPD, they do not merely participate in a task; they also appropriated it as meaningful to themselves. For students with disabilities, such moments of agency are particularly significant, as they reflect not just inclusion in activity, but recognition as active and capable learners using an emerging technology.
GenAI as a Mediated Dialogic Partner
The findings in this study refine how GenAI may be understood as a dialogic partner in inclusive classrooms. In the interactions, student–GenAI exchanges were rarely sustained independently and was often structured through educator mediation. The findings thus suggest that GenAI functioned less as an autonomous dialogic partner and more as a dialogic resource whose potential was realized through educators’ support in getting students to initiate questions, interpret responses, and continue the interaction. This qualification is particularly important in the context of students with disabilities, whose interactions with GenAI were shaped not only by the tool’s affordances but also by the scaffolding required to make those affordances accessible. As shown in the episodes, dialogic engagement was distributed across the student, the GenAI tool, and the supporting adult, rather than residing in a simple one-to-one exchange between student and AI. In this sense, we can argue that GenAI’s dialogic function in the classroom was realized through a mediated process in which the tool, the student, and the educator together shaped the unfolding exchange.
Implications for GenAI Design, Inclusive Pedagogy and Teacher Education
While this study used ChatGPT as a pilot tool, the lessons learned point to broader implications for the design of future GenAI systems and the pedagogy surrounding their use, especially in the context of inclusive education. Although ChatGPT includes some accessibility features such as voice input and audio playback, these were evidently not sufficient for the diverse students’ needs observed in this study. For example, as shown in episode 2, R1 had to intervene and guide S2 manually to activate the audio feature. This action, though helpful, revealed the limitations of the tool’s interface for users who may not easily discover or activate such functions on their own. It should be noted that this study was carried out in late 2024 and the students were using the free version of ChatGPT 3.5. With the rapidly changing technological capabilities and user interfaces of GenAI tools, the students’ interactions documented in this study may be different. Nevertheless, while the specific applications may change, the taxonomy of support types emerged in this study may continue to offer analytic relevance for future studies examining more advanced GenAI systems.
For educators of students with disabilities, the findings underscore the importance of professional learning that prepares them to scaffold AI-mediated learning effectively, which is a crucial area in inclusive education (Waterfield et al., 2024). This includes not only technical familiarity with the tool but also pedagogical awareness of how to interpret students’ needs as they interact with GenAI. Future research can build on the taxonomy developed in this study and combine it with models of technology integration, such as the Technological Pedagogical Content Knowledge (Mishra, 2019; Mishra & Koehler, 2006) and the Technology Integration Matrix (Harmes et al., 2016). Such research may develop guidelines for teacher training, highlighting specific scaffolding practices that may be required in classrooms with learners with disabilities.
Practical Application of the Taxonomy
To support educators in applying this taxonomy of support, we outline some practical ways for each support to be embedded into inclusive classroom practice. For instructional support, teachers can model how to use GenAI, explain its relevance within the learning goals, and frequently relate a GenAI activity to the immediate inquiry tasks. For literacy support, teachers can co-construct prompts with students by highlighting key vocabulary, providing sentence starters, or decoding AI-generated texts with the students. For emotional-behavioural support, teachers can affirm student contributions, offer encouragement, and provide low-stakes choices to maintain engagement and reduce anxiety. Lastly, for technological support, teachers can assist students in navigating features such as audio playback or typing questions when motor or executive functioning challenges arise. By anticipating and embedding these types of support, educators can create more inclusive and responsive learning environments that support students with disabilities in engaging more meaningfully with GenAI.
Conclusion
This study responds directly to the urgent need to understand how GenAI can be made accessible and meaningful for students with disabilities. As these tools become increasingly embedded in classrooms and broader society, it is essential that their design and implementation do not reproduce or exacerbate educational inequities. By identifying the types of scaffolding that enable students with disabilities to engage meaningfully with GenAI, this research offers early insights into how GenAI might serve not only as a tool for literacy and inquiry, but also as a personalized learning assistant that can potentially support participation, confidence, and developmental growth. If the potential of GenAI as a learning technology is to be realized for all learners, then the integration of GenAI in inclusive classrooms must be a priority in current educational research.
Footnotes
Acknowledgements
During the preparation of this manuscript, OpenAI ChatGPT-4o was used to improve the readability and language of sentences and paragraphs. After using this tool, we reviewed and edited the content as needed and took full responsibility for the content of the article.
Ethical Considerations
Ethical approval for the study was obtained from the Human Research Ethics Committee of Curtin University in Australia (Project number: HREC2023-0413).
Consent to Participate
Informed consent was secured from both students and their guardians, with assurances that all interactions would remain anonymous and confidential. All personal names used in this paper are pseudonyms.
Author Contributions
Kok-Sing Tang: Conceptualization, Project administration, Formal analysis, Investigation, Methodology, Validation, Writing – original draft, Writing – review and editing
Natasha Anne Rappa: Data curation, Project administration, Validation, Writing – review and editing
Karen Nonis: Data curation, Project administration, Validation
Khansa Ilyas: Data curation, Formal analysis, Validation
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This paper draws on data from the research project: “Transforming Teaching, Learning and Assessment with Generative AI in a Western Australian School” funded by the Innovation & Excellence Award, School of Education, Curtin University.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
Data are available on request due to privacy/ethical restrictions.