CoChat
Sign Up
Cited by 0Artificial Intelligence in EducationEducational Technology in LearningEthics and Social Impacts of AI

The Use of Generative AI by Students with Disabilities in UK Higher Education: A Mixed-Methods Survey of 125 Students

Open access
Shepherd, Nicki James·2025·Journal of Knowledge Learning and Science Technology ISSN: 2959-6386 (online)
DOI 10.60087/jklst.vol4.n4.016
Open this paper in CoChat — free accountOpen PDF

Abstract

Higher education's response to generative artificial intelligence (AI) has moved quickly from curiosity to policy, yet questions of who benefits from these tools and who is excluded remain under-examined. Students with disabilities — a population long under-represented in higher education and demonstrably reliant on assistive technologies — may have much to gain from generative AI, but also face specific risks from it; empirical evidence on their actual practices is scarce. This study asks how students with disabilities at a UK university use generative AI in academic writing, what concerns they hold about it, and what institutional support they want. A mixed-methods online survey, anchored in a five-element generative AI literacy framework, was distributed to all 1,797 students registered as disabled at one UK university during February and March 2024. The survey returned 125 valid responses, a 7.0 % response rate. Closed items were analysed descriptively; open-text responses were coded inductively through content analysis. Seventy-eight percent of respondents reported using generative AI in their studies, with ChatGPT (53 %) the dominant tool, supplemented by rewriting applications such as Grammarly (10 %) and translation services. Respondents used these tools across the full writing pipeline — explaining concepts, structuring arguments, summarising readings, and refining prose. Concerns clustered around the accuracy of AI outputs, threats to academic integrity, and the inequity introduced by paid subscriptions. Participants voiced strong demand for involvement in institutional AI policy (M = 4.03 on a five-point scale) and for university-provided training (M = 3.95). The paper closes with policy and pedagogical recommendations for building a more inclusive institutional response to generative AI.

AI snapshot

Written from the paper’s full text.

TL;DR

A mixed-methods survey of 125 students registered as disabled at one UK university found that 78% used generative AI in academic writing, chiefly ChatGPT, and wanted institutional involvement and training rather than bans. Students reported informed concerns about accuracy, academic integrity, and paid-access inequity, while using AI across the writing pipeline.

Key findings

  • Of 125 respondents, 97 (78%) reported using generative AI; 28 (22%) reported infrequent or no use, and 91% did not pay for generative AI services.
  • ChatGPT was the dominant tool (n = 66, 53%), followed by Gemini (n = 11, 9%), Copilot (n = 5, 4%), and Claude (n = 2, 2%); Grammarly was the main rewriting tool (n = 13, 10%).
  • ChatGPT uses spanned refining writing (n = 25, 20%), structuring writing (n = 22, 18%), explaining topics (n = 19, 15%), summarising learning materials (n = 18, 14%), and identifying learning resources (n = 10, 8%).
  • Top concerns were misuse for “bad purposes such as deepfakes” (M = 4.03), inaccuracy (M = 3.69), academic integrity (M = 3.57), and pay-for-access equity (M = 3.15); support items rated highest were student involvement in AI policy (M = 4.03) and university-provided training (M = 3.95), while support for a ban was low (M = 1.88).

Methods

  • Mixed-methods online survey distributed census-style to all 1,797 students registered with a disability at one UK university during February and March 2024; 125 valid responses were returned, a 7.0% response rate.
  • The instrument covered respondent characteristics, disability information using HESA terminology refined with institutional Disability and Dyslexia Support Services, and generative AI use, costs, perceptions, concerns, and recommendations; open-text questions probed disability impact on writing, AI uses, and desired support.
  • Closed items were analysed descriptively; open-text responses were coded inductively through content analysis, anchored in a five-element generative AI literacy framework: pragmatic, safety, reflective, socio-ethical, and contextual understanding.

Limitations

  • The paper states that non-response bias is plausible, that the 125-response sample and 7.0% response rate constrain disaggregation by disability category and cross-sectional comparison, that the female skew limits gender-based analysis, and that the study is single-institution and single-country.
  • Evident from the text: the design is cross-sectional and self-report, so it cannot establish causality or track changes over time; the sample is self-selected from one institution, limiting generalisability.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.60087/jklst.vol4.n4.016

Cites

35 of 35 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

12 of 29 cited sources checked · 2 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Visually impaired learners may struggle to interact with text-based chatbots.

    “Visually impaired learners may struggle to interact with text-based chatbots (Tlili et al., 2023).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Tlili et al., 2023 source

      In “Personality and emotions” (page unavailable), the paper says ChatGPT is limited to a textual interface. “Nothing should be taken for granted” (page unavailable) calls for research on whether chatbots meet the needs of students with disabilities. It does not specifically say that visually impaired learners struggle to interact with text-based chatbots. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
  • Premisecitation

    Generative AI can scaffold reading for students whose disabilities, particularly dyslexia, make text processing slow or error-prone.

    “Generative AI can scaffold reading for students whose disabilities -dyslexia in particular -make text processing slow or error-prone (Brewer, Urwin, & Witham, 2023; Schiavo, Mana, Mich, Zancanaro, & Job, 2021; Tamdjidi & Billai, 2023).”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Brewer, Urwin, & Witham, 2023 source

      The Impact on education section says disabled students may need additional time to process information, and the Introduction mentions assistive technology as a learning accommodation. But the paper does not say that generative AI scaffolds reading, or specifically link dyslexia to slow or error-prone text processing. Its discussion of AI concerns disability-disclosure administration, not reading support (Impact on education; page unavailable). (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites Schiavo et al., 2021 source

      The paper describes slow and inaccurate word reading in dyslexia (Attention-driven Read-aloud Technology, p. 4) and explains how text-to-speech and read-aloud tools support readers with dyslexia (Related works, p. 7). It studies assistive read-aloud technology, not generative AI, so it does not establish the claim about generative AI specifically. (LLM, conf=0.90)

      Experimental — accuracy not yet measured
    • cites Tamdjidi & Billai, 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Premisecitation

    There is plausibly real scope for generative AI to support reading, summarisation, and interactive scaffolding for students with dyslexia and related conditions.

    “There is, plausibly, real scope for generative AI to support reading, summarisation, and interactive scaffolding -domains where students with dyslexia and related conditions experience pronounced difficulty (Zhao, Cox, & Cai, 2024).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Zhao, Cox, & Cai, 2024 source

      The Findings (page not available) report students using AI tools to summarise readings and ChatGPT for tasks they found difficult; a participant also describes using its feedback to develop paragraphs. This supports the proposed reading, summarisation and interactive-help uses in general, but the paper does not address dyslexia or related conditions or establish that these tools support students with those conditions. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisereasoning

    None of the existing surveys and adoption studies of student generative AI use has focused on students with disabilities as the primary unit of analysis.

    “None of these, however, has focused on students with disabilities as the primary unit of analysis”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Almasaad et al., 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Acosta-Enriquez et al., 2024 source

      The Method > Participants section identifies the unit of analysis as university students in a 499-person sample from Peru; it does not identify students with disabilities as the study’s focus. This supports the claim for this survey, but the paper does not establish that none of the other existing surveys or adoption studies focused on students with disabilities. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites Chan & Hu, 2023 source

      The Methodology and Demographic information (p. 6) show that this paper surveys university students across disciplines, not students with disabilities as its primary group. Its review of earlier surveys (p. 5) does not establish that none of the other studies focused on students with disabilities, so the broader claim is unsupported by this source. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
  • Interpretationcitation

    Unauthorised use of unsanctioned generative AI tools risks producing unequal exposure to risk among an already marginalised group (students with disabilities).

    “unauthorised use of unsanctioned tools risks producing unequal exposure to risk among an already marginalised group (Almufareh et al., 2024).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Almufareh et al., 2024 source

      The 'Addressing algorithmic bias' section says AI bias could lead to unequal access for people with disabilities, and 'Preserving privacy and data security' identifies risks involving their sensitive data. Neither section links unequal risk exposure among students with disabilities to the unauthorised use of unsanctioned generative AI tools. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
  • May combine two claims. 31
  • Classified as probably not a substantial claim. 32
  • The quote only partly states this claim. 8

  • May combine two claims.

    Premisecitation

    Linguistic models have been shown to encode disability-related terms with negative valence, reinforcing the bias problem.

    “Linguistic models have been shown to encode disabilityrelated terms with negative valence (Venkit & Wilson, 2021), reinforcing the bias problem (McMurtrie, 2023; Rocky Mountain ADA Center, 2023).”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Venkit & Wilson, 2021 source
    • cites McMurtrie, 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Rocky Mountain ADA Center, 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    Quality-of-teaching gains include using AI to check accessibility of materials, identify early indicators of learning difficulty, and produce personalised teaching strategies.

    “Quality-of-teaching gains include using AI to check accessibility of materials, identify early indicators of learning difficulty (Johnson, Smart, & Mahar, 2023), and produce personalised teaching strategies (Bozkurt et al., 2023).”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Johnson, Smart, & Mahar, 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Bozkurt et al., 2023 source

  • May combine two claims.

    Premisecitation

    Generative models trained on under-representative datasets are prone to producing generic outputs that fail the specific needs of marginalised users, increasing cognitive load.

    “Generative models trained on under-representative datasets are prone to producing generic outputs that fail the specific needs of marginalised users (Bender, Gebru, McMillan-Major, & Shmitchell, 2021; Jafry & Vorstermans, 2024), increasing rather than reducing the cognitive load on those users (Botchu et al., 2023;Venkit & Wilson, 2021).”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Bender et al., 2021 source
    • cites Jafry & Vorstermans, 2024 source

      The Literature Review (page not provided) says that biased training data shape AI outputs and that Western-centric data can exclude or misrepresent non-Western perspectives. It does not say that generative models produce generic outputs that fail marginalised users’ specific needs. The sentence cites other works, not this paper, for the cognitive-load claim. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites Botchu et al., 2023 source

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    The possibility of net benefit of generative AI for students with disabilities has been raised but not systematically tested.

    “Within this contested terrain, the possibility of net benefit for students with disabilities has been raised but not systematically tested.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites Addy, Kang, Laquintano, and Dietrich (2023)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Evmenova, Borup, and Shin (2024) source

      The Abstract raises potential benefits for students with disabilities while saying generative AI’s full impact is unclear. The Purpose and Method describe an exploratory study of teachers’ perceptions, not a test of net student benefit. The paper does not establish the broader claim that net benefit has not been systematically tested elsewhere. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites Upadhyay (2024) source

  • May combine two claims.

    Interpretationown data

    Students' accounts show well-developed pragmatic and safety dimensions of AI literacy, but less developed socio-ethical and contextual dimensions.

    “pragmatic and safety dimensions are well developed in students' accounts; socio-ethical and contextual dimensions are less so.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Zhao, Cox, and Cai's (2024) source

      The Discussion and Table 3 describe substantial generative AI literacy among the student interviewees, good awareness of privacy risks, and limited awareness of societal impacts. They do not clearly characterize contextual understanding as less developed: Table 3 says appropriate use was at the forefront of participants’ minds, though they sought institutional clarification. The paper also notes limits to students’ pragmatic skills, including unsophisticated prompt engineering. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Concerns clustered around the accuracy of AI outputs, threats to academic integrity, and the inequity introduced by paid subscriptions.

    “Concerns clustered around the accuracy of AI outputs, threats to academic integrity, and the inequity”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Of the 125 respondents, 105 described the way their disability shaped their academic writing; 18 declined to answer the open question, and two reported no effect.

    “Of the 125 respondents, 105 described the way their disability shaped their academic writing; 18 declined to answer the open question, and two reported no effect.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    ChatGPT was the dominant generative AI tool, used by 53% of respondents.

    “with ChatGPT (53 %) the dominant tool”
    • Also stated in: § 4.2. Use of generative AI · boosted
      “ChatGPT (n = 66, 53 %) far ahead of Gemini”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The verb 'struggle' appeared 38 times in the open-text responses, and feeling overwhelmed recurred across accounts.

    “The verb struggle appeared 38 times and feeling overwhelmed recurred across accounts.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Ten respondents (8%) specified training on ethical use.

    “27 (22 %) and 10 (8 %) specified training on avoiding misuse and ethical use respectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Grammatical accuracy was a stated reason for using rewriting applications, reported by 10 respondents (8%).

    “grammatical accuracy (n = 10, 8 %)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Search and referencing applications was requested by 12 respondents (10%).

    “search and referencing applications (n = 12, 10 %)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The role of rewriting applications was often complementary to ChatGPT, with Grammarly applied to polish and ChatGPT applied to restructure.

    “The tool's role was often complementary to ChatGPT -Grammarly applied to polish, ChatGPT applied to restructure.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Google Translate was used by 5 respondents (4%).

    “Google Translate (n = 5, 4 %)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    DeepL was used by 4 respondents (3%).

    “DeepL (n = 4, 3 %)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Writing prompts was requested by 15 respondents (12%).

    “writing prompts (n = 15, 12 %)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The wider concerns above were echoed by respondents who reported infrequent or no use, who additionally cited their own lack of expertise as a barrier.

    “The wider concerns above were echoed by these respondents, who additionally cited their own lack of expertise as a barrier.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Precision of word choice was a stated reason for using rewriting applications, reported by 3 respondents (2%).

    “precision of word choice (n = 3, 2 %)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Awareness of generative AI's controversies was high, and concerns were patterned.

    “Awareness of generative AI's controversies was high, and concerns were patterned (Figure 3).”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data

    Training on generative AI emerged as a priority for one-third of respondents.

    “Training emerged as a priority for one-third of respondents.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Two respondents specifically described a worry that ChatGPT would erode the development of their own voice in writing.

    “Two respondents specifically described a worry that ChatGPT would erode the development of their own voice in writing.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The five-point scales on concerns returned high means on accuracy (M = 3.69) and academic integrity (M = 3.57).

    “The fivepoint scales on concerns return high means on accuracy (M = 3.69) and academic integrity (M = 3.57);”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Translation tools were used predominantly by students who needed to comprehend English-language readings.

    “were used predominantly by students who needed to comprehend English-language readings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Quillbot was used by 3 respondents (2%).

    “Quillbot (n = 3, 2 %)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Use of generative AI among respondents was dominated by ChatGPT, supplemented by Grammarly and translation tools, with very limited paid subscription uptake.

    “The pattern is dominated by ChatGPT and supplemented by Grammarly and translation tools, with very limited paid subscription uptake.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    Open-text responses were saturated with affect, including struggle and feeling overwhelmed.

    “The open-text responses are saturated with affect -struggle, overwhelmed -”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The most highly weighted concerns are about effects on students themselves directly; wider societal harms (environmental, supply-chain, voice) score lower.

    “but the most highly weighted concerns are about effects on themselves directlywider societal harms (environmental, supply-chain, voice) score lower.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    One unspecified translation product was used by 1 respondent (1%).

    “one unspecified product (n = 1, 1 %)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Sentence coherence was a stated reason for using rewriting applications, reported by 3 respondents (2%).

    “sentence coherence (n = 3, 2 %)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Summarisation was requested by 7 respondents (6%).

    “summarisation (n = 7, 6 %)”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data

    Identifying learning resources was a function of ChatGPT, reported by 10 respondents (8%).

    “Identifying learning resources (n = 10, 8 %).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Claude was used by 2 respondents (2%).

    “Claude (n = 2, 2 %)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Translation tools were used by respondents working with non-English source material and by international students refining English-language output.

    “Translation tools -Google Translate (n = 5), DeepL (n = 4), and one unspecified -were used by respondents working with non-English source material and by international students refining English-language output.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Wordtune was used by 1 respondent (1%).

    “Wordtune (n = 1, 1 %)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Exposure to a wider range of tools beyond the most familiar was requested by 4 respondents (3%).

    “exposure to a wider range of tools beyond the most familiar (n = 4, 3 %)”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingnone given

    Students hold informed concerns about accuracy and academic integrity.

    “hold informed concerns about accuracy and academic integrity”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. The quote only partly states this claim.

    Findingnone given

    Students want to be consulted on policy rather than face a ban.

    “want to be consulted on policy rather than face a ban.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ChatGPT is central to generative AI use by students with disabilities, used across the full writing pipeline from concept-explanation to final polish, supplemented by rewriting and translation tools.

    “ChatGPT, used across the full writing pipeline from concept-explanation to final polish, is central to that practice, supplemented by rewriting and translation tools.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ChatGPT was used to formulate search terms and broaden search scope.

    “ChatGPT formulated search terms and broadened search scope.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingnone given

    Students value the support generative AI tools offer.

    “Students value the support these tools offer”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The survey was distributed census-style to all 1,797 students registered with a disability at the host UK university during February and March 2024.

    “Distribution was census-style.All 1,797 students registered with a disability at the host UK university during February and March 2024 received the survey.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Summarising long texts into concise, jargon-light passages aids comprehension of core concepts and supports engagement with course materials.

    “Summarising long texts into concise, jargon-light passages aids comprehension of core concepts (Say, 2024) and supports engagement with course materials (Tamdjidi & Billai, 2023).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Say, 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Tamdjidi & Billai, 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Ethical approval was issued by the University of [Anon], and informed consent was obtained at the survey's start.

    “Ethical approval was issued by the University of [Anon], and informed consent obtained at the survey's start.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Facial-recognition components lack adequate training data on autistic users and consequently misinterpret affect, producing tailored support failures.

    “Facial-recognition components lack adequate training data on autistic users and consequently misinterpret affect, producing tailored support failures (Nacheva & Czaplewski, 2024).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Nacheva & Czaplewski, 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    No identifying information was collected.

    “No identifying information was collected.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Early ChatGPT versions confined interaction to text, a limitation for students with dyslexia.

    “Early ChatGPT versions confined interaction to text -a limitation for students with dyslexia (Botchu et al., 2023).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Botchu et al., 2023 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Under-representation of disabled students in UK higher education has been a long-running policy concern, and the statutory duty on universities to advance equality of opportunity for this group remains only partly fulfilled.

    “Under-representation of disabled students in UK higher education has been a long-running policy concern, and the statutory duty on universities to advance equality of opportunity for this group remains only partly fulfilled (Hubble & Bolton, 2021).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Hubble & Bolton, 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • The quote only partly states this claim.

    Interpretationown data

    Motivations for using generative AI clustered around future utility rather than enjoyment.

    “Motivations clustered around future utility rather than enjoyment.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    The survey yielded 125 valid returns, a 7.0% response rate.

    “The response yielded 125 valid returns, a 7.0 % response rate.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Dysgraphia and related conditions create difficulty with planning, organising ideas, and revising for meaning and form.

    “Dysgraphia and related conditions create difficulty with planning, organising ideas, and revising for meaning and form (Graham, Harris, & McKeown, 2013”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Graham, Harris, & McKeown, 2013

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Qualitative data were coded inductively, with codes generated from the data rather than imposed from a pre-existing scheme.

    “qualitative data were coded inductively, with codes generated from the data rather than imposed from a pre-existing scheme.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Generative AI is functioning as effective assistive technology for many disabled students, often outside institutional knowledge or sanction.

    “Generative AI is, on this evidence, functioning as effective assistive technology for many disabled students, often outside institutional knowledge or sanction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    The absolute n of 125 constrains disaggregation by disability category and limits cross-sectional comparison (e.g., between international and home students, or between undergraduate and postgraduate cohorts).

    “the absolute n still constrains disaggregation by disability category and limits crosssectional comparison (for example, between international and home students, or between undergraduate and postgraduate cohorts).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The dominance of ChatGPT, with very limited use of alternatives, suggests an under-developed pragmatic literacy in tool selection.

    “The dominance of ChatGPT, with very limited use of alternatives, suggests an under-developed pragmatic literacy in tool selection”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The policy environment, available institutional tools, and existing support architecture shape what student practice looks like in any particular setting.

    “the policy environment, available institutional tools, and existing support architecture all shape what student practice looks like in any particular setting.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The absence of disability-specific guidance on generative AI is a concrete and addressable gap.

    “the absence of disability-specific guidance is, given these findings, a concrete and addressable gap.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Non-response bias is plausible: only students using or interested in generative AI, or with strong opinions about it, may have responded.

    “Non-response bias is plausible: only students using or interested in generative AI, or with strong opinions about it, may have responded.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    AI tools function not only as productivity aids but as scaffolds against the anxiety and disability-linked exhaustion of academic writing for students with disabilities.

    “the AI tools function not only as productivity aids but as scaffolds against the anxiety and disability-linked exhaustion of academic writing.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A recommended product that students do not adopt cannot deliver the safety and policy benefits universities intend it to.

    “a recommended product that students do not adopt cannot deliver the safety and policy benefits universities intend it to.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductreasoning

    The study is single-institution and single-country.

    “The study is also single-institution, single-country;”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The sample has a female skew that reflects the institution's overall student population.

    “The female skew reflects the institution's overall student population”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    If institutions do not act, disabled students will keep using generative AI tools outside institutional knowledge or sanction.

    “disabled students will keep using these tools outside institutional knowledge or sanction”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Generative AI has become a routine part of academic writing for students with disabilities in UK higher education, often without institutional knowledge.

    “Generative AI has become a routine part of academic writing for students with disabilities in UK higher education, often without institutional knowledge.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The equity and safety risks created by disabled students using generative AI outside institutional knowledge or sanction will persist if institutions do not act.

    “the equity and safety risks that pattern creates will persist.”
    Quote in the paperSubstantial claim
64 of 64

  • The paper relies on this without support.

    Students with disabilities are demonstrably reliant on assistive technologies.

    “a population long under-represented in higher education and demonstrably reliant on assistive technologies”
1 of 1

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Discussion misidentifies the highest-rated concern

    Minor

    The discussion says the most highly weighted concerns are about effects on students themselves, but the results identify misuse for purposes such as deepfakes as the top-ranked concern.

  • Structuring is called the most-cited function despite a larger refining group

    Minor

    The paper calls structuring writing the most-cited single ChatGPT function at 22 respondents, but reports 25 respondents for refining writing and calls that the largest use cluster.

2 of 2

Cites

Flagged first, then by citations

15 of the 50 references couldn’t be matched to a known paper.

Research works printed but not linked: 3

  1. Addy, T., Kang, T., Laquintano, T., & Dietrich, V. (2023). Who benefits and who is excluded? Transformative learning, equity, and generative AI in education. Journal of Transformative Learning, 10(2), 92-103.
  2. Graham, S., Harris, K. R., & McKeown, D. (2013). The writing of students with learning disabilities and a meta-analysis of SRSD writing intervention studies: Redux. In H. L. Swanson, K. R. Harris, & S. Graham (Eds.), Handbook of learning disabilities (2nd ed., pp. 405-438). Guilford Press.
  3. Johnson, C., Smart, K., & Mahar, J. (2023). Is there a place for artificial intelligence in special education? National Social Science Technology Journal, 48.

Authors
Shepherd, N. J.
Corresponding author
Shepherd, Nicki James
Venue
Journal of Knowledge Learning and Science Technology ISSN: 2959-6386 (online)
Publisher
Open Knowledge
Published
2025-12-25
Volume
4
Issue
4
Pages
142-151
ISSN
2959-6386 (ISSN-L)
DOI
10.60087/jklst.vol4.n4.016
OpenAlex
W7164658632
Semantic Scholar
f182a64b015e1c164273f0fd099d6ae9dcfd0807
Cited by
0
References
50
FWCI
0.00
Type
journal article
Open access
Licence
Locations
Journal of Knowledge Learning and Science Technology (publishedVersion, cc-by)
Topics
  • Physical Sciences › Computer Science › Information Systems › Artificial Intelligence in Education
  • Physical Sciences › Computer Science › Information Systems › Educational Technology in Learning
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
Keywords
generative AI, students with disabilities, higher education, academic writing, ChatGPT, AI literacy, academic integrity, assistive technology, institutional AI policies, inclusive education, disability inclusion, AI training for students

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-20.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-20.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review