CoChat
Sign Up
Cited by 2Digital Accessibility for DisabilitiesAssistive Technology in Communication and MobilityArtificial Intelligence in Education

Exploring Generative AI to Support Disability Service Professionals in Writing Image Descriptions for HCI Science Figures

Open access
Iwamoto, Yugo, Raees, Muhammad, Heard, Jamison, Tigwell, Garreth W.·2025·Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility
DOI 10.1145/3663547.3759763
Open this paper in CoChat — free accountOpen PDF

Abstract

Alternative text (alt text) or image descriptions for scientific figures make them accessible to screen reader users.Yet, generating high-quality alt text remains a significant challenge, particularly for people lacking subject-matter expertise.Emerging AI tools, such as generative AI, can help people understand contexts where they lack subject expertise.This work explores how university Disability Services Office (DSO) professionals write alt text for scientific figures without subject expertise.We conducted a user study with 12 DSO professionals who authored alt text for scientific figures, first without and then using generative AI.We assessed the participants' processes for alt text generation and their confidence in the resulting output.We also conducted semi-structured interviews to understand DSO professionals' experiences with the alt text generation.Our findings reveal that AI assistance improved participants' confidence and efficiency, but introduced new challenges for interaction and trust.Participants expressed caution about relying on AI-generated descriptions without sufficient domain knowledge and editing.These insights highlight the need for AI-augmented tools that better scaffold the alt text writing process for accessibility professionals.

AI snapshot

Written from the paper’s full text.

TL;DR

In a user study with 12 university Disability Services Office professionals, AI assistance (ChatGPT-4o) improved confidence and efficiency when writing alt text for HCI science figures, but introduced interaction and trust challenges. Participants produced significantly longer descriptions with AI and all preferred AI assistance for final output, while cautioning about relying on AI without domain knowledge and editing.

Key findings

  • AI assistance improved confidence and efficiency: all 12 participants completed four figures within 20 minutes with AI support, whereas without AI all used the full 50 minutes and two could not complete all figures; all 12 agreed to pick AI assistance over manual effort for the final output.
  • Human-AI descriptions were significantly longer than human-only descriptions: paired-samples t-test, human-only mean = 104.16, SD = 66.21 vs. human-AI mean = 140.70, SD = 79.65; t(36) = -2.09, p = 0.044.
  • Participants reported higher confidence and text quality with AI, especially for complex or unfamiliar content, but struggled with interaction and trust: prompt wording, right level of detail, over-reliance, and less experienced participants accepting outputs with minimal interaction.
  • Some participants missed key information when they only uploaded figures and captions without prompting for accessibility context; others prompted AI that they needed alt text for students with vision impairment. Trust varied, e.g., P12: “I don’t trust it enough.”

Methods

  • IRB-approved user study with 12 DSO professionals (1–15 years of alt text experience, mean = 5.91, SD = 4.62; age 20–59, mean = 33.66, SD = 13.49; nine female, one male, two nonbinary; 11 reported English as primary language, one opted out). Each attended a 90-minute session.
  • Within-subjects procedure: participants wrote alt text for four HCI figures without AI (10–12 minutes per figure), then four figures from another article using AI assistance (ChatGPT 4o; expected 20–25 minutes). Four articles were counterbalanced, yielding six versions of alt text per article. Think-aloud, semi-structured interviews, and a closing questionnaire with Likert confidence and task-load ratings were used.
  • Analysis: thematic analysis of transcribed sessions; descriptive analysis and a paired-samples t-test on produced alt text word counts.

Limitations

  • Stated by the paper: participants wanted more time due to unfamiliarity with content and AI assistance; only one subject domain (HCI) was used; some participants noted they would perform better in their own offices or workstations; confidence and quality reflect self-reported experiences and may not have been analytically assessed for fact-checking; a single LLM was used, and different models may elicit different results.
  • Evident from the text: the sample is small (12 DSO professionals), and the study did not evaluate alt text quality with domain experts; the paper states future work will evaluate human vs. AI-generated alt text with HCI domain experts on quality and fact-checking.

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

Integrity signals

Registry

Crossref · 10.1145/3663547.3759763

Cites

32 of 33 references verified1 citation errors · 0 not checkable

Statistics

Claims the paper makes

0 of 50 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Some participants only uploaded figures and captions and ended up missing some key information in the generated answer because they did not prompt the LLM for accessibility or provide accessibility guidelines.

    “In contrast, some participants only uploaded figures and captions, and ended up missing some key information in the generated answer, as they did not prompt the LLM for accessibility or provide accessibility guidelines.”
    Quote in the paperNot a substantial claim
  • Findingown data

    With AI assistance, participants struggled with determining the right level of detail and wording in prompts to achieve the desired output.

    “With AI assistance, participants struggled with determining the right level of detail and wording in prompts to achieve the desired output.”
    Quote in the paperSubstantial claim
  • Findingown data

    Participants expressed caution about relying on AI-generated descriptions without sufficient domain knowledge and editing.

    “Participants expressed caution about relying on AI-generated descriptions without sufficient domain knowledge and editing.”
    • Also stated in: § 3.3 Controlling the Output · hedged
      “However, some participants expressed caution about AI that omits visual details and creates ambiguity.”
    Quote in the paperSubstantial claim
  • Findingown data

    Participants faced challenges in understanding and articulating complex terms and content without adequate background knowledge.

    “Participants faced challenges in understanding and articulating complex terms and content without adequate background knowledge.”
    Quote in the paperSubstantial claim
  • Findingown data

    Participants appreciated that when they interacted with AI, it helped improve the structure, contextual relevance, and concise expression of the figure.

    “Participants appreciated that when they interacted with AI, it helped improve the structure, contextual relevance, and concise expression of the figure.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 16
  • May combine two claims. 22

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

    Findingown data

    A couple of participants noted that they would perform better if they did it in their offices or workstations.

    “Although we provided participants with materials and computer screens, a couple of participants noted that they would perform better if they did it in their offices or workstations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Participants highlighted AI's value in summarizing content and expediting content generation.

    “Participants highlighted AI's value in summarizing content and expediting content generation.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Participants expressed the need for more time due to unfamiliarity with the content and AI assistance.

    “For example, participants expressed the need for more time due to unfamiliarity with the content and AI assistance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Participants generated and accepted longer responses for alt text with AI assistance.

    “Participants generated and accepted longer responses for alt text with AI assistance.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Some participants expressed that their iterative prompting and previous experience with generative AI made them more confident in generating alt text.

    “Some participants expressed that their iterative prompting and previous experience with generative AI made them more confident in generating alt text.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    All participants expressed satisfaction with LLM-generated responses and appreciated their assistance.

    “Still, all participants expressed satisfaction with LLMgenerated responses and appreciated their assistance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Participants reported higher assistance support and text quality in creating alt text with AI assistance, especially with complex or unfamiliar content.

    “As shown in Figure 2, participants report higher assistance support and text quality in creating alt text with AI assistance, especially with complex or unfamiliar content.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Some participants spent time refining outputs but acknowledged that the editing process was faster and less taxing than starting from scratch.

    “While some participants spent time refining the outputs, they acknowledged that the editing process was faster and less taxing than starting from scratch.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    Less experienced participants tended to accept the alt text, showing a very minimal interaction.

    “Less experienced participants tended to accept the alt text, showing a very minimal interaction.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Most participants asked generative AI to analyze and describe the article by uploading it, to ensure participants' knowledge and AI's understanding of the article were aligned better.

    “Most participants asked generative AI to analyze and describe the article by uploading it, to ensure participants' knowledge and AI's understanding of the article were aligned better.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Without AI assistance, participants found describing the provided figures challenging, citing struggles with content comprehension and limited time availability.

    “For writing alt text without the aid of AI, and with years of expertise in their roles, participants still found describing the provided figures challenging, citing struggles with content comprehension and limited time availability.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Lack of subject expertise and limited institutional support increases the cognitive and procedural burden on DSO professionals in developing high-quality alt text for complex content.

    “This lack of subject expertise and limited institutional support increases the cognitive and procedural burden on DSO professionals in developing high-quality alt text for complex content [14,23,34].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [14]

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

    • cites [23]

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

    • cites [34]

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

  • May combine two claims.

    Premisecitation

    Describing scientific figures requires both technical literacy and familiarity with accessibility standards, which can be challenging for DSO professionals working with diverse subjects.

    “Hence, describing scientific figures requires both technical literacy and familiarity with accessibility standards, which can be challenging for DSO professionals working with diverse subjects [5,29].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [5]

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

    • cites [29]

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

  • May combine two claims.

    Premisecitation

    Emerging generative AI tools, such as large language models (LLMs), have shown potential to support non-experts with writing assistance and can assist accessibility professionals in understanding scientific figures without having deep expertise in the subjects or mastering accessibility guidelines.

    “However, emerging generative artificial intelligence (AI) tools, such as large language models (LLMs), have shown potential to support non-experts with writing assistance, and can assist accessibility professionals in understanding scientific figures without having deep expertise in the subjects or mastering accessibility guidelines for writing alt texts for their students [9,27,30,36].”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [9]

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

    • cites [27]

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

    • cites [30]

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

    • cites [36]

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

  • May combine two claims.

    Premisecitation

    Screen reader users struggle to understand visuals, getting around 60% less information due to insufficient alt text, particularly in academic contexts.

    “Screen reader users struggle to understand visuals, getting around 60% less information [25], due to insufficient alt text, particularly in academic contexts.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25]

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

  • May combine two claims.

    Premisecitation

    DSO staff resort to time-consuming workarounds or rely on inconsistent faculty collaborations, ultimately impacting the learning experience of students.

    “As a result, DSO staff resort to time-consuming workarounds or rely on inconsistent faculty collaborations, ultimately impacting the learning experience of students [5,23,37].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [5]

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

    • cites [23]

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

    • cites [37]

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

  • May combine two claims.

    Premisecitation

    Automated methods such as computer vision or LLMs do not apply well to science figures, requiring refined control by end users to synthesize the content through interaction.

    “In addition, these automated methods, such as computer vision or LLMs, do not apply well to science figures [5,26], requiring refined control by end users to synthesize the content through interaction.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [5]

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

    • cites [26]

      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.

    Premisecitation

    Singh et al. reported that interactive LLMs significantly improve the quality of alt text while lowering the required cognitive load.

    “Their work reported significantly improving the quality of alt text while lowering the required cognitive load, as reported by involved participants.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27]

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

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

    Premisecitation

    Understanding visual content requires rich image descriptions that go beyond explaining surface-level figure objects.

    “Understanding visual content requires rich image descriptions that go beyond explaining surface-level figure objects [18,19].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [18]

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

    • cites [19]

      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

    The IRB-approved study followed explicit protocols for tasks and data collection.

    “The IRB-approved study followed explicit protocols for tasks and data collection.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Poorly labeled graphics with missing or vague alt text lead to user frustration.

    “Lazar et al. [14] emphasize that poorly labeled graphics with missing or vague alt text lead to user frustration.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [14]

      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

    Minor differences in analysis were resolved through discussion and review of the transcripts.

    “Minor differences were observed and resolved through discussion and review of the transcripts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Limited expertise in effective AI interaction showed underdeveloped prompting strategies and over-reliance without analytically assessing AI outputs.

    “Limited expertise in effective AI interaction showed underdeveloped prompting strategies and over-reliance without analytically assessing AI outputs.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    Articles between the two tasks were counterbalanced to avoid stereotypes, and no additional knowledge was provided that may shape the results between the two tasks.

    “Provided articles between the two tasks were counterbalanced to avoid stereotypes, and no additional knowledge was provided that may shape the results between the two tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Study design and conductnone given

    The confidence in the generated text and its quality reflects participants' self-reported experiences, and they may not have analytically assessed it while writing the alt text quality for fact-checking.

    “The confidence in the generated text and its quality also reflects participants' selfreported experiences, and they may not have analytically assessed it while writing the alt text quality for fact-checking.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    AI support can enhance DSO professionals' expertise in content, while they integrate their best practices to produce effective alt text for scientific figures.

    “AI support can enhance their expertise in content, while they integrate their best practices to produce effective alt text for scientific figures.”
    Quote in the paperSubstantial claim

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

    Study design and conductreasoning

    The use of four articles was counterbalanced to ensure that no participant uses any article twice.

    “To ensure that no participant uses any article twice, we counterbalanced the use of four articles, resulting in six versions of alt text (three human-only generated and three human-AI-generated) produced for each article.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The experimental settings used only one subject domain (i.e., HCI), which may not capture the full spectrum of challenges faced by the participants.

    “The experimental settings used only one subject domain (i.e., HCI), which may not capture the full spectrum of challenges faced by the participants.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    VisText utilizes metadata information spread across multiple paragraphs and text sections to generate alt text, but it often overlooks key information from the visuals.

    “VisText [30] utilized metadata information spread across multiple paragraphs and text sections to generate alt text, which often overlooks key information from the visuals.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30]

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

  • May combine two claims.

    Interpretationown data

    Participants struggled to write alt text without assistance, indicating that interpreting HCI figures was cognitively demanding, showing that their current support is not sufficient to meet BLV students' alt text needs.

    “Participants struggled to write alt text without assistance, indicating that interpreting HCI figures was cognitively demanding, showing that their current support is not sufficient to meet BLV students' alt text needs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Previous research has mostly focused on readers' perspective or authors of papers, and little is known about alt text writing by non-subject experts such as DSO professionals.

    “While previous research has mostly focused on readers' perspective [18,24], or for authors of papers [13,27], little is known about the alt text writing by non-subject experts, such as DSO professionals.”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [18]

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

    • cites [24]

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

    • cites [13]

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

    • cites [27]

      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.

    Interpretationreasoning

    Trust in AI varied among participants, possibly due to inexperience in evaluating LLM output.

    “Trust in AI varied among participants, possibly due to inexperience in evaluating LLM output.”
    Quote in the paperNot a substantial claim

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

    Study design and conductcitation

    The study used a single LLM, and different models may elicit somewhat different results.

    “We also used a single LLM for user studies, and different models may elicit somewhat different results [4,17].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [4]

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

    • cites [17]

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

33 of 33

  • The paper relies on this without support.

    Scientific visuals require more contextual and domain-related interpretation to make them meaningful for their consumers.

    “Unlike general images, scientific visuals require more contextual and domain-related interpretation to make them meaningful for their consumers.”
1 of 1

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • End-users need to verify, revise, and effectively prompt LLMs for better suggestions, the investigation of which is unexplored.

    “Hence, end-users need to verify, revise, and effectively prompt LLMs for better suggestions, the investigation of which is unexplored [35].”

    From How University Disability Services Professionals Write Image Descriptions for HCI Figures Using Generative AI (2026) · cites this paper as [35]

    Sections 3.2–3.3 describe participants evaluating and refining AI-generated descriptions, editing answers, and struggling to write effective prompts. The paper calls AI assistance for non-expert accessibility professionals underexplored, but it investigates these interactions itself; describing their investigation as wholly unexplored overstates the source. Page numbers are not available in the supplied passages. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • These professionals often lack the tools, support, or training needed to consistently produce high-quality and pedagogically meaningful alt text.

    “Despite their crucial role in ensuring access, these professionals often lack the tools, support, or training needed to consistently produce high-quality and pedagogically meaningful alt text [35].”

    From How University Disability Services Professionals Write Image Descriptions for HCI Figures Using Generative AI (2026) · cites this paper as [35]

  • Relatively little research has examined how automated alt text generation tools might support non-expert writers in organizational roles, such as DSO professionals.

    “Despite increasing interest in automated alt text generation, relatively little research has examined how these tools might support non-expert writers in organizational roles, such as DSO professionals, who write alt text for others' academic work [35].”

    From How University Disability Services Professionals Write Image Descriptions for HCI Figures Using Generative AI (2026) · cites this paper as [35]

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentt(36) = -2.09reported p = .044 · recomputed p = .044

    “t(36) = -2.09,p = .044”

1 of 1

Self-check

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

Self-check results appear here once an independent check has confirmed them.

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. Yung-Chen Jen Chiu Hsiao-Ying Vicki Chang Ann Johnston Mauricio Nascimento James T Herbert and Xiaoyue Maggie Niu. 2019. Impact of Disability Services on Academic Achievement among College Students with Disabilities. Journal of Postsecondary Education and Disability 32 3 (2019) 227–245.
  2. OpenAI. 2025. ChatGPT by OpenAI. Retrieved August 1, 2025 from https://chat.openai.com
  3. Zoom. 2025. Zoom video conferencing. Retrieved August 1, 2025 from https://www.zoom.com/en/products/virtual-meetings/

Cited by

Sorted by citations

Authors
Iwamoto, Y., Raees, M., Heard, J., & Tigwell, G. W.
Affiliations
Rochester Institute of Technology (US)
ORCID
Venue
Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility
Publisher
ACM
Published
2025-10-22
Pages
1-6
DOI
10.1145/3663547.3759763
OpenAlex
W4415445019
Semantic Scholar
64f7da2012b04faa7a9146c85974d75c99ee30f4
Cited by
1
References
30
Citations by year
2026: 1
FWCI
1.44
Type
conference paper
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor, 365 d)
Locations
International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS) (publishedVersion)
Funders
National Science Foundation (DGE-2125362)
Topics
  • Social Sciences › Human Factors and Ergonomics › Digital Accessibility for Disabilities
  • Health Sciences › Health Professions › Occupational Therapy › Assistive Technology in Communication and Mobility
  • Physical Sciences › Computer Science › Information Systems › Artificial Intelligence in Education
Keywords
alt text, picture description, scientific figures, generative AI, accessibility, screen readers, AI-generated text quality, trust in AI, user confidence, domain knowledge

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.

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