Lost in Instructions: Study of Blind Users' Experiences with DIY Manuals and AI-Rewritten Instructions for Assembly, Operation, and Troubleshooting of Tangible Products
Monalika Padma Reddy, Aruna Balasubramanian, Jiawei Zhou, Xiaojun Bi, IV Ramakrishnan, Vikas Ashok · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790955
Summary
Blind users increasingly reach for AI tools like Be My AI (Be My Eyes' GPT-4-powered assistant) and ChatGPT to interpret the printed and PDF product manuals that accompany consumer goods - flat-pack furniture, alarm clocks, timers, ring lights, smart thermostats. Prior accessibility work has looked at AI tools in cooking, navigation, and general information access, but not at the harder case of product-manual-mediated do-it-yourself (DIY) tasks, which demand strict step ordering, precise part identification, spatial reasoning, and accurate manipulation where a single missed step can render the product unusable. Padma Reddy and colleagues fill this gap with two studies. An interview study with 15 blind participants (ages 33-73, all regular manual users and AI-tool users) probed how they obtain, interpret, and act on manuals for assembly, learning-to-operate, and troubleshooting tasks using critical-incident methodology and reflexive thematic analysis of 734 transcript pages. A complementary usability study with 7 blind participants observed think-aloud execution of four representative tasks - assembling a 15-piece desk organiser (A1) and a ring light (A2), and learning plus troubleshooting a lamp timer (T1) and an alarm clock (T2) - using only the manuals and AI tools they normally rely on, with NASA-TLX workload measures and coded failure factors. The design contribution is two parallel guideline tables: one for manufacturers redesigning manuals, one for AI-tool builders.
Key findings
Manuals remain the authoritative reference for blind DIY-ers (all 15 interviewees treated them as essential for assembly), but multi-column layouts, multilingual fold-outs, image-only PDFs, and pop-up ads in web manuals routinely break screen readers and force trial-and-error. AI tools are the primary workaround but do not close the gap. Across the four usability tasks, only 1 of 28 task attempts completed successfully without human intervention (overall 2.4% completion); step-level accuracy with AI-generated instructions stayed below 50%, and for Tasks A1, T1, and T2 it was 0%. NASA-TLX was very high (A1 M=86.81; T1 M=78.62; T2 M=74.00) - only A2, a simple four-single-action ring-light assembly, was low-workload. The authors identify four recurring AI failure modes: (1) vision-default bias - 80% of responses still used colours, labels, or figure references even after the user said 'I am blind'; (2) failure to maintain task-aware context - when users uploaded product photos to ground a question, AI defaulted to generic image captioning rather than answering the step-specific query; (3) inconsistent granularity - multi-action steps were rarely decomposed into single-action sub-steps, and terminology drifted across turns; (4) hallucinations in 64% of AI responses, spanning semantic, factual, content, and contextual types. An experimenter-delivered 'ideal' instruction pattern - product overview, prelude/orientation step, spatial-alignment cues, confirmation cues - reduced re-clarifications by about 80%. Paper manuals were preferred over digital, single-column layouts strongly preferred, and blind users routinely combined 3+ AI apps (Be My AI + Seeing AI + ChatGPT) to cross-check, stretching DIY tasks to 3-4x the time a sighted peer would take.
Relevance
For accessibility practitioners, procurement teams, and product-support designers, this paper is a stark warning that shipping a PDF manual plus 'Be My AI can read it' is not accessibility. AI-rewritten manuals are readable but not operationalisable: they lack tactile references, confirmation cues, before/during/after orientation framing, and atomic single-action steps. Manufacturers who want genuinely independent DIY completion by blind customers should ship single-column paper booklets, tactile progress checklists, NFC tags on parts that announce orientation, and QR codes to context-sensitive conversational Q&A. For AI-tool builders, the recommendations are concrete: layered macro/meso/micro instruction generation, user-relative reference frames, retrieval-augmented prompting, and chain-of-thought templates. Limitations: small samples (15 interviewed, 7 in usability), English-only, desktop/PDF-heavy OEM manuals, and the study was run in mid-2025 on GPT-4-era models - newer multimodal models may shift the error profile. Longitudinal work and non-English populations are open. The finding that even expert blind users with well-crafted 'don't use colours' prompts still hit below-50% step accuracy is the sharpest evidence yet that current multimodal AI is not a substitute for accessible source materials.
Tags: blind and low vision · generative AI · large language model · assistive technology · document accessibility · PDF accessibility · usability testing · hallucination · spatial reasoning · inclusive design
Standards referenced: PDF/UA · ISO 14289-1