B Hari

October 31, 2026

The Extra Work of Asking AI for Help

AI visual assistance can expand independence. Its success should include how much effort people must spend making it useful—and whether employers keep removing the barriers around them.

On October 29, the University of Maryland highlighted research (https://today.umd.edu/apps-ai-eyes-for-blind-users-fall-short-umd-researchers-find) into blind people's use of Be My AI, an app that describes photographs and answers questions about them. The underlying study appeared earlier this year; this is fresh attention to an existing finding, not a new product test. Its central tension is consequential: assistance can provide information while leaving the user responsible for making that information usable. AI should reduce the work of securing access, while institutions continue to repair the barriers that create it.

The study (https://pmc.ncbi.nlm.nih.gov/articles/PMC13227612/) involved 11 blind participants and analysed 418 interactions across laboratory and everyday use. Short questions often produced much longer spoken answers. Participants developed ways to redirect the system towards what mattered to them. The researchers studied a fixed stage of the app's development, not every visual assistant or its present version. These observations reveal design problems; they do not establish that AI assistance generally fails.

The distinction matters because the benefit is substantial. Being able to inspect unfamiliar material without waiting for someone else can give a person more privacy, spontaneity and control. An imperfect tool may still improve on the alternatives available. Demanding flawless AI before allowing people to choose it would impose another barrier.

But availability and usable access are different achievements.

Imagine a blind employee receiving a scanned workplace form. An assistant could help identify its fields. If the employee must repeatedly photograph, question and check it, however, part of the institution's accessibility work has moved onto that employee. The organisation may see a completed form; the worker experiences the additional effort required to complete it. This is a hypothetical example, not a finding about any employer.

That effort belongs in the productivity calculation. Minutes spent clarifying an answer, recovering from an error or finding human help are costs, even when no invoice records them. Conversely, genuinely useful assistance can free attention for the work someone was hired to do. Employers should assess both effects before changing workloads or support budgets.

The commercial choices already make distribution relevant. Be My Eyes describes (https://support.bemyeyes.com/hc/en-us/articles/46666301222673-Free-desktop-app-vs-Be-My-Eyes-Workplace) a free desktop app and a paid Workplace service sold to organisations, with specialised tools and access to company-specific sighted agents. That is a description of its offering, not independent evidence of improved employment outcomes. Paid services can finance dependable support. They can also make the quality of assistance depend on an employer's purchasing power.

A worker at a well-funded company might gain both software and responsive help, while a freelancer must supply more troubleshooting time. That is a plausible distributional risk, not an established result of this product. The relevant question is whether a business model rewards reduced user effort or merely more institutional subscriptions.

India provides a useful practical counterpoint. At October's EMPOWER conference at IIT Delhi, Vision-Aid reported (https://visionaidindia.org/news/vision-aid-at-empower-2026-ai-and-digital-accessibility/) an AI assistant intended to help users operate interfaces that conventional screen readers struggle to interpret. Its account describes encouraging early testing, while acknowledging variable reliability and that formal accuracy and response-time measurements remain future priorities. This is promising development work, not proof of access at scale.

Such tools could help people work around today's inaccessible software. The danger would be allowing the workaround to become an excuse for leaving the software inaccessible.

The surrounding services matter too. WHO identifies (https://www.who.int/news-room/fact-sheets/detail/assistive-technology) funding, procurement and workforce capacity among barriers to assistive technology. UNDP's 2025 Human Development Report (https://www.undp.org/sites/g/files/zskgke326/files/2025-05/human_development_report_2025.pdf) explains why technologies designed in richer settings may fit poorly where devices, connectivity and accessible training are constrained. For Indian institutions, testing should therefore include the languages, equipment and support conditions people actually use.

Three futures are plausible. In a better one, disabled people are paid to shape products, employers repair inaccessible systems, and AI expands the tasks people can choose to undertake. In an uneven one, good assistance follows institutional budgets. In a harmful one, organisations withdraw human support because an app exists, then treat difficulties using it as an individual's deficiency. Purchasing and management decisions will help determine which emerges.

Questions for founders and investors

Can users control how much information they receive, and reach suitable human help when it is insufficient? Does evaluation measure completed tasks and total user effort? Who pays for training, support and repairs to the underlying service?

Funders should require answers from disabled users themselves. Employers should pay for access and learning time, retain needed support, and fix inaccessible forms and software. Public buyers can pool testing across languages and devices. The objective is a person with more practical freedom—and less work to do before ordinary work can begin.

Sources

- Zeraati and colleagues, CHI 2026 author manuscript (https://pmc.ncbi.nlm.nih.gov/articles/PMC13227612/).