Artificial Intelligence (AI) and Assistive Technology (AT) Users

By Nelson Tang

Whether or not artificial intelligence is a positive or negative thing for humanity can be debated endlessly back and forth between both sides of the argument. However, it can not be denied that AI has increasingly become incorporated into everyday life and every level of society from the workplace to the home. With each day that passes, AI continues to grow and learn from what they are exposed to by the developers whether for the better or worse. One of the most common forms of everyday AI usage is generative AI tools and platforms such as ChatGPT, which saw a boom in popularity following the 2020 pandemic. Since then, it has paved the way for other platforms like it and the increased use of OpenAi to develop things.

ChatGPT and like many of its counterparts primarily serve a tool/platform for people to gather information online whether it be for research purposes or other reasons. The appeal of this technology is that it can gather information at rates that would otherwise be humanly impossible if an individual were to manually research a topic. Other uses of generative AI include generating prompts, ideas, or artwork, which in itself is a whole other argument of human creativity versus artificial intelligence replacing it. Moreover, it depends on the intention or purpose of using these platforms and the people who are behind the development. Thus, even artificial intelligence has been proven to not always be 100% accurate and give out false information. Based on several studies done by different groups around the world, there is data that supports this claim. For example, an international study conducted by the European Union and British Broadcasting Channel (BBC) found that there was a 45% error regarding major AI tools. In another source, an article titled “AI Is Becoming a Go-To for Data Questions. How Reliable Are the Answers?” from Urban Institute, it involved an experiment containing 100 questions being asked. Results showed some error and even the AI giving answers to different questions or an answer that was not related to the prompt. The ones who conducted this experiment recorded their findings stating “From this exercise, we found four significant limitations in current LLM performance: Models explained concepts well but struggled to retrieve and use specific data. Incorrect information was difficult to detect, Models answered different questions than the ones asked.Pointing models to the right sources and tools didn’t improve results” (Tyagi et al. 2026).

If artificial intelligence is supposed to be smarter, then why does it still often make mistakes? One simple fact is that AI is reliant on preexisting information as it cannot truly create something on its own or at least technology has not progressed that far enough yet. In an article from Humboldt Institute for Internet  by Katharina Mosen, she explains the process. Mosen states “AI systems often rely on historical data shaped by existing power structures and social inequalities. These biases are adopted by the systems and incorporated into their predictions. As a result, old inequalities not only persist but are also projected into the future. This creates a vicious cycle where the past dictates the future, cementing or, in the worst case, exacerbating social inequalities. The use of AI-based predictive systems leads to self-fulfilling prophecies, in which certain groups are favoured or disadvantaged based on historical preferences or disadvantage.”(Mosen, 2024). In other words, AI learns from what data is given to them by the developers and engineers and the databases from the Internet. As a result, this is essentially how generative AI platforms generate things for users through pulling from all over the internet. Therefore, this can create another flaw in the design through bias or skewed data as mentioned.

Continuing on the topic of AI bias, it has unfortunately been proven true as it could have pulled from information that was discriminatory towards race,gender,religion,sexual orientation, or disabilities. That being said, AI technology has become more and more intertwined with assistive technology and those who use it. There has been some studies on AI being biased towards individuals living with IDD and the disabled community as a whole. One such case is a peer reviewed study conducted by the Special Olympics in a press release titled “Is AI Fair? New Evidence Suggests Bias Against People with Intellectual Disabilities Is Built In” In this article, they pulled from another study from Oregon State University examining how AI learning systems are just adding to the existing stereotypes towards those living with intellectual disabilities. Thus, creating technology that is ableist towards the users. Issues that arise include stereotype reinforcements such as portraying those with IDD as “symbols of inspiration” or one dimensional similar to how media portrays the “token minority” troupe. In other words, this portrays those with IDD as people without full autonomy. The study specifically looked at Chat AI and found that AI viewed those with intellectual or developmental disabilities as individuals who could be independent. “Among the core findings, five leading AI systems consistently generated stories portraying people with ID as more dependent, childlike, and in need of supervision than people without disabilities” (Stroppel, 2026). It concluded with all of this resulting from preexisting information and the path forward was designing AI to be more inclusive.

Another case study was conducted by the United Nations in an article titled “Building an accessible future for all: AI and the inclusion of Persons with Disabilities” Such takeaways include the risks involved such as the hiring process of those living with disabilities. There is software that is used to track eye or body movements and decisions are made based on this. This type of technology is commonly used in law enforcement or for anti-cheating for online exams. This technology is applicable towards virtual interviews, which creates the issue of discrimination and AI tracking to misinterpret body language and movement, Therefore, it can cause those with disabilities not being hired for a job. This is reinforced by algorithms and statistical data that is already discriminatory. UN special reporter Heba Hagrass highlighted how tackling the issues of technology and power could help people realize the rights of those with disabilities stating “One of the urgent matters is how artificial intelligence will change our world. The world for everybody. When computers came in and then the Internet, people did not notice that having computers and the Internet instead of creating a haven for everybody has built lots of obstacles and unbeatable barriers for many disabilities. It took us at least from 10 to 15 years to combat this. The problem is to convince people who are providing such programs to make them accessible” (“Building an accessible future for all”, 2024). Another way to combat the bias, is to make sure no one is left behind in the process of advancing technology.

What are ways to alleviate these issues and work towards a more progressive step in AI and inclusiveness for those living with IDD and assistive technology users. It got me thinking that the best course of action was to test out a simple thought exercise for several different AI platforms such as Claude, Gemini, ChaptGPT, and a few others. Furthermore, this experiment involved asking a series of questions pertaining to assistive technology users to each of the AIs to see any similarities and differences in the answers that they generated for me. The questions presented and answers given by each platform are shown below. How will each of the platforms compare in terms of how they answer the prompts? Will some give more detailed or accurate responses compared to the others?

ChatGpt:

How can AI coexist with existing assistive technologies without replacing them?	AI should be designed to work alongside existing assistive technologies, rather than replace them. For example, AI could integrate with screen readers, speech-to-text software, hearing aids, communication devices, or mobility technologies to make them more effective. Users should also be able to choose whether and how AI features are used. The goal should be to enhance the user's independence and existing tools, while keeping familiar assistive technologies available as a reliable backup.
How can AI improve the day to day lives of those who use assistive technologies?	AI can make everyday tasks more accessible and reduce the amount of effort required to complete them. For example, AI can provide image descriptions for people who are blind or have low vision, generate captions for people who are deaf or hard of hearing, help people communicate through speech or text, summarize complicated information, and assist with navigating digital environments. AI can also personalize assistance based on an individual's preferences, potentially making technology more flexible and easier to use.
Could there be any unintentional discrimination by AI towards those with disabilities and how could it be solved?	Yes. AI can unintentionally discriminate if it is trained on data that does not adequately represent people with disabilities. For example, speech-recognition systems may perform poorly with atypical speech patterns, facial-recognition systems may have difficulty recognizing people with certain disabilities, or automated hiring systems may unfairly disadvantage applicants whose disabilities affect how they communicate or complete assessments. These problems can be reduced by including people with disabilities in the development and testing process, using diverse training data, conducting accessibility and bias testing, and providing ways for users to report problems and request human review. Developers should also avoid assuming that one accessibility solution will work for everyone.
How can AI be helpful to assistive technology users in the workplace and the potential risks of AI for employees with disabilities?	AI can help employees with disabilities by automating repetitive tasks, transcribing meetings, generating captions, summarizing documents, organizing information, and providing alternative ways to interact with workplace software. It can potentially make workplaces more flexible and allow employees to perform tasks that might otherwise be difficult. However, there are risks. Employers could use AI to monitor employees excessively, evaluate productivity unfairly, or make hiring and promotion decisions based on biased algorithms. AI systems may also misunderstand the needs of employees with disabilities. Organizations should therefore make sure that AI is used to support employees rather than penalize them, and employees should have reasonable accommodations and access to human decision-making when automated systems affect important workplace decisions.
What are the privacy concerns that assistive technologies users should know about when using AI and how should sensitive data regarding the disabilities be protected/handled?	Assistive technologies can process highly personal information, including a person's voice, medical information, location, communication patterns, or information about their disability. Users should understand what information an AI system collects, where it is stored, who can access it, and whether it is used to train AI models. Sensitive information should be minimized and collected only when necessary. It should be protected through strong encryption, access controls, secure storage, and clear privacy policies. Users should also have meaningful choices about whether their data can be stored or used for other purposes. Organizations should treat disability-related info
What are some overlooked features when designing AI platforms and how can developers better involve assistive technology users in the testing phase?	Developers sometimes overlook things such as keyboard-only navigation, compatibility with screen readers, captions, alternative text, adjustable text sizes, color contrast, voice input, and the ability to customize how information is presented. They may also overlook cognitive accessibility, such as making interfaces predictable and avoiding unnecessarily complicated instructions.  Developers should involve people who use assistive technologies from the beginning, rather than waiting until the product is finished. This could include interviews, focus groups, accessibility audits, usability testing, and beta testing with people who have different disabilities and use different assistive technologies. Importantly, developers should compensate participants for their time and actually incorporate their feedback into the product.
What are the questions developers should be asking assistive technology users before development?	Developers should ask questions that focus on the user's actual experiences and needs, rather than assuming what people with disabilities want. Examples include: What assistive technologies do you currently use? What do you find difficult or frustrating about existing technology? What tasks would you most like AI to make easier? What features of your current assistive technology are essential to you? What would make you uncomfortable about using an AI-powered system? What information would you consider too private to share with an AI system? How much control would you want over the AI's decisions or suggestions? What should happen if the AI makes a mistake? What accessibility features are essential for you to use the platform independently? How do you prefer to receive information—text, audio, visual information, captions, or another format? Which assistive technologies and devices should the AI be compatible with? How would you want to report an accessibility problem? Would you be willing to test early versions of the product, and what would make that testing accessible to you?

Gemini:

How can AI coexist with existing assistive technologies without replacing them?	AI should function as an enhancer (an overlay or plugin API) rather than an overrider of established AT tools. Traditional assistive technologies (such as screen readers like NVDA/JAWS, refreshable Braille displays, switch access devices, and eye-trackers) are deeply fine-tuned to individual motor and sensory needs. API Interoperability & Clean DOMs: AI tools should operate through standardized accessibility APIs (like ARIA, Accessibility Insights, or OS-level accessibility trees) so that outputs seamlessly pass through the user’s existing screen reader or display hardware rather than injecting custom, non-standard UI layers. Respect User Preferences: Automated AI accessibility tools (e.g., overlay widgets) often break custom contrast, magnification, or key-binding settings if they force top-level script overrides. Coexistence requires AI to respect pre-existing OS and browser preferences. Modular Augmentation: Rather than building a "new screen reader," AI should act as a background service—processing raw visual/audio input into structured, semantic metadata that existing AT tools can announce or translate.
How can AI improve the day to day lives of those who use assistive technologies?	AI bridges the gap between static assistive software and unstructured, non-accessible real-world environments: Advanced Visual Recognition: AI models (e.g., Be My AI, Seeing AI) describe photo layouts, scan complex physical documents, interpret handwritten notes, and navigate urban surroundings in natural language. Contextual Speech & Communication: For non-verbal individuals or those using Augmentative and Alternative Communication (AAC) devices, AI can predict context-aware phrases, synthesize natural voice tones, or reconstruct dysarthric speech into clear audio in real time. Automated Content Refactoring: AI can convert dense, jargon-heavy text into Plain Language (simplifying cognitive load), dynamically auto-generate alt text for unlabeled web graphics, or auto-caption video content with high accuracy. Predictive Task Automation: Intelligent agents can execute complex multi-step digital tasks (e.g., booking flights, navigating cluttered web forms) via simple voice commands or single-switch triggers.
Could there be any unintentional discrimination by AI towards those with disabilities and how could it be solved?	Yes, AI systems frequently introduce algorithmic bias against individuals with disabilities, primarily due to non-representative training data. Examples of Bias Atypical Speech Patterns: Voice-recognition AI trained exclusively on neurotypical, fluent speech often fails to process speech from individuals with dysarthria, stutters, or accents. Facial Recognition & Emotion Detection: AI used in automated interviews often flags non-standard eye contact, facial ticks, or atypical gestures (common in neurodivergent or mobility-impaired individuals) as "lack of engagement" or "dishonesty". Screening Out Edge Cases: AI hiring algorithms trained on historical "successful employee profiles" may automatically penalize resume gaps caused by medical treatment or penalize alternative educational/career paths. Solutions: Inclusive Datasets: Actively incorporate diverse training data, including varied vocal patterns, motor inputs, and neurodivergent behaviors. Algorithmic Audits: Conduct regular third-party bias audits specifically focused on disability and intersectional demographics before deployment. Human-in-the-Loop Override: Ensure high-stakes decisions (e.g., hiring, medical access, financial approvals) always have a human review path and explicit alternative application formats.
How can AI be helpful to assistive technology users in the workplace and the potential risks of AI for employees with disabilities?	Workplace Benefits: Meeting Comprehension: Real-time AI transcription, speaker identification, and dynamic summary notes reduce cognitive strain for deaf, hard of hearing, and neurodivergent employees. Executive Function Support: AI tools assist with task breakdown, drafting routine emails, summarizing long documents, and managing calendars. Communication Translation: Translates alternative input or non-standard syntax into polished workplace communication without changing the user's core meaning. Potential Risks Inaccessible AI Tools: Companies adopting internal AI portals (e.g., proprietary chatbots, reporting platforms) without WCAG compliance lock out screen reader and switch users. Performance Monitoring Bias: AI productivity trackers measuring keystrokes, eye movements, or mouse clicks unfairly penalize workers who use alternate input methods or take structured medical breaks. Automated HR Discrimination: AI resume screeners or video interview analyzers filtering out qualified disabled candidates before a human recruiter ever sees them.
What are the privacy concerns that assistive technologies users should know about when using AI and how should sensitive data regarding the disabilities be protected/handled?	AI systems often require vast amounts of personal telemetry (biometrics, eye tracking, voice recordings, medical history) to function effectively as assistive tools. Key Privacy Concerns Unintended Disclosure: AI models learning from user prompts might inadvertently leak sensitive health or disability disclosures into public/shared training sets. Biometric Surveillance: Video and audio feeds used for gaze-tracking or voice control capture continuous personal data and environmental background details. Best Practices for Handling Data Zero Data Retention for Training: Prohibit vendor models from using health/accessibility telemetry or prompt inputs for public LLM fine-tuning. On-Device Processing (Edge AI): Whenever possible, run computer vision, voice recognition, and eye-tracking locally on the user's device rather than streaming raw data to cloud servers. Granular Consent & Anonymization: Implement clear opt-in models for data collection, stripping all Personally Identifiable Information (PII) and Protected Health Information (PHI) before any cloud transfer.
What are some overlooked features when designing AI platforms and how can developers better involve assistive technology users in the testing phase?	Frequently Overlooked Features Keyboard-Only & Switch Navigation for AI Interfaces: Designers build chat windows or visual nodes that look great but lack visible focus indicators or proper HTML landmarking. Rate Limits & Timeout Controls: AI response latency or strict idle timeouts disproportionately affect users who communicate at slower typing or reading speeds. Adjustable AI Verbosity: Screen reader users need conciseness; an overly conversational AI rambling for three paragraphs before reaching the point creates severe screen-reader bottlenecks. Better User Involvement in Testing Paid Co-Design & Testing: Hire disabled testers as paid subject-matter experts, not voluntary focus group participants. Shift Left (Early Testing): Involve AT users during wireframing and prototype phases rather than running an accessibility audit right before launch. Test with Native AT, Not Simulators: Automated accessibility checkers catch only ~30% of barriers. Real manual testing using actual hardware (screen readers, braille displays, eye-gaze systems) is essential.
What are the questions developers should be asking assistive technology users before development?	Before writing code, product teams should engage users with the following foundational questions: Workflow Integration: "How does your current assistive technology setup (hardware, software, shortcuts) handle this specific task today, and where does it fail?" Control & Autonomy: "What parts of this process do you want automated by AI, and which parts do you strictly want to retain manual control over?" Output Format Preferences: "How do you prefer AI-generated information to be presented (e.g., concise bullet points, structured headings, audio descriptions, raw plain text)?" Error Recovery: "When the AI makes a mistake or produces a hallucination, what is the fastest way for you to detect, correct, or bypass it using your input setup?" Data Comfort Boundaries: "What personal data or telemetry are you willing to share locally vs. with cloud models in exchange for higher personalization?"

Co-pilot:

How can AI coexist with existing assistive technologies without replacing them?	Takeaway: AI works best as an augmentation layer — not a replacement — by enhancing what assistive tools already do well. Interoperability — AI systems should integrate with screen readers, AAC devices, mobility aids, and browser accessibility APIs rather than bypass them. Modular design — AI features should be optional add‑ons, not mandatory replacements. User control — Assistive tech users must be able to turn AI features on/off or choose the level of automation. Respect for established workflows — Many users rely on long‑practiced routines; AI should support them rather than force new patterns.
How can AI improve the day to day lives of those who use assistive technologies?	Takeaway: AI can reduce cognitive load, automate repetitive tasks, and personalize support. Contextual assistance — AI can interpret context (e.g., reading complex PDFs aloud with structure preserved). Real‑time adaptation — Dynamic font resizing, captioning, gesture recognition, or predictive text tailored to the user’s patterns. Environmental interpretation — Computer vision can describe surroundings, read signs, or detect obstacles. Communication support — Predictive AAC suggestions, emotion‑aware phrasing, or faster symbol retrieval.
Could there be any unintentional discrimination by AI towards those with disabilities and how could it be solved?	Takeaway: Bias emerges when AI is trained on datasets that underrepresent disabled users or misinterpret disability‑related behavior. Dataset bias — Models may misclassify atypical speech, movement, or interaction patterns. Accessibility drift — AI updates sometimes break compatibility with assistive tools. Algorithmic exclusion — Systems may deny access or flag disabled users as “anomalies.” Solutions: Inclusive datasets — Collect data from diverse disability groups with consent. Accessibility audits — Regular testing with disabled users before deployment. Transparent models — Explainable AI helps detect discriminatory patterns.
How can AI be helpful to assistive technology users in the workplace and the potential risks of AI for employees with disabilities?	Takeaway: AI can level the playing field at work, but it can also introduce new vulnerabilities. Benefits Task automation — Summaries, scheduling, transcription, code generation. Adaptive interfaces — Personalized UI adjustments based on user needs. Communication support — Real‑time captioning, translation, or meeting summaries. Risks Surveillance creep — AI monitoring tools may misinterpret disability‑related behavior as low productivity. Disclosure pressure — AI systems may infer disability traits without consent. Automation bias — Managers may over‑trust AI assessments that disadvantage disabled employees.
What are the privacy concerns that assistive technologies users should know about when using AI and how should sensitive data regarding the disabilities be protected/handled?	Takeaway: Disability‑related data is among the most sensitive categories and must be protected with strict controls. Key concerns Inference risk — AI can infer disabilities from typing speed, voice patterns, or navigation behavior. Data sharing — Some apps share usage data with third parties. Biometric vulnerabilities — Eye‑tracking, voice recognition, or gait analysis can expose disability traits. Protection strategies Minimization — Collect only what is necessary. Encryption — End‑to‑end encryption for biometric and behavioral data. Consent granularity — Users choose exactly what data is shared and for what purpose. Local processing — On‑device AI avoids cloud exposure.
What are some overlooked features when designing AI platforms and how can developers better involve assistive technology users in the testing phase?	Takeaway: Developers often miss accessibility needs because they design for “average” users. Commonly overlooked features: Keyboard‑only navigation Screen reader labeling Caption accuracy Cognitive load reduction Customizable automation Better involvement Co‑design — Disabled users participate from the earliest design phase. Accessibility sprints — Dedicated testing cycles with disability groups. Real‑world scenario testing — Evaluate AI in noisy environments, low‑light conditions, or with multiple assistive tools active.
What are the questions developers should be asking assistive technology users before development?	Takeaway: Developers should ask questions that uncover real needs, constraints, and lived experiences. Essential questions: Daily challenges — “What tasks take the most effort or time?” Tool compatibility — “Which assistive technologies do you rely on?” Failure points — “Where do current tools break down?” Control preferences — “How much automation feels comfortable?” Privacy boundaries — “What data do you consider too sensitive to share?” Contextual needs — “Does your accessibility need change depending on environment?”

Claude:

How can AI coexist with existing assistive technologies without replacing them?	AI works best as a layer on top of existing tools rather than a wholesale substitute. Screen readers, switch access devices, eye-tracking systems, and AAC (augmentative and alternative communication) devices have been refined over decades in direct partnership with disabled users — that accumulated knowledge shouldn't be discarded. AI can coexist by: Enhancing existing tools (e.g., AI-powered image descriptions feeding into an existing screen reader like JAWS or VoiceOver, rather than replacing the screen reader itself) Filling gaps rather than overlapping — AI is often best used for tasks that were previously impossible (real-time captioning, object recognition) rather than tasks a mature technology already does well Preserving user choice and interoperability, so people aren't forced to abandon tools they've trained on for years, sometimes at great personal cost in time and effort Being opt-in and adjustable, not a default that overrides established workflows
How can AI improve the day to day lives of those who use assistive technologies?	Real-time transcription and captioning for Deaf and hard-of-hearing users in conversations, calls, and video content Image and scene description for blind and low-vision users (AI can describe photos, menus, product labels, surroundings) Predictive text and AI-assisted communication for AAC users, dramatically speeding up communication that would otherwise be slow letter-by-letter or symbol-by-symbol Voice control and natural language interfaces for people with limited mobility Personalized cognitive support — reminders, task breakdown, simplified summaries — for people with cognitive disabilities Navigation assistance (indoor/outdoor wayfinding) for blind or mobility-impaired users
Could there be any unintentional discrimination by AI towards those with disabilities and how could it be solved?	Yes, and this is a well-documented risk: Training data bias: Disabled people are underrepresented in datasets, so speech recognition may fail on atypical speech patterns, facial recognition may misread expressions from people with facial differences, and hiring algorithms may penalize resume gaps or nontraditional career paths common among disabled applicants Proxy discrimination: AI hiring or insurance tools might use variables that correlate with disability (gaps in employment, certain typing patterns, browsing behavior) even without explicitly considering disability status Accessibility-blind design: AI products tested only on non-disabled users may simply not function for many disabled users (e.g., a voice assistant that can't parse dysarthric speech) Solutions include: diverse and representative training data, mandatory bias audits specifically for disability, involving disabled people throughout development (not just as an afterthought), transparency about how AI decisions are made, and legal/regulatory accountability frameworks specifically addressing algorithmic disability discrimination.
How can AI be helpful to assistive technology users in the workplace and the potential risks of AI for employees with disabilities?	Benefits: AI note-takers and meeting summarizers reduce burden for people who struggle with real-time processing or memory. Automated captioning and translation in workplace communication tools AI-assisted writing tools help people with dyslexia, dysgraphia, or motor impairments produce polished work faster. Task and calendar management support for people with executive function challenges. Risks: AI monitoring/productivity tools (keystroke tracking, activity monitoring) can penalize workers whose disabilities affect pace or work patterns, even when output quality is equal. AI resume screening can filter out qualified disabled candidates due to biased pattern-matching. Over-reliance on AI accommodations by employers could become an excuse to avoid other necessary accommodations or make disabled employees seem "less independent" in performance reviews. Job displacement risk may disproportionately affect roles disabled workers have historically been able to access.
What are the privacy concerns that assistive technologies users should know about when using AI and how should sensitive data regarding the disabilities be protected/handled?	Key concerns: Assistive AI often requires highly personal data — speech patterns, biometric data (eye tracking, facial data), health information, location data, and behavioral patterns that can reveal a disability even when not explicitly labeled as such. Disability status is a protected and highly sensitive category; a data breach could expose someone's disability to employers, insurers, or others without consent. Many AI assistive tools rely on cloud processing, meaning intimate data (like AAC conversations) may pass through third-party servers. Data collected for accessibility purposes could later be repurposed (e.g., sold to advertisers or used for algorithmic profiling) without clear consent. How it should be handled: Strong data minimization — only collect what's essential. On-device processing where possible instead of cloud-based. Clear, accessible consent processes (not buried in inaccessible legal text) Encryption and strict access controls. Legal protections akin to HIPAA-level safeguards for disability-related AI data, with disabled users given control over deletion and portability of their own data.
What are some overlooked features when designing AI platforms and how can developers better involve assistive technology users in the testing phase?	Overlooked features often include: Compatibility with existing AT (many AI apps aren't tested with screen readers or switch devices at all) Customizable response speed/pacing for cognitive or processing differences Multiple input/output modalities (not everyone can type, speak, or see equally well) Error tolerance — AI systems that assume "typical" input patterns often fail ungracefully for atypical speech, movement, or typing Offline/low-bandwidth functionality, important for people who can't afford premium data plans or high-end devices Better involvement in testing: Paid, ongoing partnerships with disabled testers (not one-off, unpaid usability sessions) Testing across a range of disabilities and severities, not just one representative case Including AT users from the earliest design stages, not just post-launch QA Co-design sessions where disabled users help shape requirements, not just react to finished prototypes
What are the questions developers should be asking assistive technology users before development?	What assistive tools do you currently rely on, and how would this product need to interact with them? What tasks are hardest for you right now, and where does existing technology fall short? What would make you trust this product with sensitive personal or health-related data? What input/output methods do you actually use day to day (voice, switch, eye gaze, text, symbols)? What has "accessible" technology gotten wrong for you in the past? Would this feature save you time/effort, or create a new dependency that feels risky if it fails or is discontinued? How should errors or misunderstandings by the AI be communicated to you, and what's an acceptable failure mode? Who else needs to be consulted — caregivers, employers, disability organizations — and how do we make sure their input doesn't override your own preferences?

Perplexity: (note from it): AI should complement, not replace, established assistive technology. The best approach is to make AI optional, interoperable, controllable by the user, and designed with disabled people throughout development and testing.

How can AI coexist with existing assistive technologies without replacing them?	AI can work alongside tools such as screen readers, speech-to-text, captions, switch controls, braille displays, hearing aids, and alternative keyboards rather than trying to substitute for them. - Build AI features to follow accessibility standards and work with existing operating-system accessibility settings and assistive devices. - Keep familiar non-AI methods available. For example, a screen-reader user should still be able to navigate a website by headings and keyboard even if an AI “summarize page” button is offered. - Let users turn AI features on or off, correct them, and choose how much automation they want. - Avoid assuming AI output is always correct. It should provide assistance, not take control of important actions without confirmation. - Use open formats and APIs so an AI tool can share information with a person’s preferred assistive technology rather than locking them into a new system.
How can AI improve the day to day lives of those who use assistive technologies?	AI can reduce barriers and effort in everyday tasks when it is accurate, affordable, and user-controlled. - Communication: Real-time captions, transcription, text simplification, predictive text, and speech synthesis can support people with hearing, speech, language, or cognitive disabilities. - Vision access: AI image descriptions, object and text recognition, scene explanations, and navigation support may help blind and low-vision users interpret visual information. - Physical access: Voice control, gaze control, and adaptive interfaces can reduce the number of movements needed to use devices. - Organization: AI can help break tasks into steps, create reminders, summarize long documents, and tailor information density or reading level. - Personalization: A user might ask an AI assistant to present the same information in plain language, large text, audio, or a concise checklist. These functions need a clear way to report mistakes, because a wrong description, caption, or instruction can be more harmful than no automation.
Could there be any unintentional discrimination by AI towards those with disabilities and how could it be solved?	Yes. AI can unintentionally discriminate when its data, assumptions, or testing exclude disabled people. For example, a speech-recognition system trained mostly on typical speech may work poorly for people with speech disabilities; a video-interview system may penalize eye contact, facial expression, movement, or response time that is unrelated to job performance. Ways to reduce this risk include: - Include disabled people with varied access needs in design, training-data review, and evaluation. - Test performance separately across relevant disability-related use cases—not just average accuracy across all users. - Use human review and an appeal process for consequential decisions, such as hiring, benefits, education, health, or housing. - Give users ways to correct an AI’s output and explain why it may have failed. - Evaluate whether the system measures a real, necessary task rather than a proxy such as typing speed, voice pattern, webcam behavior, or mouse movement. - Continuously monitor errors after release, since models and user contexts change over time. NIST’s AI Risk Management Framework is intended to help organizations manage risks to people and society from AI systems, including harmful bias and other trustworthiness concerns.  W3C also emphasizes considering who may be at risk, examining historical bias in data, and actively working with affected communities to define and measure fairness. [nist](https://www.nist.gov/itl/ai-risk-management-framework)
How can AI be helpful to assistive technology users in the workplace and the potential risks of AI for employees with disabilities?	AI can make work more accessible by supporting real-time captions in meetings, document summaries, drafting assistance, transcription, task planning, communication aids, and flexible interfaces for people who use voice, keyboard-only input, screen readers, or alternative input devices. | Workplace use | Potential benefit | Main risk  | AI captions and transcripts | Makes spoken meetings more accessible | Incorrect captions can distort important information | | Writing and summarizing tools | Reduces cognitive, reading, or fatigue barriers | Confidential work information may be exposed | | Voice, chat, or alternative-input tools | Supports different communication and motor needs | Systems may not recognize atypical speech or input patterns | | Hiring and performance systems | Can organize applications and reduce repetitive work | May screen out qualified people because of disability-related traits | | Productivity monitoring | May identify workflow barriers if used supportively | Can unfairly treat slower input, breaks, or accommodation use as poor performance | In the United States, the DOJ and EEOC have warned that algorithmic employment tools can unlawfully disadvantage applicants and employees with disabilities. Key concerns include failure to offer reasonable accommodations, screening out qualified people, and improperly requesting disability or medical information. [justice](https://www.justice.gov/archives/opa/pr/justice-department-and-eeoc-warn-against-disability-discrimination) Employers should therefore provide an accessible alternative assessment, a clear accommodation-request process, meaningful human review, and an appeal route. They should not rely solely on AI to hire, promote, discipline, or dismiss someone.
What are the privacy concerns that assistive technologies users should know about when using AI and how should sensitive data regarding the disabilities be protected/handled?	AI systems may collect highly sensitive information: disability status, medical details, speech recordings, facial video, mobility patterns, eye-gaze data, assistive-device settings, support needs, and work-performance data. Even information that seems harmless—such as repeated use of captions or a voice-control tool—can sometimes reveal a disability indirectly. Good safeguards include: - Collect only data necessary for the specific feature; do not collect disability information merely because it might be useful later. - Obtain clear, accessible, informed consent before collecting or sharing sensitive data. - Explain what is collected, why, where it is stored, who can access it, and whether it is used to train AI models. - Use strong encryption, access controls, short retention periods, and secure deletion procedures. - Separate accommodation and medical information from managers’ routine employment records and from performance data where possible. - Give users practical controls to view, correct, download, delete, or opt out of secondary uses of their data. - Do not use disability-related data for profiling, advertising, insurance decisions, employment monitoring, or eligibility decisions without a lawful, necessary, clearly communicated basis.
What are some overlooked features when designing AI platforms and how can developers better involve assistive technology users in the testing phase?	Frequently overlooked features include: - Keyboard-only operation and clear visible focus. - Screen-reader-compatible labels, headings, status messages, and error messages. - Captions, transcripts, audio descriptions, and controls that do not depend on sound alone. - Adjustable timing, no forced time limits, and the ability to pause or repeat AI output. - Plain-language explanations, predictable layouts, low-distraction modes, and support for memory or attention differences. - User control over speech rate, voice, caption placement, contrast, text size, and level of detail. - Clear statements of uncertainty—such as “I may be wrong”—rather than presenting AI guesses as facts. - A usable fallback when an AI feature fails or declines a request. Developers should involve paid participants who actually use assistive technologies from the earliest research stage through prototype testing, pre-release testing, and ongoing monitoring. Recruit a diverse group rather than treating “disabled users” as one category: include people with visual, hearing, motor, speech, cognitive, learning, and multiple disabilities, and people using different devices and techniques. Automated accessibility checkers are useful, but they cannot replace testing with people; guidance on accessibility testing specifically recommends including users with different access needs and not relying solely on automated tools. [universaldesign](https://universaldesign.ie/communications-digital/web-and-mobile-accessibility/web-accessibility-techniques/content-providers-introduction-and-index/always-test-your-content-with-users-who-have-access-needs
What are the questions developers should be asking assistive technology users before development?	Developers should ask questions that identify goals and barriers without demanding unnecessary medical details. 1. What are you trying to accomplish, and what would success look like? 2. Which devices, assistive technologies, browsers, and operating-system accessibility settings do you use? 3. What parts of the current task are hardest, slowest, most tiring, or most error-prone? 4. What existing workarounds already help, and what do you want to keep unchanged? 5. Would AI assistance be useful here, or would a reliable non-AI improvement solve the problem better? 6. What level of automation feels comfortable, and which actions must always require your confirmation? 7. What types of mistakes would be inconvenient, harmful, embarrassing, unsafe, or costly? 8. How should the system show uncertainty, ask clarifying questions, and let you correct it? 9. What formats should information be available in—text, audio, braille, captions, plain language, icons, or another form? 10. What information should the system never collect, store, share, or use for training? 11. What accessible method should exist if the AI feature fails? 12. How would you prefer to give feedback, report accessibility barriers, or appeal an AI-generated decision? The central principle is: “Nothing about us without us.” People who use assistive technology should be compensated as expert contributors, not included only at the end as testers.

For the most part, the answers were similar and consistent enough with much of the responses complementing or in general agreement. Something that all of these platforms point at is that AI should aid or complement existing assistive technologies with no issue rather than replace them. As I assumed, the level of response and detail did differ between the AIs with some focusing on more specific aspects or giving a broader general overview in terms of answers. In terms of a balanced and straightforward response to each of the prompt questions, ChatGpt and Co-pilot fit this criteria. Out of these two, Co-pilot gave a little more insight regarding subjects such as user autonomy and personalization. Claude also emphasized these aspects taking a look at the user-centered perspective and experience. In terms of the most comprehensive and detailed responses, it came down to Perplexity and Gemini. Perplexity took a look at the standards delving into the areas of laws and privacy practices as well as very detailed questions developers should ask users. Gemini primarily focused heavily on the technical aspects such as software, accessibility testing, the technical risks, and on-device processing. In conclusion, the main takeaway of everything shows that inclusive design and input is needed in order for a more progressive AI as more advancements are being made in the field. 

Sources:

https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste https://www.urban.org/urban-wire/ai-becoming-go-data-questions-how-reliable-are-answers 

https://www.hiig.de/en/why-ai-is-currently-mainly-predicting-the-past/ 

https://www.specialolympics.org/about/press-releases/is-ai-fair-new-evidence-suggests-bias-against-people-with-intellectual-disabilities-is-built-in

https://health.oregonstate.edu/news-and-stories/2026-06/ai-fair-new-evidence-suggests-bias-against-people-intellectual

https://unric.org/en/building-an-accessible-future-for-all-ai-and-the-inclusion-of-persons-with-disabilities/