Ai for Accessibility: Tools for Disability

• A. To replace human caregivers entirely
• B. To help people with disabilities interact with technology and the world more independently ✓
• C. To make devices more expensive for premium users
• D. To eliminate the need for assistive technology

Why: AI accessibility tools aim to expand independence and participation for people with disabilities, such as through captioning, image description, and speech recognition, rather than replacing human support or assistive devices.

• A. Computer vision
• B. Automatic speech recognition ✓
• C. Reinforcement learning
• D. Robotic process automation

Why: Automatic speech recognition (ASR) converts spoken audio into text, enabling real-time captions that help deaf and hard-of-hearing users follow spoken content.

• A. By physically enlarging printed text
• B. By using computer vision to generate spoken or text descriptions of images and scenes ✓
• C. By changing the color contrast of a screen
• D. By sending images to a human reviewer only

Why: Computer vision models can identify objects, people, text, and scenes in images and produce natural-language descriptions that screen readers can speak aloud.

• A. AI tools always work perfectly for everyone
• B. Training data may underrepresent people with disabilities, leading to biased or less accurate performance ✓
• C. Accessibility features make software slower by default
• D. Disabled users never need privacy protections

Why: If training data lacks diverse examples, such as speech from people with atypical speech patterns or images reflecting varied disabilities, AI systems can perform poorly for those users, making inclusive data and testing essential.

• A. Predictive text and word completion that reduces the number of keystrokes needed ✓
• B. A printed Braille book
• C. A wheelchair ramp built into a building
• D. A manual transcription service staffed by volunteers

Why: Predictive text uses language models to suggest likely words, reducing typing effort for users with limited dexterity, which makes it a direct AI-based aid for motor disabilities.

• A. To replace all human interpreters in every setting
• B. To convert spoken language into text so people who are deaf or hard of hearing can read what is being said ✓
• C. To synthesize natural-sounding voices for people who cannot speak
• D. To translate written documents into braille automatically

Why: Speech recognition, also called automatic speech recognition (ASR), transcribes spoken words into written text in real time, allowing deaf and hard-of-hearing users to follow conversations, lectures, and meetings by reading. Voice synthesis and braille translation are different technologies.

• A. Speaker-independent recognition
• B. Speaker-dependent recognition ✓
• C. Continuous recognition
• D. Keyword spotting

Why: Speaker-dependent systems are trained on a specific individual's voice, which typically yields higher accuracy for that person. Speaker-independent systems are designed to work across many different voices without individual training.

• A. These users always speak too loudly for microphones to pick up
• B. Most models are trained on data from typical speech patterns, so atypical articulation, pacing, or pronunciation can be misrecognized ✓
• C. Speech recognition only works in English, which excludes these users
• D. Dysarthria prevents the vocal cords from producing any sound at all

Why: Mainstream ASR models are usually trained on large datasets of typical speech. Atypical speech patterns are underrepresented, which is a major accessibility challenge and a reason for ongoing efforts to collect more diverse training data.

• A. It eliminates the need for any teacher instruction
• B. It gives students who are deaf or hard of hearing near-immediate access to spoken content as text, supporting participation ✓
• C. It automatically grades student spoken responses
• D. It converts textbooks into audio for blind students

Why: Real-time captions let students follow lectures as they happen, reducing the lag of waiting for later transcripts. Captioning supports access but does not replace teaching, grading, or other tools.

• A. Speaking as quickly as possible to reduce processing time
• B. Using a good quality microphone placed consistently and reviewing or correcting the system's output when needed ✓
• C. Shouting every word to ensure the system hears it
• D. Turning off all punctuation settings to simplify recognition

Why: Clear audio input from a well-placed microphone significantly improves accuracy, and users can often train or correct the system. Shouting usually distorts audio and reduces accuracy.

• A. Semantic segmentation
• B. Image captioning ✓
• C. Optical flow estimation
• D. Depth estimation

Why: Image captioning generates a natural-language description of a scene, which can be read aloud by a screen reader, making it directly useful for blind users. Segmentation and depth estimation produce pixel-level or geometric outputs that are not immediately meaningful without further interpretation.

• A. To enhance the color contrast of images
• B. To convert printed or handwritten text in images into machine-readable text ✓
• C. To detect faces in a crowded environment
• D. To track the user's eye movements

Why: OCR identifies text characters within images and converts them into digital text that can be read aloud, enlarged, or translated. This is foundational for reading mail, menus, signs, and documents without sight.

• A. It makes the app run faster on mobile devices
• B. It helps users judge how much to trust a description, since models can be wrong ✓
• C. It is required by law for all AI products
• D. It reduces the file size of the output

Why: Computer vision models can make mistakes, sometimes confidently. Exposing confidence or uncertainty lets users decide when to verify information with another source, which is critical when decisions involve safety or money, such as reading medication labels.

• A. They cannot process any images in color
• B. They may perform unevenly across lighting conditions, camera angles, and image quality ✓
• C. They only work with images stored on a desktop computer
• D. They cannot generate any spoken output

Why: Vision models are sensitive to poor lighting, blur, motion, unusual angles, and cluttered backgrounds. Users often need to retake photos, so designers should provide guidance such as framing or lighting prompts.

• A. A single overall description of the room's mood
• B. Bounding boxes with labels and positions for items such as doors, people, or obstacles ✓
• C. A map of the room's wifi signal strength
• D. A list of colors used in the room's decoration

Why: Object detection identifies what objects are present and where they are located via bounding boxes. This spatial information can be translated into audio cues like 'chair about two meters ahead on your left,' supporting navigation and obstacle avoidance.

• A. To compress video files so they load faster on slow connections
• B. To convert spoken audio into on-screen text so people who are deaf or hard of hearing can access the content ✓
• C. To translate video into a different visual style for artistic effect
• D. To remove background noise from recordings before publication

Why: Automatic captioning uses speech recognition to transcribe spoken words into text displayed on screen, giving deaf and hard-of-hearing viewers access to dialogue and audio information.

• A. Optical character recognition
• B. Speech-to-text transcription ✓
• C. Text-to-speech synthesis
• D. Natural language generation

Why: Speech-to-text transcription, also called automatic speech recognition, converts spoken audio into written text. Text-to-speech does the reverse, and OCR reads text from images.

• A. Captions are always perfectly accurate, so this is not a concern
• B. Training data may underrepresent certain accents, dialects, or speech patterns, which can lead to errors that exclude some users ✓
• C. AI captions only work in silent films, which limits their use
• D. Captioning software cannot process audio longer than one minute

Why: If speech recognition models are trained mostly on certain accents or speaking styles, they perform worse on others. Accessibility depends on reliable output, so these gaps must be identified and corrected, often with human review.

• A. Teach hearing people to read lips
• B. Recognize signed gestures through cameras and translate them into text or speech, or generate sign language avatars from written text ✓
• C. Replace all human sign language interpreters in every setting
• D. Convert sign language into Morse code for telegraph transmission

Why: These tools aim to bridge communication between Deaf signers and non-signers by recognizing signs and producing text or speech, or by animating signing avatars from text. They are supportive tools rather than full replacements for qualified interpreters.

• A. Live captions are impossible to create with any technology
• B. Real-time captions often have more errors than post-produced captions because the system has no chance to review or correct its output ✓
• C. Live captions can only be displayed in black and white
• D. Live events never need captions because speech is always clear

Why: Real-time captioning must process speech as it happens, leaving no opportunity for editing, so errors and delays are more common than in pre-recorded content that can be reviewed and corrected before publication.

• A. To automatically translate text into another language
• B. To suggest likely next words or phrases so users can type with fewer keystrokes ✓
• C. To read aloud everything a user types on screen
• D. To correct spelling errors only after a document is finished

Why: Predictive text analyzes what a user has typed so far and offers likely word or phrase completions. This reduces the number of keystrokes or selections needed, which can be especially valuable for people with motor impairments or fatigue.

• A. Switch scanning
• B. Eye gaze input ✓
• C. Voice dictation
• D. Chorded keyboard

Why: Eye gaze input tracks where a user is looking on a screen and uses that position to select keys or items. It is often used by people with severe mobility impairments who cannot use their hands or voice reliably.

• A. By removing all punctuation automatically
• B. By learning the user's frequently used phrases and personal vocabulary to offer more relevant suggestions ✓
• C. By limiting the user to a fixed list of twenty words
• D. By requiring the user to speak every word aloud before it appears

Why: Personalized AI models can adapt to a user's own vocabulary, common phrases, and writing style over time. This makes suggestions more relevant and reduces the effort needed to communicate, which is the core benefit of personalization in augmentative communication.

• A. It always makes typing slower than manual entry
• B. Suggestions may not reflect the user's intended meaning, so users must check them carefully ✓
• C. It prevents the device from connecting to the internet
• D. It permanently deletes the user's previous messages

Why: Predictive suggestions can be wrong or steer a sentence toward an unintended meaning. Users, and sometimes their communication partners, need to review suggestions so the final message accurately reflects what the person wants to say.

• A. Speaking commands to change the screen brightness
• B. Using one or two switches to highlight options in sequence and select the desired item when it is highlighted ✓
• C. Typing with both hands on a full-sized keyboard
• D. Moving a cursor freely with a mouse

Why: In switch scanning, the system automatically highlights items such as letters or words one at a time. The user presses a switch to select the highlighted item. This lets people with very limited movement access text entry using just one or two reliable movements.

• A. Optimizing AI only for the most common user patterns to maximize average performance
• B. Ensuring the system works effectively for people with diverse abilities, backgrounds, and needs from the start of design ✓
• C. Adding accessibility features only after the product has been launched and tested by the majority
• D. Limiting the system's functionality to reduce the number of edge cases developers must handle

Why: Inclusive design means considering diverse users from the beginning, rather than treating accessibility as an afterthought or optimizing only for majority users.

• A. It reduces the cost of collecting data from real users
• B. Models trained only on typical speech, images, or movement patterns may perform poorly for users with atypical speech, visual impairments, or different motor patterns ✓
• C. Diverse data makes models run faster on mobile devices
• D. It is only needed for text-based applications and not for sensory or motor tools

Why: If training data underrepresents people with disabilities, such as speech with dysarthria or images from blind users, the system will be less accurate and less reliable for those users.

• A. Testing the system only with developers and internal staff who are familiar with the technology
• B. Involving people with disabilities as co-designers and testers throughout the development process ✓
• C. Relying on automated accessibility checkers as the only form of evaluation
• D. Assuming user needs based on what designers believe is helpful

Why: Participatory design with people with disabilities surfaces real needs and barriers that designers and automated tools often miss.

• A. It will always be more expensive than multimodal systems
• B. Users who cannot use that modality, such as people who are deaf or have speech impairments, may be excluded ✓
• C. It will automatically violate privacy laws
• D. It cannot be used on any device with a screen

Why: Offering multiple ways to interact, such as voice, text, touch, and visual output, helps ensure users with different abilities can access the system.

• A. Transparency is only a legal formality with no practical benefit for users
• B. Users need to understand, question, and correct system outputs, especially when errors disproportionately affect people with disabilities ✓
• C. Control features are needed only when the system is running in a laboratory setting
• D. Explanations slow down the system and should always be removed for efficiency

Why: Clear explanations and user control allow people to detect and correct errors, which matters greatly when a system's mistakes can disproportionately harm users with disabilities.

• A. The tool may run slightly slower on older devices
• B. Voice recordings may contain sensitive personal information that could be stored or shared without the user's clear consent ✓
• C. The tool might use too much battery power
• D. The interface may be difficult to customize for different font sizes

Why: Voice data can reveal identity, health conditions, location, and private conversations. Users need to know whether recordings are stored, for how long, who can access them, and whether they are used to train models. Clear consent and data minimization are key ethical safeguards.

• A. Latency in cloud processing
• B. Encryption failure
• C. Algorithmic bias arising from unrepresentative training data ✓
• D. Normal variation in camera quality

Why: When training datasets underrepresent certain groups, models learn patterns that work better for overrepresented groups. This algorithmic bias produces unequal accuracy and can cause real harm, such as misdescribing people or situations to users who depend on the tool.

• A. Assuming consent is implied because the user installed the app
• B. Providing the disclosure only in dense legal terms the user may not understand
• C. Explaining in accessible language what data is collected, who sees it, and how the user can change or withdraw permissions ✓
• D. Obtaining consent only once from a caregiver and never from the user

Why: Informed consent requires that information be understandable and that the person can act on it. Accessible explanations, ongoing options to adjust or revoke permissions, and involvement of the user (with support where needed) respect autonomy. Assuming consent, or bypassing the user, undermines their rights.

• A. It reduces the cost of marketing the product
• B. It allows companies to avoid legal responsibility entirely
• C. Lived experience reveals failure modes and needs that developers without disabilities may overlook, reducing bias and improving usefulness ✓
• D. It is only necessary for tools intended for physical disabilities

Why: Developers often cannot anticipate how a tool fails in real use. People with disabilities can identify biased outputs, unsafe defaults, confusing workflows, and missing features. Participatory design is a core ethical practice that improves both fairness and effectiveness.

• A. Transparency and the right of families to know how decisions about a child are made ✓
• B. The principle of maximizing software speed
• C. The requirement to use the latest hardware
• D. The obligation to publish open-source code

Why: Using AI to make inferences about a child's disability or learning needs without disclosure denies families the chance to understand, question, or contest the process. Transparency and accountability are central ethical requirements, especially when decisions affect a minor's education and labeling.