Artificial Intelligence (AI) is rapidly shaping digital experiences, including accessibility. This article explores both the opportunities and risks of AI in accessibility work.
Let’s start with the basics.
What is Artificial Intelligence (AI)?
AI refers to technologies that can perform tasks typically requiring human intelligence, like understanding language, recognizing images, or making decisions based on data.
One advantage of AI that is often referenced is its ability to automate human tasks and perform them much quicker than a human can. However, automation is just the result of the massive computing power available to us these days and it doesn’t represent or replace human intelligence.
Another advantage often referred to is AI’s ability to access and draw on massive datasets, including – in theory – all of the Worldwide Web. Again, however, just being able to access data and information on a huge scale does not bring about intelligence.
A third advantage of AI is supposed to be autonomy, AI’s ability to take very little information, infer what a user needs and then set about autonomously meeting those needs without the need for human intervention. This is about as AI gets to what could be called intelligence – but is it desirable?
Bear in mind that AI has been in use since the 1940s. It’s only in recent years that the technologies using AI have created a revolution based not only on speed, automation, autonomy, and large amounts of data, but also the introduction of user friendly interfaces that seem to mimic human interactions and algorithms that produce outputs that seem to resemble the kinds of choices and instructions that human intelligence can deliver.
AI Technologies
There are several AI technologies you’ll encounter in accessibility work. Often, they work together or are combined into specific tools that can enhance digital accessibility or present accessibility challenges.
- Machine Learning (ML) allows systems to improve over time. ML can learn user preferences and adjust content in real time.
- Natural Language Processing (NLP) helps computers understand and generate human language.
- Computer Vision (CV) enables systems to interpret images and visual content.
- Automatic Speech Recognition (ASR) converts speech into readable text, enabling voice-controlled, hands-free interaction with technology.
- Generative AI (GenAI) creates new content such as text, images, voice, video, and code by interacting with users in conversational exchanges.
- Generative UI (GenUI) generates user interfaces dynamically, customised to user preferences.
- Convolutional Neural Networks (CNN) analyse visual data and detects edges, shapes, and textures.
The application of these AI technologies often overlap and combine to undertake tasks having an impact on digital accessibility. You’re probably already using AI in accessibility without realizing it. Tools that generate alt text for images, create captions for videos, or scan websites for accessibility issues all rely on AI. AI technologies are also now commonly used in language translation, search, maps, and chatbots.
Limitations
These tools can save time and help scale accessibility efforts, but they have limitations.
Limitation Example 1: Overlays
Automated accessibility overlays are widgets placed on websites that claim to identify and remediate all accessibility issues without the site owner having to do much more than click a button.
They often claim to be “AI powered” and offer a solution to a website owner’s accessibility problems by using JavaScript to modify markup and code.
The reality is that these widgets can only address a certain number of accessibility issues and often cause greater problems for users, such as when they prevent assistive technology like screen readers operating properly.
Many website owners who have installed automated accessibility overlays have endured a litany of complaints from users with disabilities and some have become the subject of legal action.
A distinction should be made between automated accessibility overlays and widgets that are designed to allow a user with disabilities to customise their web experience to meet their specific needs. The latter provide at least some benefit to users by choice, while automated overlays provide little benefit to users or website owners.
Limitation Example 2: Code Suggestions
If you ask a GenAI tool like ChatGPT or Claude, “How can I make this <div> accessible?”, it will typically respond with a suggestion to use ARIA, one that is generally accurate as far as it goes.
What it won’t do is question the basic premise of the question – “Is a <div> the best element to use here? Is there a semantic element available that can fill the same role without having to use ARIA, that is natively recognised by assistive technology, and that avoids <div> soup?”
ChatGPT and other GenAI will try to tell you what it assumes you want to hear. Learning how to prompt GenAI is part of the issue – “What’s the best element to use here for accessibility?” – but it’s important to be aware that AI doesn’t truly understand context the way humans do.
It makes predictions based on patterns in data, which means it can misunderstand meaning, miss nuance, or produce incorrect results. That’s why it’s important to think of AI as an assistant, not an expert.
Let’s explore some of the pros and cons that AI brings to digital accessibility.
Bias
AI systems learn from data, and that data is not always inclusive.
If an AI system is trained on data that lacks diversity, it will reflect those gaps. This can lead to biased outcomes.
For example, speech recognition systems may struggle with non-standard accents or speech patterns. Image recognition tools may misinterpret assistive devices or fail to represent people with disabilities accurately.
Bias can also show up in language. AI generated descriptions may unintentionally reinforce stereotypes or exclude certain cultural perspectives.
These issues aren’t just technical, they directly impact accessibility. When AI doesn’t recognise or represent users accurately, it creates barriers instead of removing them.
AI bias against people with disabilities in digital accessibility manifests in many ways.
- High error rates or complete failure in voice recognition may be evident for people with non-standard speech patterns. This is common in people with various disabilities and impairments.
Stephen Hawking used a speech synthesizer that AI might have trouble understanding. - AI facial recognition may not recognise non-standard faces. AI vision recognition tools struggle with “non-standard” or diverse physical appearances. Again, this is common in people with various disabilities and impairments.
AI facial recognition may not recognize non-standard faces. - AI interaction that relies on touch interfaces is often inaccessible and unforgiving of people with mobility impairments and a lack of digital dexterity.
AI can make touch targets too small and place them too close together (Image: Apple.) - GenAI that responds to user prompts frequently – at this stage, perhaps even predominantly – ignores the lives of people with disability in its responses.
- AI trained on broad LLM largely treat people with disabilities as exceptions and don’t include, consider, or even recognise them unless specifically referenced.
This kind of bias has very serious implications when AI is used to generate models of products or services intended to meet the needs of the whole community.
The risk of AI bias is not limited to people with disability. It extends to anyone who does not fit the profile of “average”. Significantly, though, AI bias in digital accessibility has a disproportionate and often critical impact on people with disability.
Since bias in AI tools comes from the data on which it is trained, there are two basic ways to address this problem.
Diversify Training Data
One way of reducing bias is to ensure that AI systems are trained on diverse and inclusive data.
If people with disabilities are not fully and accurately represented on the web and the web as a whole is not an inclusive enough data source to avoid producing biased output, then AI tools must be instructed to compensate, to take particular note of data sources that are inclusive and representative of disability and that understand what accessibility is and how to implement it.
Limit Training Data
In the same way – only in reverse – a GenAI chat tool can be trained only on a specific set of data sources. For example, Vispero (where I work) has built a chat tool that refers only to our own Knowledge Base:
- the rules and solutions we use for accessibility testing and remediation
- articles we’ve written as guidance for our accessibility engineers
- our digital accessibility training courses
- public blog posts we’ve written
In this way, we can trust that the information and functionality our AI tools provide are reliable, authoritative, and inclusive.
Assistive Technology
Let’s look at one of the most exciting areas where AI is making a difference.
Assistive Technology (AT) is any item, piece of equipment, software, or product system used to increase, maintain, or improve the functional capabilities of individuals with any kind of disability. Specific AT is designed to counteract the negative impact of all sorts of impairments.
AT enables greater independence and safety in daily life, ranging from low-tech tools (e.g., magnifying glasses, walkers, and modified eating utensils) to high-tech solutions (e.g., screen readers, voice-controlled devices, and power wheelchairs).
AI has enormous potential to make AT more efficient, functional, and accurate.
Screen readers
AI can help screen readers interpret context, describe images more effectively, and present information in ways that are easier to understand.
For example, JAWS has used Computer Vision since 2019 to power Picture Smart, which analyses images, photos, graphs, and app screenshots to describe people, text, landscapes, and chart data.
Further, JAWS developers have used AI to create FS Companion, a custom-trained model using specialised Machine Learning algorithms to provide detailed and responsive step-by-step instructions for using JAWS.
JAWS Page Explorer leverages Large Language Models to analyse and summarise webpage structures.
The next step for JAWS is the introduction of JAWS AI Agent, which combines PictureSmart, Page Explorer, and FS Companion into a single, seamless, conversational interface. Instead of switching between tools, users can simply ask questions and get help through one unified assistant.
Speech recognition
AI improves speech recognition software by utilizing Deep Learning and Neural Networks to understand context, accent, and nuance better, dramatically increasing accuracy, reducing latency, and handling background noise.
It moves beyond simple word-matching to interpret conversational meaning, filler words, and speaker intent in real-time.
Image recognition
AI improves image recognition software by using DL, CV, and CNN to automatically detect, classify, and interpret visual data with high speed and accuracy.
By training on massive, labelled datasets, AI identifies intricate patterns, edges, and shapes, reducing reliance on manual feature extraction.
Image recognition tools can generate alt text automatically, helping to fill gaps where descriptions might otherwise be missing.
Captions and Transcripts
AI improves captioning and transcription by using ASR and NLP to generate fast, low-cost, and scalable text from audio.
It can boost accuracy to nearly 100% in optimal conditions, handles multilingual translation, and provides real-time captioning for live events.
When you see live captioning correct itself in real time, that’s AI at work, understanding context and editing its own output to correct mistakes.
Plain language
AI improves language simplification by leveraging LLM to instantly reduce vocabulary and sentence structure complexity while preserving the original meaning.
Unlike traditional rule-based methods, AI can understand context, enabling it to rephrase, summarise, or add explanations to make dense text accessible to diverse audiences, including learners, people with cognitive impairments, or general readers.
AI can process complex web content and present it in simplified language, making it easier for users with various types and levels of cognitive impairments to understand and interact with, for example, medical, legal, and financial content.
More
- AT adapts over time to individual user needs rather than offering one-size-fits-all solutions.
- Instead of waiting for reactive assistance, AI powered AT delivers support in real time.
- AI enables voice-controlled, gesture-based, or gaze-tracking AT interactions for those with limited mobility.
- AI can identify potential hazards and alert AT users to avoid accidents.
All of these advancements can significantly reduce barriers and increase independence for people with disabilities who use Assistive Technology.
As powerful as the effects of AI on AT are, they still require oversight to ensure accuracy and usability, but it’s fair to say AI can be a life-changing way to support human control of AT.
Reliability
All this sounds great, but let’s acknowledge a current reality: AI can be helpful, but it isn’t always reliable.
AI driven digital accessibility tools promise efficiency but can introduce significant reliability issues, often creating an illusion that can be more detrimental than no accessibility features at all.
Misinterpretation
AI tools can fail to understand and convey the real or full purpose of content, resulting in misleading text summaries or inaccurate automated alternative text for images.
Automated captions can misinterpret words, especially names, technical terms, jargon, or accents, leading to user confusion or misinformation.
Hallucinations
While conversational GenAI tends to present with great confidence and certainty, it can provide false information to users who rely on these tools to digest content.
This tends to happen when AI is trained on data that includes incorrect information – in which case it will relay the incorrect information – or lacks the required information – in which case it will create a plausible response and present it as fact.
Experts on a topic might pick up on these “hallucinations” but often these tools are being used in the first place by people who are not experts and don’t detect incorrect information.
Context
While AI powered image recognition is getting better at deducing and conveying not just the content of images but their purpose in digital content, they are still limited in understanding context.
Automated alt text, for example, might describe an image as “a person outdoors” when the image contains important details that are missing from that description.
AI generated text summaries may lack nuance, omit critical context, or oversimplify complex information.
Security
Automated AI tools like overlays may identify and fix minor errors, leading to the assumption that a website is fully standards conformant and thus compliant with legal requirements, while failing to address complex interactive issues like keyboard navigation.
This false sense of security can lead website owners into serious trouble when users are unable to access digital content and take legal action.
Code
AI tools often rely on digital content being coded and marked up correctly. When there are errors in code or a lack of semantics in markup, they struggle to convey proper content and functionality.
Rather than identifying and fixing these errors, some AI tools will misinform users, while some others “give up” and become unusable. This demonstrates that foundational accessibility principles still need to be implemented for AI to work properly.
Ableism
We’ve already mentioned that AI draws on large datasets that often don’t include sufficient representation of people with disabilities.
This lack of awareness can make AI tools unreliable for users with disabilities, particularly those with complex needs. Not only do they perpetuate ableist bias, but they lack an ability to apply human judgement based on an understanding of real needs when trying to resolve digital accessibility issues.
The reliability of AI matters in digital accessibility. AI users must have confidence that content is conveyed correctly and completely, and that full digital functionality is available regardless of any disability they may have.
That’s why human validation is essential. AI can speed up the process, but it should never be the final authority. Always review AI generated content, especially when it impacts digital accessibility.
That seems like a good place to leave Part I of this two part series. Tomorrow, I’ll publish Part II.