In Part I, we set the scene for the impact AI is having on digital accessibility. We defined what AI is, we explored some AI technologies used in digital accessibility, we looked at two examples of limitations of AI in digital accessibility, explored the impact of AI bias, how AI can enhance Assistive Technology, and how reliable AI is.
Let’s continue by exploring some more pros and cons of using AI in digital accessibility.
Shifting Left
Here’s another upside: AI can bring accessibility considerations earlier into the product development process.
Traditionally, accessibility has often been addressed late in the development process, sometimes just before release. This approach can be costly and inefficient.
AI can help to change that by enabling a “shift left” approach — bringing attention to accessibility into earlier stages of design and development.
The term comes from imagining product development as a left-to-right process from ideation to creation. Instead of considering accessibility as an afterthought at the end of the process, shifting left places accessibility at the beginning, and keeps attention on it throughout the product development process.
Designers
AI powered tools can analyse early designs, prototypes and wireframes for color contrast issues, suggest accessible layouts, and check that WCAG conformance requirements are met.
- Figma AI
- Figr (Figma)
- Stark (Figma, Adobe XD, Sketch)
- Evinced
Developers
AI tools can scan code as it’s written, pointing out accessibility errors and omissions, and recommending fixes before the product even starts to be built.
- GitHub Copilot
- Codex
- Evinced
- Axess AI (VS Code)
- Accessibility Assistant (WordPress)
Content Authors
AI tools can prompt for and even automatically generate alternative text for images, create captions and transcripts for audio and video content, check text for readability, and suggest content improvements to enhance accessibility.
- Grammarly
- Hemingway
- ReadEasy
- Descript
- Silktide
- Captions
- Otter
All of these tools should be considered supplementary to the core activities of web creators. The point of using them is to “nudge” web creators to consider accessibility early in the product development.
None of them are in themselves complete solutions to ensuring accessibility, but they are all very good for prompting web creators to treat accessibility as a primary, fundamental concern.
Most of these AI tools also support continuous monitoring, identifying issues as products evolve.
New Barriers
We mentioned earlier that while AI can help remove barriers to digital accessibility, it can also introduce new ones. Let’s look at that more closely.
Much of the AI powered functionality used in digital content comes in self-contained or third-party widgets and tools. Web creators and owners are then dependent on outside agents being aware of accessibility and lack the ability to modify the tools to take accessibility into consideration.
Website owners should be aware that relying on third-party content and functionality does not excuse them from conforming to accessibility requirements.
Chatbots
Complex AI powered chatbot interfaces theoretically make it easier for users to obtain information and have questions answered without the need for human intervention.
However, chatbots may be difficult to navigate with a keyboard and therefore create barriers for people who can’t use a mouse, including screen reader users.
If interactions aren’t carefully and properly structured, users with cognitive disabilities may struggle to understand or control the conversation.
The end result may be that users with disabilities end up having have less access to support from chatbots rather than more.
Search
AI powered search tools use NLP to understand user intent rather than just keywords, theoretically offering superior search results and handling complex queries.
Again, however, keyboard access and search engine compatibility can be poor or lacking completely, often due to poor labelling, creating barriers for many users with disabilities.
Recommendations
Widgets that act as recommendation engines for ecommerce shopping sites are based on users’ previous choices but generally don’t take disability and accessibility needs into account.
On video streaming services, a user might regularly choose movies with captions or audio description yet be offered choices by recommendation engines that don’t offer these accessibility options.
Voice-only
AI powered voice-only systems in digital content can exclude users who can’t speak or who have speech impairments.
When voice control functionality is implemented, it’s essential that alternatives to voice-only control are made available.
Language Complexity
While AI can simplify text, it can also default to complex language patterns that create barriers for people with cognitive disabilities or lower literacy levels.
Language complexity modification should be a user choice, allowing people to choose a reading level that suits their needs.
Unexpected Behavior
AI driven, dynamic content can change content and context unexpectedly, making it difficult for users who rely on consistent structures to navigate and operate digital content.
Consistency in the presentation of digital content is a key need for people with various kinds of disabilities and accessibility needs, including blind screen reader users, people with low vision, keyboard only users, and people with cognitive impairments.
Over-automation
Autonomous AI powered interfaces can reduce user control, making it harder for individuals to adjust settings or override AI decisions.
While the automation that AI tools offer can provide great benefits to all users, including people with disabilities, it can also negatively affect users who need to customize their online experiences.
AI should never automatically override assistive technology or accessibility settings.
AI Relies on Average
All of these barriers to accessibility come from AI depending on all users to fit an “average” model. AI is, so far, not good at considering the needs of people other than some mythical average.
The reality is that there are is no such thing as an average user, and AI tools must be configured to understand and meet the needs of all kinds of users, including those with disabilities.
Accessibility principles don’t stop at interfaces built with AI tools — they must apply to AI driven experiences as well.
Customization
Let’s dig deeper into one of AI’s greatest strengths: its ability to personalize user experiences.
Individual needs
One of the greatest potential benefits for the use of AI in making digital content accessible is its ability to let users customize and personalize their experiences.
Instead of designing for an “average” user, AI can adapt interfaces to meet individual needs, including those of people with disabilities and accessibility needs from simple to complex.
By analysing real-time data such as navigation paths, clicks, and scroll depth, AI such as GenUI can dynamically adapt the visual appearance, content structure, and page and screen layouts of digital content, and make recommendations to achieve individual user intent.
The potential for creating accessible, individualized experiences for users with disabilities is clear.
Agentic AI and Generative UI
Agentic AI refers to autonomous systems that can plan, reason, and take multi-step actions using tools to achieve goals without continuous human supervision.
Unlike GenAI, which creates content based on prompts, Agentic AI proactively makes decisions, adapts to new information, and collaborates across systems to complete complex workflows.
Assistive Technology will increasingly use an Agentic AI approach to provide users with a single tool built into their software that draws on combined AI functionality to make digital content accessible.
Agentic AI is the “brains” behind GenUI, a more general use AI tool that, as the name implies, generates user interfaces on the fly based on user intent.
When a user expresses a need to arrange a business trip, Agentic AI will search flights and hotel rooms, and check them against the user’s calendar.
GenUI will then present an interface that includes a flight selector and hotel options with relevant details that suit the user’s dates.
It’s not hard to see that such an approach can also take into account a user’s accessibility needs, both in the interface and the content it presents.
While they are designed to act autonomously, Agentic AI and GenUI must remain user-focused tools that understand and support user choices, especially in the context of digital accessibility.
Ethics
This is a big one. As AI becomes more integrated into daily life, digital content, and accessibility, ethical considerations become increasingly important.
We’ve already seen how AI can be used to include but also to exclude people with disabilities from full participation in digital life.
For this reason alone, ethical considerations are essential in AI and digital accessibility to prevent AI from perpetuating existing barriers or creating new ones.
The use of AI in digital accessibility and, indeed, in life in general must be accompanied by ethical frameworks to ensure fairness, transparency, and accountability.
Consent
Users need to understand when AI is being used in digital content, how and why.
They must be informed when AI is being used in order to be able to consent to its use, they must be provided with the means to express a lack of consent when AI negatively affects digital accessibility, and a lack of consent expressed by users must be acted on.
Control
Users must be provided with control over functionality that is AI powered. That might seem to defeat the advantage of autonomy but it’s necessary to ensure safety and lessen risk.
Some of this is already built into the Web Content Accessibility Guidelines (WCAG) in specific requirements such as allowing users to turn off, adjust, or extend time limits and to pause, stop or hide moving, blinking, scrolling or auto-updating content.
In fact, most of WCAG is about letting users control the way they consume digital content in accordance with their accessibility needs.
When AI is used in digital content, it must conform to these requirements.
Fairness
Users must have confidence that AI powered systems are designed to treat them equitably.
It’s the responsibility of web designers, developers and content authors to ensure that the AI they use in digital content does not discriminate against people with disabilities, neither in the information or functionality it presents, nor the way in which that content is presented.
AI systems must be trained on datasets that don’t reflect societal biases against people with disabilities and they must not perpetuate ableist language or discriminatory stereotypes.
Monitoring
The ethical use of AI in digital content requires ongoing monitoring.
Outputs should be reviewed by humans, and systems should be updated when issues are identified.
Just as WCAG has evolved over time to include new technologies such as mobile and to reflect changing priorities such as the inclusion of people with cognitive impairments when considering digital accessibility, so must it continue to evolve to include the implications of AI powered systems.
This is in turn will provide benchmarks that underpin disability discrimination legislation, and all governments and organizations must monitor standards conformance to ensure legal compliance.
Responsibility
Ultimately, ethical responsibility lies with the people designing and deploying AI — not with the technology itself. It’s not going too far to say that AI in digital accessibility can only be considered intelligent when it can understand, acknowledge, empathize with, and meet the needs of people with disabilities.
Scale
We seem to be swinging from positives to negatives and back again, so let’s consider another powerful AI advantage: scalability.
Digital accessibility work can be time-consuming, especially for large digital systems, in identifying issues, finding fixes, and applying remediation.
Consider that up until now, automated auditing tools are estimated to be able identify as few as 30% of accessibility errors in digital content. The rest have to be found and fixed manually, which takes time and expertise.
This is an area where AI can play a transformative role.
Instead of being limited in scope to a representative set of digital content, sometimes just a handful of web pages and workflows, AI powered tools can automatically interrogate thousands of web pages and app screens in very little time, identifying WCAG conformance failures and even going beyond WCAG to other standards.
Drawing on lightning fast access to massive datasets, it can identify and suggest appropriate remediation strategies in minutes.
As AI gets better and becomes more trustworthy and reliable, it can apply those fixes in real time, rewriting code and markup, adding alternative content, simplifying text, and making functionality more manageable for users with disabilities.
Website owners and digital practitioners can leverage ML, CV, and NLP to move from reactive, labor-intensive, periodic audits to proactive, continuous, and scalable accessibility conformance.
Benefits of scale
- Speed: AI can scan thousands of webpages or documents in minutes, a task that might take human testers days or weeks.
- Coverage: It’s estimated AI tools can identify 70–80% of common accessibility barriers, such as missing alternative text, improper heading structures, or poor color contrast.
- Monitoring: AI can replace one-off audits with round-the-clock monitoring, automatically checking newly published content for accessibility issues, at scale.
- Cost: AI tools can mitigate high costs associated with manual accessibility audits, lessen the need for expensive accessibility experts, leave owners less vulnerable to costly legal action, and eliminate post-production accessibility retrofitting.
Global Impact
At a global level, the scalability of AI powered digital tools has the potential to close the accessibility gap and make digital experiences more inclusive and accessible worldwide.
Not all of these AI powered capabilities are at a point yet where they can be trusted to operate completely automatically.
Many are improving rapidly, but scalability should not come at the expense of quality: human oversight remains essential.
Dependence
One last potential downside: as AI becomes more integrated into accessibility workflows, there’s a risk of us becoming too dependent on it.
We’ve covered some of the associated risks already, but it’s worth looking at them again in this light.
Skills Erosion
As digital designers, developers and content creators start to rely on AI tools, there’s a risk they will let their accessibility skills and expertise wither. There is already a tendency for web creators to pay insufficient attention to accessibility and dependence on AI may make this worse.
Digital accessibility professionals may fall into the same trap, letting their skills erode as they rely on AI automation for testing and remediation. They may then not have sufficient skills to properly assess AI output, nor know what to do when AI fails.
Overconfidence
Overlays that provide automated, “one-line-of-code” solutions often instill a false sense of security in website owners. Digital properties may appear to pass automated tests but still be unusable for people using assistive technologies, leading to possible legal action.
When web creators trust AI outputs without question, they may be relying on tools and systems that are inconsistent, inaccurate, or become unavailable.
Erasure
When AI systems completely replace rather than support human management of digital accessibility, the real lived experience of people with disabilities can be marginalised even more than it already is, leading to erasure of disabled voices and the principle of “Nothing about us without us”.
Quality
As AI vendors release new models and versions, they may pay less attention to digital accessibility issues, regarding them as edge cases that don’t need to be addressed. There are already reports that some newer AI tools produce lower quality results than earlier models when it comes to accessibility.
For example, while some AI tools are getting better at providing accurate alternative text for images by taking context into account, others regard merely putting anything into an `alt text` attribute as being sufficient to conform to standards. Automated testing will not pick this up without human attention.
Too much dependence on AI tools without monitoring and assessing automated output is a trap for website owners, web creators, digital accessibility professionals, and users with disabilities.
Digital accessibility work requires human judgment: an understanding of context, empathy, and user needs.
AI can support this work, but it cannot replace it.
Summary
When it comes to digital accessibility, AI is powerful but imperfect.
- AI can be biased against people with disabilities.
- AI can enhance Assistive Technology.
- AI is not always reliable.
- AI can flag accessibility issues earlier in development.
- AI can create new barriers to accessibility.
- AI can support the customization of digital experiences.
- AI must always be used in an ethical way.
- AI offers great scale advantages to digital accessibility.
- AI can create over-dependence on automation.
The bottom line is that AI in digital accessibility must always be used in combination with human judgement. Aim for a balanced use of AI to support web creators, digital accessibility practitioners, and users with disabilities, while maintaining strong human oversight and expertise.
Some resources
- Artificial Intelligence (AI) and Accessibility Research Symposium (W3C)
- AI and Accessibility: abdicating engagement? (UNESCO)
- Artificial Intelligence Has One Chance To Get Accessibility Right (Forbes)
- What AI Can Do For (and To) Existing Web Accessibility Tools (Vispero)
- Revolutionising Accessibility: The Role of AI in Assistive Technology (DLI)
- Digital Accessibility Ethics (book)
- Disabling Intelligences (book)