The Evolution of AI in Digital Accessibility

Introduction

Digital Accessibility is the theory and practice of making digital services and products accessible to people with disabilities. This traditionally includes involvement in multiple stages of the process including design, implementation and distribution. Digital includes digital websites, documents, media, applications and increasingly immersive experiences. Whilst the requirement for accessibility is large, with roughly 16% of the global population, or 1.3 Billion people, having some kind of disability, the actual industry is small and requires a large breadth and depth of knowledge and consistent awareness of emerging technologies.

Overview of Artificial Intelligence (AI) and Machine Learning (ML)

Artificial Intelligence (AI) is the latest buzzword on the block. Simply it is a program or application that is able to do many tasks - it can be embedded into existing or new software, web applications and robotics to name just a few! Machine Learning (ML) is a method of achieving this by providing an algorithm with a large amount of data and allowing it to learn itself. There are many other methods of creating AI but ML is one of the most prevalent at the time. For the sake of this article the terms are used interchangeably.

Both of these things have been around for a long time but have seen massive advancements in recent years.

Role of AI/ML in Enhancing Accessibility

Accessibility has been slow to adopt AI and both users and specialists have often had increasingly negative experiences.. This has led to a further resistance amongst adoption whilst the rest of the world has exponentially adopted AI. This resistance is not without good reason - Overlays and Auto-Remediation tools that have prioritised ease of fix at the cost of accuracy, inclusion and human-centric innovation have been the most public implementation of AI so far.

Historical Perspective of AI/ML in Accessibility

Early Adoption of AI/ML in Accessibility

Whilst AI has become more prevalent in recent years, the ethos currently underpinning the development and adoption of previous technical solutions and tools has been a driving force behind the current usage of AI within accessibility.

Historically accessibility tools, specifically overlays and auto-remediation tools have been used since the 1990s with varying success and mostly failure. It is no surprise to those with exposure to these tools why they have been early adopters of AI.

Challenges and Criticisms

To understand the challenges and criticisms of existing solutions it’s first important to understand their purpose, intent and methodology. OverlayFactsheet.com will firstly provide a more comprehensive history of Overlays than any existing resource, as a crowd contributed and edited resource it is currently the leading resource for understanding overlays. Whilst it is not required reading to understand this article, it is highly recommended and will be referenced.

As explained Overlays are usually provided in the form of widgets and/or auto-remediation tools that manipulate the technology underneath, with the intent of fixing accessibility issues.

This is usually done by running a script on the page/application and attempting to fix issues or provide extra functionality (like font and colour manipulation) without providing lasting changes to the product/service underneath.

Issues with Overlays and Auto Remediation Tools

Reading OverlayFactsheet it’s instantly obvious the current shared opinion of these tools by accessibility specialists, product companies and most importantly, people with disabilities. They often fail to identify issues, they fail to remediate issues, they can be an active hindrance for users compared to no overlay at all, and they can often lead to false positives.

To be clear, considering Overlay companies can often sue accessibility professionals as seen in Adrian Roselli and French Case, these are not the individual opinions of a11y Boost or it's founders - they’re the outcomes of an EU Commission finding.

The EU commission found that overlays could solve no more than 30% of issues and are NOT a replacement for the consideration of technology. An echo of the estimation provided by specialists that 70%+ issues could not be discovered, yet alone rectified, using these tools. Previously, the most common example was it was impossible to rely on a computer to identify the contents of an image.

Over-reliance on Technology

The core of the issue here is not the use of technology to provide accessible outcomes. In essence Digital Accessibility is intrinsically about using technology to provide access to technology. The leading global guidelines for accessibility Web Content Accessibility Guidelines which are used to implement legislation in many countries, are intended to guide developers, designers and engineers to build accessible technologies.

People with disabilities rely on automated tools, successfully, all the time. Screen readers including Voiceover by iOS and voice interaction tools like VoiceAccess by Google have a mixture of commercial, paid, open-source and free equivalents on nearly every major platform and operating system.

The argument against current implementations such as overlays is the over-reliance of technology at the cost of universal design, expert implementation, user testing and tangible results. This cost is due to the perceived economic benefits of marginalising the human input.

Marginalization of Human Input

This perception of marginalising human input or human labour has been a long discussed element of accessibility and it’s been interesting see it become replicated as one of the greatest controversy points of the “Effective Accelerationism (e/acc) Vs Deceleration (decel)” debates driving the discussion of AI ethics today.

For those not in the know, a very simple explanation of a very complex ideological debate is e/accs belief that the benefits provided means that technological progress should be pursued relentlessly with decels holding the belief that the political, social and economic instabilities caused by technological progress should either hinder, slow or stop the development of future technology completely.

To say there is an ‘extreme’ implementation of ideology amongst the movement would be an understatement. Movement by e/accs is met immediately with fear mongering from decels, and any reasonable discussion about guardrails from decels is met immediately with pessimism and disbelief by e/accs.

Within accessibility there is often a push for innovation which when done without complete perfection is ostracised and permanently viewed with distrust. The claim made by Overlay companies that they could solve all issues AND remove the humans involved, has so far been completely unfounded and now anybody with exposure of the ‘early adopters’ greatly distrust future attempts at refining this technology.

We are fast approaching a point where selective use of AI/ML will be essential in digital accessibility. This is due to the increasing scale of services, products and society moving to digital environments and doing so in increasingly complex, innovative and fulfilling methods.

This selective use should never marginalise human input but rather focus on the augmentation and enhancement of human input - from both creators and users.

The Turnaround: Improvements in AI for Accessibility

Technological Advancements in AI/ML

The recent advances in machine learning have led to the realisation of more advanced tools and solutions in recent years. Whilst machine learning is older than the internet it had a resurgence in the 80s alongside the invention of the internet in neural networks. In the mid-to-late 2000s we saw the introduction of deep learning and a wider adoption of machine learning particularly in academia. But it wasn’t until the 2018 we saw significant progress in the artificial intelligence we’ve seen, ahem, emerge as Natural Language Processing (NLP).

NLP, once again simply put, is the ability for computers to learn how to read, write and communicate in plain language. This incredible jump led to OpenAI releasing ChatGPT3 - a model fine tuned for conversations by November 2022. Not even 18 months from writing this article.

Increased Understanding and Awareness of Accessibility Needs

Alongside the decades-long development of AI we’ve also seen a wider understanding and awareness of accessibility requirements in digital systems.

WCAG has progressed to 2.2 with WCAG3.0 under development and many countries including Australia, USA, UK, Canada and those inside the EU have implemented laws for minimum accessibility requirements both within the private marketplace and in relation to government tenders.

Simultaneously we’ve seen open-source contributions to assistive technology such as NVDA with a wide array of personal and corporate sponsors and contributors.

People with disabilities have increasing access to employment in many sectors, including software development directly but empowered by technology in many others. Increasingly private services and products, government services, education and health have enabled, or restricted, access through digital products.

Current AI Implementations in Accessibility

There are already companies implementing AI into their existing tools or workflows that begun as human-centric products and continue to augment or improve the current systems instead of replacing them. This includes screen readers, voice recognition systems and adaptive interfaces.

Improved Screen Readers

Screen readers are examples of the earliest forms of assistive technology, parsing text and mark up elements to provide visual information to a user through audio. Windows had already introduced Narrator by 2000 and Apple followed suit with VoiceOver first introduced on the Mac in 2005 becoming the first operating system native screen reader - requiring no extra installation or payment.

More recently they’ve incorporated early, optional, AI to allow automatic image recognition. Whilst it’s too early to see any significant progress in this area, it is encouraging to see Apple’s own guidelines providing descriptive alt text for their products has not changed.

Other popular screen readers also already use forms of machine learning algorithms to assist with accuracy particularly surrounding different languages and accents - although you’ll rarely see them discussing this publicly.

Enhanced Voice Recognition Systems

Voice recognition systems may be the most prevalent form of assistive technology, moving over to the coveted mainstream convenience product. They inherently rely on machine learning to learn what a user is saying so the system can implement the user’s instructions. Alexa, Google Assistant and NaturallyDragon are just a few, but the list is extensive. AI’s use and ability to easily understand natural language has seen companies such as Amazon move quickly to implement generative AI into their existing pipelines

Adaptive Interfaces

Adaptive interfaces have a near identical definition to overlays - a tool that manipulates the original input to provide contextual benefits to the end user. The clear difference is the intent and output of the tool, whilst overlays were promising to fix core accessibility issues, adaptive interfaces offer additional options whilst aiming to not interfere or Band-Aid the existing product. Adaptive interfaces intrinsically rely on the new abilities provided by AI.

Microsoft Seeing AI is a mobile application that allows users with low or no vision to receive audio descriptions of their surroundings. Whilst Google Live Caption allows for those with low or no hearing to receive visual descriptions of audio.

The Future: A Human-Centric Approach to AI in Accessibility

Impact on Various Disabilities

It’s clear that assistive technologies and those incorporating AI are here to stay and will become more widely adopted, particularly as they increasingly solve real user problems.

Visual Impairments

Those with visual impairments are able to rely increasingly on technology such as screen readers, tactile feedback devices, audial mapping and data to navigate the digital and built environment.

Hearing Impairments

Those with hearing impairments similarly can rely on technologies such as hearing aids, closed captions and visual alternatives to sounds both digitally and in their physical surroundings.

Mobility and Physical Disabilities

Adaptive devices, assistive robotics and much of the discussion points of Harriet McBryde Johnson’s future of Independent Care discussed in Unspeakable Conversations are becoming an increasing reality. With many specialised automated home and medical care devices already available, these products and services will see exponential growth not just due to AI but due to the robotics innovations required and driven by AI such as Figure.

Cognitive Disabilities

Those with cognitive disabilities may be most vastly impacted by future developments in AI than other disability group - this could either be positive or negative. The use of natural language could allow for enhancements in plain language tools such as HemingwayApp, we’ve seen AI implemented into major organisational tools like Notion and for years major businesses have implemented both light and ‘dark’ patterns that can benefit or hinder those with cognitive disabilities. Services and Products specifically created to benefit people with cognitive disabilities may impact areas including education, career development, and social experiences.

Legal and Ethical Considerations

The enforcement of dark patterns in unknown ways through AI are just one, insidious and purposeful negative possibility of AI. Dark patterns are not limited to just AI, neither are the inequalities caused by the mass adoption of inaccessible technology. The callous handling of private, sensitive or commercially protected data has been magnified by the impact of data bias and training methods but is once again, not limited to developments in AI. These issues though, particularly when targeted at a more vulnerable group, need to be considered.

Importance of Human Involvement in AI Development

This is where we move from history and possibilities, to a place closer to controversy. a11y Boost believes that AI will be fundamental to enabling equal access to society.

Whether it’s due to requiring access to AI or AI enabling access to other areas it is integral that accessibility is included and prioritised in the conversation surrounding AI. Simultaneously. we should not decry any attempts at progress because they are not perfect - no innovation is perfect. WCAG itself will soon reach its fifth major version, and is constantly being improved.

Just as we’ve seen digital accessibility take route in the past decades we’ll see an increasing shift of reliance on technology for more facets of our life. If this technology is not provided equally, people with disabilities could be further excluded from society.

To achieve this we must prioritise human-centric AI that does not replace, but augments, human input.

Balancing Technology and Human Touch

The Overlay Ship has sailed, firmly come under cannon fire and, in a cloud of smoke and fury, begun it’s slow descent to the bottom of the depths. Rising in its place are a fleet of AI-driven technologies both those that drive immeasurable benefit and surely many that are or will only be a net negative for people with disabilities.

We’ve seen automated captions, image descriptions and spatial audio instantly unlock tangible benefits. We’re seeing a rise in automated robotics that provide both cognitive and mobile solutions.

We’ve also seen a rise in rapidly built and essential, yet inaccessible websites, applications and media. We’ve seen entire regulated industries filled with misinformation, damaging systems and lack of expertise. Lastly, we’ve seen a wave of services and products specifically targeting people with neurodiversity.

All of these, whether good or bad, just slightly effective or incredibly powerful, are still in their infancy and have much, much more potential left. To ensure the ships with potential manage to sail safely to port and to circumnavigate the dangers of those ships with dangerous loads, humans must be in the lighthouse. We can not allow the lighthouse to guide autonomously.

Predictions and Trends for Human-Centric AI Solutions

Human-centric AI within accessibility over the next decade is both a continuation of current trends and likely new innovations society is yet to begin adopting.

More accessibility specialists will rely on a mixture of automated AI testing tools and their own knowledge and human testing abilities. As more countries increasingly regulate accessibility and the accessibility industry fails to expand its recruitment pipelines the current method of testing, auditing and developing accessible solutions must be augmented to keep up. 96% of the top 1,000,000 homepages are inaccessible. Even if the accessibility industry manages to 100x its resource count and impact overnight, it would still not solve for half of that subset within a subset.

Workflows must augment and to remain effective accessibility specialists should embrace and refine their adoption.

More users will rely on a mixture of existing traditional accessibility aids, convenience technologies and new AI innovations. As disability and accessibility awareness increase more people with disabilities are able to access aids and services. The stigma of accessing these services has also increasingly decreased as they are adopted by mainstream consumers.

As many populations globally face an aging population, more and more people will face difficulties due to permanent or temporary disabilities. Traditionally as a population ages they are less likely to adopt new and challenging technologies; the very same technologies that are becoming increasingly embedded in our economy, healthcare, education, politics and social currency. Natural language AI and adaptive interfaces will be integral.

Universal design is an essential ethos of accessibility, the correct human-centric use of AI does not and should not remove accessibility fundamentals, web guidelines and regulation. It should empower a diverse range of solutions for the diverse range of disabilities, including those currently unaccounted for.

Potential Challenges and Solutions

Accessibility Specialists have a fierce challenge ahead of them.

Internally there should be a commitment, in line with OverlayFactsheet, that we do not rely on automated accessibility tools that do not solve the underlying, or tangible, barriers people with disabilities face.

But, there also needs to be participation from Accessibility Specialists and people with disabilities in creating and using AI-led automated tools for auditing, remediation, and usability. Without specialists and disabled folk becoming involved in the conversation it will be implemented without us and the output will most likely not include us.

Simultaneously this is a call to those within the e/acc movement and those taking up the label techno-optimist or techno-capitalists to take the unprecedented opportunity - the first and most monumental societal wide adoption of innovation - in a world where accessibility awareness already exists, don’t add it in as an afterthought. Take the opportunity to not reinvent the bus, but instead listen to people with disabilities.

Prompt Engineering

Whilst the ethos of computer automation has not been widely positive within accessibility. We have a time-boxed opportunity to ensure AI is beneficial to every facet of accessibility and inclusion, we should not waste it.

By being deeply involved in a mutually beneficial partnership those developing AI should understand the need for inclusive design and testing with people with disabilities.

So a11y Boost is prompting you to reach out and get engaged.

If you are an accessibility specialist working, or looking to work on, AI tools for accessibility a11y Boost wants to hear from you.

If you are developing, or looking to develop, Artificial Intelligence and want to ensure it is accessible, a11y Boost needs to hear from you.