Conversational AI is blind: The parallels between AI and accessibility challenges
Believe it or not, people with visual impairments share many of the same challenges as conversational AI when it comes to interpreting digital content.
Interestingly, these two seemingly disparate entities — people with visual impairments (or more broadly speaking, print disabilities*) and conversational AI — both navigate a world of digital content that is primarily designed for visual interaction. In this regard, conversational AI mirrors the blindness of visually impaired users, relying on structured, accessible content to understand and provide meaningful responses.
* A print disability is a difficulty or inability to read printed material due to a perceptual, physical, or visual disability. The reasons for print disability vary but may include vision impairment or blindness, dyslexia, physical dexterity problems such as multiple sclerosis, Parkinson's disease, arthritis, or paralysis.
A shared visual impediment: Understanding AI's ”blindness”
Despite its advanced capabilities, conversational AI struggles to comprehend visual content and tables. Much like people who are visually impaired and cannot perceive an image without descriptive text, AI models rely heavily on textual information to generate responses or summaries, leaving non-textual content like images, videos, and other multimedia unprocessed.
This is where the importance of accessibility becomes paramount. Features like alternative text (alt-text) for images, appropriately used headings in documents or websites, and structured bullet points provide a text-based interpretation of the content, enabling AI to — in essence — visualize the content in an understandable format.
The dependence on screen readers
Just as AI fumbles with non-textual content, people with visual impairment or other print disabilities rely on screen readers to browse the web and read documents. Screen readers convert text into speech, but when faced with poorly structured or inaccessible content, they often provide incomplete or disjointed information, leading to potential confusion or misunderstanding.
For example, a screen reader might start reading an invoice from the buyer's address, struggle to interpret a vendor's logo if alt-text is not provided, then fail to establish the relationship between item names, costs and the total invoice amount. Screen readers may also fail to extract pertinent information from an uncaptioned video, an infographic that contains text only as visual component, or to convey the structure and hierarchy of web content that lacks proper headings or lists. They also struggle with element names that have not been labelled with text or the “aria-labelledby” attribute — such as a button containing an SVG image or an icon font without any text.
The imperative of accessibility
Incorporating accessibility frameworks and features not only make content accessible for those with disabilities, but also enhance the capabilities of conversational AI.
When technical digital accessibility standards (like WCAG 2.1 - the Web Content Accessibility Guidelines, AAA or AA level) are applied effectively, documents and digital content have a cohesive semantic structure with proper use of headings and lists. This structure allows both screen readers and AI systems to discern the hierarchy and relationships within the information, leading to a more accurate interpretation of the content. The same principles apply to the use of alt-text or long descriptions for complex images, enabling both humans and AI to understand the content of images.
However, we still have a long way to go. In 2023, 98.1% of the top million website home pages had detectable WCAG 2 failures according to WebAIM, falling short of the technical accessibility standards set by the World Wide Web Consortium (W3C). This inaccessibility not only excludes people with visual impairments and print disabilities from participating fully in the digital space, but also limits the quality of responses from AI systems. While many might be quick to blame AI “hallucinations” for their inadequate responses, the real culprit could be the inaccessible content that they are trying to consume.
As we progress, it's crucial to champion accessibility by design in all digital content, rather than retrofitting solutions (by AI system providers) to address machine limitations.
Organizations need to embrace inclusive practices when creating digital content. In doing so, they will ensure that their employees with print disabilities can use assistive technologies and contribute at an equal level as colleagues that do not rely on such tools. Accessibility empowers these employees to be more productive and have a more fulfilling work life. Accessibility also enhances the quality of responses generated by the conversational AI systems available to all employees.
Inclusivity by default: The way forward
People with print disabilities and conversational AI both face considerable challenges when interpreting digital content. However, through by adopting standardized accessibility practices, these hurdles can be overcome.
A commitment to digital accessibility is not only a step towards inclusivity and a culture of belonging, but also a crucial stride towards improving the capabilities of AI systems. Therefore, let's work towards making accessibility the default way of working in all digital content creation. It is a true win-win for everyone involved.
Posted on: October 18, 2023
Neil Milliken
Vice President, Global Head of Accessibility & Digital Inclusion, Atos
Member, Atos Research Community
View details >
