Accessibility expert warns of AI 'bollocks' while highlighting benefits

Léonie Watson, Director at TetraLogical and Chair of the W3C Board of Directors, delivered a comprehensive analysis of artificial intelligence's impact on accessibility during FFConf 2024 on November 8, 2024. The presentation examined both transformative benefits and significant shortcomings of current AI implementations for users with disabilities.

Watson outlined several established AI applications that have provided genuine accessibility improvements. Speech recognition technology enables individuals with motor disabilities to interact with computers without traditional input methods like keyboards or mice. According to Watson, this technology has evolved from early tools like Dragon NaturallySpeaking to mainstream implementations in Siri and Alexa.

Object recognition capabilities deliver particular value for blind users. Microsoft's Seeing AI application allows Watson to identify currency denominations, text content, and environmental features through smartphone cameras. "It's really, really good at telling me what stuff is, which is quite frankly, pretty useful," Watson noted during her presentation.

Face recognition technology, popularized by the iPhone 10 in 2017, provides security benefits for users who struggle with touchscreen interactions. Character recognition AI converts printed text into digital format readable by screen readers, helping users access information from medicine labels to store pricing.

Pattern recognition systems, including spam filtering, indirectly support accessibility by reducing cognitive load for users with disabilities who find processing large amounts of information challenging.

Generative AI creates new accessibility opportunities

Generative AI tools demonstrate remarkable capabilities for image description. Watson's screen reader can request detailed image descriptions from ChatGPT or Claude directly within the interface. "The level of detail that you can get at now independently on your own is really quite something," she explained.

Video description features have emerged in recent weeks through iOS 18 beta releases. Real-time camera description functionality allows Watson to receive immediate audio descriptions of her environment, supporting independent navigation and daily activities.

Communication assistance tools like Easy Speak, developed by Pearl Hilbert, combine text generation with AI speech synthesis to help individuals with communication difficulties participate in conversations more effectively. Voice cloning technology enables users to preserve their speech patterns before losing verbal abilities due to conditions like ALS.

Sign language translation projects demonstrate AI's potential to bridge communication barriers between different sign languages and spoken languages, though these remain in experimental phases.

Image generation and overlays present significant problems

Watson identified substantial issues with AI image generation when depicting disability. Generated images often contain anatomical impossibilities and perpetuate stereotypes about disabled individuals. "These tools have literally been hoovering up everything humanity has had to say on pretty much every subject, including all the horrible opinions that humans tend to hold," Watson observed.

Accessibility overlay vendors, including AccessiBe and UserWay, received particular criticism for promising compliance they cannot deliver. These third-party JavaScript tools claim to provide website accessibility compliance for monthly fees but lack the technical capabilities to address fundamental accessibility barriers.

"These tools can't fix that, and they can't fix it simply because to fix it, what they have to do is work out which elements on a page have JavaScript event listeners, and the browser won't let you do that," Watson explained, referencing a common accessibility issue these tools cannot resolve.

Legal challenges have emerged against overlay vendors. A dermatology clinic in New York filed a class action lawsuit against AccessiBe in recent months after being sued for website inaccessibility despite using the overlay service.

Code generation requires careful verification

Testing of AI code generation tools revealed mixed results for accessibility compliance. Watson evaluated Gemini, ChatGPT, and Fix My Code for their ability to create accessible code and identify accessibility violations.

All three tools produced code with accessibility deficiencies. Gemini's code included ARIA attributes but lacked keyboard support. ChatGPT missed ARIA roles entirely. Fix My Code, despite claims of exclusive training on accessibility data, also failed to implement proper keyboard support.

"Used well and used wisely, they're actually quite helpful," Watson concluded about AI coding tools, emphasizing the necessity of verification through testing or trusted sources.

Hallucinations remain a persistent challenge

AI hallucination issues affect accessibility applications significantly. When Watson asked ChatGPT to describe items on her office shelf, the AI insisted a telephone was beige despite corrections and failed to identify a prominent pencil in the image. These inaccuracies demonstrate the risks of relying on AI descriptions for important decisions.

Watson maintains realistic expectations about AI reliability while continuing to use these tools. "I am smart enough to know that I should not be making any kind of big decisions on what I'm being told," she stated.

Agentic AI transforms web interaction possibilities

According to Watson's LinkedIn analysis published on August 8, 2025, agentic AI systems capable of autonomous action could fundamentally change web interaction patterns. Platforms like Innosearch demonstrate early implementations where AI agents can search products, filter results, and manage shopping baskets on behalf of users.

"Why build a website designed to be clicked, tapped, or key-pressed, if someone can ask an agentic AI to carry out an action on their behalf with no need for any of that messy physical interaction business," Watson questioned in her August 8, 2025 LinkedIn post.

This transformation particularly benefits users who depend on keyboards, assistive technologies, or have difficulty processing large amounts of visual information.

Industry implications for marketing and accessibility

The accessibility community's experience with AI reflects broader marketing industry trends. Recent analysis by McKinsey identifies agentic AI as a leading transformation technology for marketing organizations, while 28% of marketing leaders prioritize AI and machine learning implementation.

X's October 2024 launch of an AI-powered advertising platform demonstrates automated targeting capabilities that could improve accessibility by reducing manual interface complexity. However, Perplexity AI's vision for agents replacing human attention in advertising raises questions about accessibility in these new interaction models.

The Interactive Advertising Bureau reports that 86% of video ad buyers use or plan to implement generative AI, highlighting the technology's rapid adoption across marketing channels. This widespread implementation necessitates accessibility considerations in AI-generated advertising content.

Copyright and environmental concerns persist

Watson emphasized two critical issues throughout her presentation. AI tools have been trained on copyrighted content without permission or compensation, violating intellectual property rights. Additionally, these systems consume excessive electricity and water resources during training and operation.

"Whatever you think of the current generation of AI tooling, the truth is, they've all been trained on our data. They've used our information indiscriminately without permission, without respect for copyright law, and certainly with no respect for our individual privacies," Watson stated.

She advocated for government intervention through policy, law, and regulation to address these fundamental problems with current AI implementations.

Future accessibility assessment challenges

Organizations transitioning to agentic AI systems face new accessibility evaluation challenges. Traditional testing methods rely on consistent output for representative sampling, but AI systems generate different responses to identical prompts.

"How do we determine with reasonable confidence that the generated content is accessible?" Watson questioned in her LinkedIn analysis. "Anyone who's used the likes of Meta, Gemini, Claude, or ChatGPT for more than 10 minutes will know that you don't get the same response twice even when you repeat the exact same prompt."

Watson's company, TetraLogical, has been approached by organizations seeking guidance on accessibility implications as they pivot toward agentic web implementations.

The presentation concluded with Watson's acknowledgment that AI tools are becoming increasingly difficult to avoid while emphasizing the importance of understanding their limitations and verification requirements for accessibility applications.

Timeline

• November 8, 2024: Léonie Watson presents AI accessibility analysis at FFConf 2024
• August 8, 2025: Watson publishes LinkedIn analysis on agentic web implications for accessibility
• July 2025: McKinsey Technology Trends Outlook 2025 identifies agentic AI as frontier technology for marketing
• July 15, 2025: IAB releases report showing 86% of video ad buyers using AI for creative development
• January 21, 2025: Comscore reports 72% of marketers plan to increase programmatic advertising investment
• January 2, 2025: Survey reveals 28% of marketing leaders prioritize AI and machine learning implementation
• January 1, 2025: Perplexity AI founder outlines vision for AI agents replacing human attention in advertising
• December 8, 2024: Perplexity AI reveals advertising implementation with sponsored questions and side media
• October 23, 2024: X launches AI-powered advertising platform with automated targeting
• 2017: Apple introduces Face ID with iPhone 10, advancing biometric accessibility
• 2013: Microsoft develops Seeing AI application for object recognition

PPC Land explains

Artificial Intelligence (AI): Computer systems designed to perform tasks that typically require human intelligence, including pattern recognition, decision-making, and learning from experience. In accessibility contexts, AI encompasses both discriminative systems that classify existing data and generative systems that create new content based on training datasets.

Accessibility: The practice of designing digital products, services, and environments to be usable by people with disabilities. This includes considerations for visual, auditory, motor, and cognitive impairments, often implemented through assistive technologies and adherence to established guidelines like the Web Content Accessibility Guidelines.

Screen Reader: Assistive technology software that converts digital text and interface elements into synthesized speech or braille output, enabling blind and visually impaired users to navigate computers and mobile devices. Modern screen readers can integrate with AI services to provide enhanced functionality like image description.

Agentic AI: Artificial intelligence systems capable of autonomous planning and execution of complex tasks without constant human supervision. Unlike traditional AI that responds to specific prompts, agentic AI can independently manage workflows, make decisions, and adapt strategies based on real-time data and changing conditions.

Generative AI: Machine learning systems that create new content including text, images, audio, and video based on patterns learned from training data. These systems can produce human-like responses and creative content but are prone to hallucinations and may perpetuate biases present in their training datasets.

Accessibility Overlays: Third-party JavaScript tools that website owners install with claims of automatically fixing accessibility barriers for monthly fees. These solutions typically cannot address fundamental accessibility issues due to technical limitations and browser security restrictions, leading to legal challenges and industry criticism.

Object Recognition: Computer vision technology that identifies and classifies items within images or video streams. For accessibility applications, this enables blind users to understand visual content through audio descriptions of physical objects, text, faces, and environmental features captured by smartphone cameras.

Speech Recognition: Technology that converts spoken language into digital text or commands, enabling hands-free computer interaction. This proves particularly valuable for users with motor disabilities who cannot easily use traditional input methods like keyboards, mice, or touchscreens.

Voice Cloning: AI technology that replicates individual speech patterns and vocal characteristics from relatively small audio samples. This capability allows people who may lose their ability to speak due to medical conditions to preserve their unique voice for future communication through assistive devices.

Web Content Accessibility Guidelines (WCAG): International standards developed by the World Wide Web Consortium that define how to make web content accessible to people with disabilities. These guidelines serve as the technical foundation for accessibility compliance requirements in legislation like the Americans with Disabilities Act.

Summary of the article

Who: Léonie Watson, Director at TetraLogical and Chair of the W3C Board of Directors, alongside AI developers, accessibility overlay vendors, and marketing technology platforms.

What: Comprehensive analysis of artificial intelligence's impact on accessibility, covering beneficial applications like speech recognition and image description, problematic implementations in image generation and overlay tools, and emerging challenges with agentic AI systems.

When: Presentation delivered at FFConf 2024 on November 8, 2024, with LinkedIn analysis published August 8, 2025, and ongoing developments in AI accessibility applications throughout 2024-2025.

Where: Analysis spans global AI implementations affecting web accessibility, digital marketing platforms, and assistive technology applications across multiple jurisdictions and market segments.

Why: AI technologies require critical evaluation for accessibility benefits and risks as they become increasingly integrated into digital platforms, with particular urgency given widespread adoption in marketing and web development despite significant limitations in current implementations.