From Access to Agency: Rethinking Digital Inclusion for Persons with Disabilities in the Age of AI

A11ySIG / IGF DCAD GAAD Event - 21 May 2026

VIDEO | AUDIO | RECAP EN / ES / FR / AR / ZH / RU | ARCHIVE | PERMALINK

Internet Governance Forum Dynamic Coalition on Accessibility and Disability (IGF-DCAD) and Accessibility Special Interest Group (A11ySIG) Global Accessibility Awareness Day Event - 21 May 2026

Speakers: Dr. Vint Cerf - Chief Internet Evangelist, Google; Dr. Muhammad Shabbir - Co-coordinator, IGF-DCAD; Alessandro Ciceri; Charles Hall; Kamran Rafiq
Moderator: Giacomo Mazzone - Global Project Director, UNDRR; Judith Hellerstein - Co-coordinator, IGF-DCAD

Opening Remarks and Accessibility Commitments

Giacomo Mazzone opened the session by welcoming participants to the IGF-DCAD webinar focused on digital inclusion, disability, and AI. He immediately handed the floor to Dr. Muhammad Shabbir for introductory remarks.

Dr. Muhammad Shabbir framed the webinar around Global Accessibility Awareness Day (GAAD), noting that while the observance takes place annually on the third Thursday of May, accessibility work must continue year-round across digital and physical environments. He emphasized that Internet accessibility is not solely for persons with disabilities, but benefits everyone. He also acknowledged the availability of live captioning and sign language interpretation for the session, reinforcing the event’s commitment to accessibility in practice.

Shabbir paid tribute to Dr. Vint Cerf for his longstanding support of the Dynamic Coalition on Accessibility and Disability and thanked the organizing team for facilitating the event. He encouraged participants to actively contribute questions and discussion points throughout the session.

Fellowship Program for IGF 2026

Judith Hellerstein introduced the IGF-DCAD fellowship program that will support participation at the 2026 Internet Governance Forum in Nairobi, Kenya, scheduled for 14–18 December 2026. She explained that the fellowship program differs from many traditional support mechanisms because applicants define the accommodations they personally require rather than receiving a fixed package.

Examples included:

• Travel companions for visually impaired participants

• Shared accommodation support

• Other individualized accessibility requirements needed to fully participate in the IGF

Hellerstein stressed that accessibility support must be tailored to individual lived experiences rather than standardized. She also highlighted procedural requirements, including membership in the Dynamic Coalition and timely registration for the IGF itself.

Vint Cerf on Accessibility, AI, and Inclusive Design

Dr. Vint Cerf began by acknowledging the leadership of Giacomo Mazzone, Judith Hellerstein, and Muhammad Shabbir in advancing accessibility discussions within Internet governance spaces. He argued that accessibility remains neglected across much of the digital world, though not necessarily out of malice. Instead, he suggested many developers lack practical understanding of disability experiences and the wide range of accessibility requirements users may have.

Cerf used his own hearing impairment as an example, explaining that disabilities exist on broad spectrums. While hearing aids work effectively for him, others may require cochlear implants or additional assistive technologies. Similarly, visual impairments range from requiring magnification to needing full scene description and text-to-speech systems.

He identified two major technological developments that could transform accessibility:

• Artificial intelligence

• Widespread access to computing infrastructure

Cerf emphasized that AI is not a universal solution, but described recent advances in “vibe coding,” where users describe desired software behavior in natural language and AI systems generate the code. He revealed that Google internally had already moved from generating roughly 25 percent of software this way in the previous year to approximately 75 percent during the current year.

He proposed that accessibility specialists and persons with disabilities help train AI systems by documenting:

• Successful accessible interfaces

• Accessibility failures

• User experiences

• Examples of effective remediation

According to Cerf, these lived experiences could become training material for large language models specialized in accessibility-focused software development.

Cerf also highlighted broader structural challenges:

• Approximately 25 percent of the global population still lacks direct Internet access

• Reliable electricity remains foundational to digital inclusion

• Low Earth orbit satellite systems such as Starlink may help close connectivity gaps

• Infrastructure investment remains essential for meaningful accessibility

He stressed that accessibility cannot exist without robust underlying infrastructure.

Another major theme in Cerf’s remarks was the importance of feedback loops. He argued that accessible systems require mechanisms for users to easily report failures and barriers, and organizations must genuinely respond to those reports rather than ignoring them.

Muhammad Shabbir and the Concept of “Apomechanic Systems”

Muhammad Shabbir introduced the concept of the “Apomechanic System” or “Apomechanic Web,” a term he coined to describe systems that actively push users away because of inaccessible design.

He explained that society already has terminology for:

• People disconnected from the Internet

• People excluded due to economic barriers

However, he argued that another category exists:

• Users who possess connectivity and modern devices but remain excluded because systems themselves are inaccessible

The term “Apomechanic” derives from Greek and Latin roots meaning systems that “send users away.” He gave the example of unlabeled buttons that prevent screen-reader users from submitting forms or accessing services. In these cases, the system itself creates exclusion.

Shabbir contrasted this framework with earlier “medical model” approaches to disability that focused on “fixing” individuals. Drawing on the UN Convention on the Rights of Persons with Disabilities (CRPD), he argued that the responsibility lies instead with society and systems to remove barriers.

He emphasized that inaccessible systems can emerge anywhere in the world regardless of infrastructure quality, and therefore accessibility must be treated as a core design principle rather than an afterthought.

AI-Generated Accessibility Problems

Shabbir shared a practical example of using AI tools to build a personal website despite having no coding experience. He described generating a complete site in approximately two hours using AI prompts and his professional documents.

However, he discovered that while the AI system technically labeled graphics and buttons, the labels lacked meaningful contextual descriptions that human accessibility practitioners would normally provide. As a result:

• The website appeared superficially accessible

• Automated accessibility checks failed to detect deeper usability problems

• Human evaluation remained necessary

This reinforced his argument that lived experiences and human testing remain indispensable even as AI systems become more sophisticated.

AI Learning, Privacy, and Persistent Context

A major discussion topic centered on whether AI systems could eventually learn continuously from interactions rather than requiring periodic retraining. Cerf explained that current large language models simulate learning through massive context windows and temporary ingestion of web information, but do not truly retain knowledge after interactions conclude.

Participants raised concerns regarding:

• Privacy

• Storage of sensitive disability-related data

• Reliability of AI outputs

• Trust in AI-generated interpretations

Shabbir specifically described concerns about AI systems reading personal documents, financial information, or private conversations for accessibility purposes. He questioned how such data is stored and whether users could truly control it.

Cerf responded cautiously, acknowledging that these problems are unresolved. He proposed the possibility of maintaining private contextual data under user control, similar to browser cookies, where information would only be temporarily shared with AI systems during interactions and not permanently absorbed into training datasets.

Bias, Machine Unlearning, and Disability

Alessandro Ciceri introduced research from his thesis work on “machine unlearning,” which aims to remove specific information from AI models in ways compatible with privacy regulations such as the GDPR.

He explained that removing data points can unintentionally worsen existing biases within AI systems. In disability contexts, attempts to protect privacy may inadvertently reduce representation of marginalized groups within training data, thereby increasing discriminatory outcomes.

Cerf responded enthusiastically, describing the idea of user-controlled contextual information stores that could be selectively shared with AI agents during interaction without permanently modifying the underlying models.

Accessibility as a Societal Responsibility

Judith Hellerstein and Muhammad Shabbir both emphasized that inaccessible environments are societal failures rather than individual shortcomings. Hellerstein compared inaccessible digital systems to buildings without wheelchair access, arguing that exclusion often results from assumptions that “no disabled people are here anyway.”

Shabbir stressed:

• Disability can affect anyone at any stage of life

• Aging introduces new accessibility needs

• Accessibility should be integrated from the beginning rather than retrofitted later

• Retrofitting increases costs and complexity

He argued that accessible design ultimately benefits society as a whole.

Accessibility Feedback and Cultural Incentives

Participants discussed the need for stronger mechanisms enabling persons with disabilities to report barriers and influence system design. Cerf described existing accessibility reporting channels within Google’s real estate and workspace teams and emphasized the need for broader societal incentives that reward inclusive design.

He acknowledged that accessibility and inclusiveness are not yet deeply embedded cultural norms and suggested that organizations must create environments where proactive accessibility work is valued rather than treated as optional.

AI, Sentiment Detection, and Adaptive Interfaces

Shabbir raised the challenge of helping visually impaired users perceive the emotional atmosphere of meetings or conferences through AI-generated contextual cues. Cerf noted that detecting sentiment is far more difficult than describing physical scenes and remains challenging even for sighted humans.

Later, Cerf speculated about AI-driven adaptive interfaces capable of modifying presentation layers in real time based on user feedback and conversational interaction. He compared this concept to existing Google Home systems that control Internet-connected devices through speech.

He suggested future systems might:

• Dynamically alter interfaces

• Adjust content presentation

• Personalize accessibility settings conversationally

• Learn preferred interaction styles over time

Accessibility Trade-Offs and AI Captioning

The final portion of the discussion focused heavily on AI-generated captioning systems. Judith Hellerstein expressed concern that organizations increasingly prefer cheaper AI captioning solutions over human captioners, despite ongoing quality limitations.

Shabbir noted that many deaf and hard-of-hearing users historically disliked AI captions due to inaccuracies, though he acknowledged improvements in recent years. He insisted that cost should never determine whether accessibility accommodations are provided.

Charles Hall referenced ongoing W3C work related to caption accuracy standards for WCAG 3.0, noting that:

• 95 percent accuracy may be acceptable for entertainment

• It is insufficient for critical information systems

Cerf described his own reliance on captioning systems, including Google’s “Live Transcribe,” developed by deaf Google engineer Dimitri Kanevsky. While acknowledging amusing and sometimes serious transcription errors, he speculated that future AI-to-AI voice communication systems could help improve speech recognition quality further.

Giacomo Mazzone added that multilingual captioning systems still perform poorly outside English contexts, particularly during scientific or technical discussions where inaccuracies can become highly problematic.

Closing Reflections

As the session concluded, participants reiterated several recurring themes:

• Accessibility must be built into systems from the beginning

• Human lived experience remains essential despite AI advances

• AI offers transformative possibilities but also introduces new risks

• Privacy, trust, and bias remain unresolved challenges

• Accessibility should not be sacrificed for convenience or cost savings

Muhammad Shabbir thanked all participants, organizers, captioners, interpreters, and technical support staff for contributing to the discussion and reaffirmed the coalition’s commitment to continuing these conversations beyond GAAD itself.

RESOURCES

• IGF Dynamic Coalition on Accessibility and Disability (DCAD) — co-host; organizes the IGF fellowship program for persons with disabilities

• Accessibility Special Interest Group (A11ySIG) — co-host of the GAAD 2026 webinar

• Global Accessibility Awareness Day (GAAD) — observed every third Thursday of May

• “The Apomechanic Web and Digital Exclusion” — Dr. Muhammad Shabbir’s article coining the term discussed in the session

• UN Convention on the Rights of Persons with Disabilities (CRPD) — the rights-based “social model” framework Shabbir referenced

• Google Live Transcribe — free Android captioning app created by deaf Google research scientist Dimitri Kanevsky, cited by Vint Cerf

• W3C Ethical Principles for Web Machine Learning — one of the W3C standards Charles Hall flagged in the chat

• W3C WCAG 3.0 (draft) — the accessibility guidelines work addressing caption accuracy noted by Charles Hall

• ITU Facts and Figures 2025 — the connectivity data behind Cerf’s point that billions still lack Internet access