Introduction
Artificial intelligence holds transformative potential for accessibility, yet skepticism remains warranted. As an accessibility innovation strategist at Microsoft and a leader of the AI for Accessibility grant program, I share many of the concerns raised in Joe Dolson’s recent piece on AI and accessibility. AI can empower or exclude, depending on how we design and deploy it. This article aims to complement that critical perspective by highlighting concrete opportunities where AI can make meaningful differences for people with disabilities. We must address real risks urgently, but it’s equally important to explore what’s possible—so we can move toward a more inclusive future.
The Current Landscape of AI and Accessibility
AI tools have advanced rapidly, yet their application in accessibility remains uneven. Computer-vision models, natural language processing, and machine learning offer new ways to bridge gaps—from generating image descriptions to enabling real-time captioning. However, these systems often operate in isolation, lacking the contextual understanding that human users instinctively apply. For example, an AI that describes an image may miss its purpose within a webpage, labeling decorative elements as informative or vice versa. These limitations underscore the need for careful implementation and continuous improvement.
Alternative Text: Progress and Pitfalls
Joe Dolson’s critique of AI-generated alternative text is well-founded. Current computer-vision models still produce descriptions that are often vague, inaccurate, or irrelevant. Because alt text relies on both the image content and its context—such as surrounding text, page layout, and user intent—today’s models fall short. They examine images in isolation, drawing from separate foundation models for text and vision that rarely communicate. This leads to descriptions that miss nuances, such as whether an image is purely decorative or essential for understanding.
Yet there is promise. As models improve, they can generate richer, more detailed descriptions. The key is not to replace human judgment but to augment it. Even flawed AI output can serve as a starting point, prompting users to refine or correct the text. When a model offers a candidate description, the human author can quickly adjust it—saving time while maintaining quality.
Human-in-the-Loop: A Pragmatic Approach
The most effective current strategy is a human-in-the-loop workflow. AI provides a draft, and a human reviews and edits it. This approach harnesses the speed of automation while preserving the accuracy and nuance that only human experience can provide. For example, an author might receive an AI-generated alt text suggestion and immediately recognize it as off-target—prompting them to write a better description. Over time, these corrections can train the model, creating a feedback loop that improves future output.
This model also accommodates different skill levels. New accessibility practitioners can learn from AI suggestions, while experts can use them as a quick starting point. The goal is to reduce friction and encourage more inclusive content creation.
Identifying Decorative vs. Informative Images
One promising avenue is training models to distinguish between decorative images (which may not require alternative text) and informative images (which do). Current AI analyzes images in isolation, but with context-aware training, a model could evaluate an image’s role within a page. For instance, a photo of a sunset used as a background might be decorative, while a chart illustrating sales trends is clearly informative. By classifying images, AI could help authors prioritize their accessibility efforts and ensure that essential visuals get proper descriptions.
This contextual understanding would also accelerate accessibility audits. Tools could flag images likely missing alt text or, conversely, highlight decorative images that are mistakenly described. As datasets and algorithms improve, this capability becomes more reliable—though human oversight remains essential.
Navigating Complex Visuals
Charts, graphs, and diagrams pose unique challenges. Even for humans, describing a complex data visualization succinctly is difficult. Current AI struggles with these, often generating overly simplistic or confusing descriptions. However, advances in multimodal models—those that combine text, vision, and layout—offer hope. Future systems could parse the structure of a chart, identify key trends, and generate a meaningful summary. Until then, complex visuals require careful human attention, but AI can assist by extracting raw data or suggesting outlines.
Looking Ahead
The opportunities for AI in accessibility are real, but they demand a thoughtful, inclusive approach. We must invest in diverse training data, prioritize user feedback, and maintain rigorous oversight. The potential—from smarter alt text to context-aware image classification—can reduce barriers for millions of people. By combining the best of human judgment and machine efficiency, we can create a digital world that truly works for everyone.