In the ongoing conversation about artificial intelligence and accessibility, skepticism is healthy and necessary. Joe Dolson's recent critique rightly highlights many shortcomings in current AI applications. However, as an accessibility innovation strategist at Microsoft and a manager of the AI for Accessibility grant program, I see both the risks and the remarkable opportunities. This list explores where AI can genuinely make a difference—not as a replacement for human judgment, but as a powerful ally when used responsibly. Each point builds on the idea that AI, like any tool, can be wielded for good or harm. Let’s focus on the positive potential while acknowledging the work still needed.

1. AI as a Constructive Tool, Not a Savior

Skepticism about AI is warranted, especially in accessibility contexts. However, dismissing it entirely overlooks its potential to assist. The key is mindset: AI can be used to augment human capabilities, not replace them. For people with disabilities, AI can offer starting points that save time and reduce frustration—for example, generating rough drafts of alternative text. The goal isn't perfection but progress. By treating AI as a collaborative partner, we can harness its strengths while remaining vigilant about its limitations. This balanced perspective allows us to explore innovations without ignoring real risks like bias or inaccuracy.

2. Human-in-the-Loop Alt Text Authorship

Joe Dolson rightly emphasizes the flaws in fully automated alt text. But a human-in-the-loop approach where AI offers a starting point—even if flawed—can still be valuable. Imagine a system that generates a description, then prompts the user to correct it. That prompt alone raises awareness and reduces the daunting blank page. As I've seen in AI for Accessibility grantees, this hybrid model enhances efficiency. The AI might suggest "a person walking" while the human refines it to "a young woman in a wheelchair navigating a curb cut." That collaboration respects human expertise while leveraging AI speed.

3. Teaching AI to Distinguish Decorative vs. Informative Images

Current vision models analyze images in isolation, missing context. But specialized training could change that. By feeding models examples of decorative images (e.g., a border) versus informative ones (e.g., a chart), AI could flag images needing descriptions. This reduces the burden on content authors. For instance, a webpage with 20 images might only have 5 that are contextually important. AI that identifies the rest as decorative saves time and reduces unnecessary alt text. This contextual understanding is a frontier with huge accessibility gains.

4. Handling Complex Visuals Like Graphs and Charts

Describing complex images is tough even for humans. AI is not there yet, but progress is accelerating. The GPT-4 announcement showed improved ability to interpret detailed visuals. For graphs, AI can extract key data points and trends, producing a summary like "the line graph shows a 40% increase in sales from Q1 to Q2." While not perfect, such descriptions can serve as a foundation for human refinement. In academic or business contexts, this can dramatically improve access to data for screen reader users.

5. Continuous Improvement of Vision Models

Computer vision models are evolving rapidly. Early versions produced generic or wrong alt text, but newer models capture richer details—colors, actions, emotions. This trajectory suggests that with more training data and better algorithms, accuracy will improve. However, improvement requires diverse datasets that include people with disabilities and varied contexts. Funding research like the Microsoft AI for Accessibility grants accelerates this. The future likely holds models that not only describe but also interpret cultural cues and user intent.

6. Integrating Text and Image Analysis for Context

One major flaw: today's models often separate text and image analysis. A foundation model that merges both could understand that a picture of a dog is decorative in an article about pet care but informative in a veterinary guide. This integration is key to generating relevant alt text. Research into multi-modal AI—where the system reads the surrounding text and analyzes the image together—promises more intelligent decisions. For accessibility, this means fewer irrelevant descriptions and more meaningful support for users.

7. Acknowledging and Mitigating Real Risks

No discussion of AI and accessibility is complete without addressing risks: bias, privacy, and automation of discrimination. These aren't hypothetical—they affect real people. For example, image models trained on biased datasets may misrepresent marginalized groups. To counter this, inclusive design principles must guide development. The AI for Accessibility program requires grant recipients to evaluate ethical implications. By openly discussing risks and building safeguards, we can deploy AI responsibly. We need to address these issues yesterday, as Joe insisted.

8. Empowering Content Creators with AI as a Starting Point

Many content creators skip alt text because they lack time or knowledge. AI can lower that barrier. Even a poor AI suggestion prompts the creator to think about the image's purpose. Over time, such nudges build habits. Tools that generate a draft and then request review can significantly increase the overall accessibility of web content. This isn't about removing human responsibility but using AI to scaffold good practices. With thoughtful UX, creators might even find the process enjoyable rather than tedious.

9. Focusing on User-Centric Evaluation

AI's value in accessibility ultimately depends on user feedback. People with disabilities should be central to testing and refining AI tools. For instance, screen reader users can reveal whether AI-generated alt text helps or confuses. Their insights drive meaningful improvements. Programs like the AI for Accessibility grant require involvement of disability communities. This ensures that solutions address real needs rather than assumptions. User-centric design turns AI from a generic tech into a personalized assistive ally.

10. A Balanced, Optimistic Vision for the Future

The path forward involves embracing AI's potential while staying grounded in ethics and human oversight. By investing in research, diverse data, and inclusive design, we can create tools that make a genuine difference. I believe we'll reach a point where AI assists with everything from image descriptions to real-time captioning to predictive navigation for people with cognitive disabilities. But this requires collective effort—from developers, policy makers, and users. The opportunity is real, and the time to act is now, with both hope and caution.

AI's role in accessibility is complex, but not hopeless. By focusing on collaborative models, contextual awareness, and ethical safeguards, we can unlock benefits that transform lives. While skepticism keeps us honest, optimism drives innovation. Let's strive for a future where technology truly leaves no one behind.