Central issues in Accessibility of Machine Learning and Generative AI

To advance the discussion before next week’s meeting, I think it would be useful to elaborate the two themes which I raised last week, as these seem to capture many of the issues considered by the Task Force.


  1.  Can any given machine learning or generative AI system produce output that is accessible? This question has different answers depending on context. In the authoring environment, in which generative AI is used to produce or modify content, the output has to satisfy the needs of a diversity of users. Thus, the accessibility question is typically answered in practice by reference to whether the AI/ML system can produce output satisfying accessibility guidelines. However, if the AI/ML is used as part of the application and invoked directly or indirectly by the user, the question is whether its output can meet the access needs of a specific person - a different problem.

In the second case, information about the user’s access needs may be available to the AI/ML system. This introduces privacy concerns, depending on what information about the user is disclosed and to whom. A system running on the user’s hardware may be more privacy-preserving, but this isn’t guaranteed.

In general, I think we should distinguish clearly between satisfying accessibility guidelines, and satisfying the needs of a specific, identified user, as the issues involved seem quite different due to the option of customizing the output for a specific individual that exists in the second scenario.

If AI/ML is invoked as part of a Web application, then we are concerned with both its ability to generate output that meets accessibility guidelines, and its ability to tailor its output to the needs of each specific user.


  1.  How well does the AI/ML system perform its task? How error-prone is it in accessibility-relevant scenarios? This question applies to all of these systems, whether classified as “generative" or not. For example, a system that assesses the accessibility of Web content may only produce an evaluation of some kind (a score, a structured report, etc.) However, the question of accuracy nevertheless arises in such cases.

It seems to me that the problem of error-proneness provides the rationale for considering how to combine human capabilities with those of AI/ML systems appropriately to complete tasks. If the AI were reliable, we wouldn’t need human review or intervention, and we could confidently give the task entirely over to the machine. Thus, some of Josh’s broader concerns, and perhaps all of them, fit within this general topic of accuracy and reliability - identifying the tasks for which AI/ML systems are sufficiently reliable to be useful, and addressing the errors that they nevertheless make.

The problem of how to evaluate and refine AI/ML systems appropriately to improve their performance is also motivated by the reliability issue.

The nature of the available human intervention to address the reliability shortcomings of AI/ML systems varies according to the social context in which they are used. For example, the authoring environment offers opportunities for evaluation and correction that typically aren’t available to the ultimate user of a Web site or application. Thus, I think the reliability issue has different solutions depending on context - and we should make this clear in our future drafts.

To summarize the main potential benefits of AI/ML for accessibility, we can identify at least two:


  1.  Improving human productivity by partly or fully automating accessibility-related tasks.

b. Offering opportunities for user interface customization and responsiveness to the needs of the user that were not achievable with previous technologies.

Comments are welcome, as is discussion at next week’s meeting.