Some thoughts on Accessibility of Machine Learning and Generative AI

In preparation for the discussion at the meeting tomorrow, I reviewed the current draft’s table of contents to refresh my memory of what is addressed in the text that we now have.

It seems to me that we could consider an expanded structure that addresses the role of machine learning and generative AI throughout the “pipeline” encompassing the creation and use of Web content, including

  *   Its role in content development, including code generation, Web site and document development, and multimedia accessibility from the author’s perspective. In short, machine learning seems likely to have a growing role in the authoring environment, including software development, where its presence is already noteworthy.
  *   Its role in Web applications. In this case, the machine learning is part of the application; it is deployed server-side or client-side, and it provides or at least influences the user interface that is ultimately experienced. Some generative AI applications would fit into this category, for example those which offer natural language interaction, and which can produce text or graphical output. Here, the machine learning is integral to the deployed Web application, and not merely a means of constructing it.
  *   Its role in enhancing the accessibility of user interfaces and in assistive technologies at the point at which Web sites and applications are being interacted with by the user. This is already occurring. Examples include automated caption and image description generation. Improved speech recognition and handwriting recognition are also illustrative.
Although I don’t think these categories are as cleanly demarcated as one would like, they’re nevertheless a useful point of departure for organizing the document. The distinction between the role of machine learning/generative AI in the authoring environment and in the user’s environment seems to me to be a useful approach to dividing up the cases so they can be described, and the accessibility-related issues documented.

Some machine learning technologies can be applied in the authoring environment or in the user’s environment. For example, automated caption generation and image description could be deployed in either environment. However, the issues are different in each case. In the authoring context, there is more opportunity for collaborative review and correction of the generated material assuming that there are multiple authors with complementary abilities involved, whereas in the user’s environment, support for identifying and correcting errors may be unavailable (e.g., no human available to verify and correct image descriptions, or no opportunity to correct automated captions in real time).

In general, I think this document would benefit from an analysis that classifies the scenarios which are relevantly different in their implications for accessibility, and then discusses each of the cases.

Comments are welcome, of course, at the meeting  or via the mailing list.