Accessible technology & responsible AI

A care experience should help people find information, understand it, and take a useful next step. Accessibility belongs in that work from the beginning. A person should not have to solve a difficult interface before they can ask a care question or read an update.

Responsible AI starts with a similarly concrete question: which part of this task is appropriate for assistance, and who remains responsible for the result? Describing a feature as intelligent or personalized does not answer either question.

Web Content Accessibility Guidelines 2.2 (opens in a new tab) supplies testable criteria for web accessibility. The December 2024 Recommendation covers needs including navigation, input, and accessible authentication. A citation to the standard is not evidence that an entire product conforms to it.

The AHRQ health literacy toolkit (opens in a new tab) addresses communication and the complexity of using care information. A technically operable page can still leave a person unsure what the words mean.

The NIST Generative AI Profile (opens in a new tab), published July 26, 2024, is voluntary risk-management guidance. It supports consideration of trustworthiness across AI design, use, and evaluation. It is not a clinical validation certificate.

We use these resources as separate starting points: access to the interface, understanding of the information, and responsibility for AI-assisted work.

Consider this hypothetical task: a family member receives an update, opens it on a phone, reads what happened, and asks a question. A useful local review would follow the entire journey.

Could they reach the content with their preferred input method? Could they enlarge the text? Was it clear which information was new? Did they understand whether a reply would reach a person? Could they complete the same task when they needed assistance or another communication format?

We recommend involving people with relevant access needs in that review and recording the barriers they encounter. An automated accessibility scan can contribute to the work, but the program still needs evidence that people can complete its intended tasks.

Imagine a staff member using an AI draft to prepare an ordinary activity summary. This is a proposed example, not a statement that a particular genusConnect feature is available in every deployment.

Before trying it, the organization should decide what information the draft may use, who checks it, and what would make the draft unsuitable to send. A reviewer should be able to compare the draft with the underlying information and correct a misleading statement. The organization should also decide what happens when no reliable draft can be produced.

Those decisions create a process that can be evaluated. Possible local questions include how often drafts need substantive corrections, what kinds of mistakes occur, how much review time is needed, and whether the final message is understandable to its intended reader.

Care communication should have a clear destination for questions that software cannot resolve. A person who is confused by an update needs a way to reach the responsible team. A staff member who spots conflicting information needs a way to stop and check it.

Our design position is to make the purpose and limits of assistance understandable within the task. The organization can then examine whether the chosen approach is useful and acceptable in practice.

For genusConnect, accessible technology and responsible AI are continuing design and evaluation responsibilities. Research and standards inform that work. Claims about a particular implementation should describe what was actually tested, with whom, and what remains outside that evidence.