As information technology becomes a necessity, rather than just a convenience, for most people, more attention is now focused on supporting the widest range of users; thus, making technology more inclusive.
At the same time as awareness of inclusive technology is increasing, radically new ways to support users are emerging from advances in generative AI. Can generative AI provide value for developers of inclusive technology? If so, in what ways? What further work will be needed to make the most of the potential? These are the questions addressed in the accompanying paper, with a specific emphasis on supporting people with dyslexia.
The authors focus on writing, where people with dyslexia report challenges at many levels, not only with spelling, punctuation, and grammar, but also in command of style and tone as well as higher-level issues like organizing and sequencing one’s ideas in a text. While all writers must deal with these matters, people with dyslexia find them to be especially difficult.
After extended consultation with people with dyslexia, the authors created LaMPost—a software tool driven by generative AI, that supports people writing email messages. LaMPost provides three kinds of support: Identify main ideas gives the user an outline of their draft email, as an aid to organization. Suggest possible changes asks the user to select a part of their draft, and then suggests revisions the user might consider, such as “make this passage more formal.” Users can then choose one of the suggestions, or suggest of their own, and submit it to the third feature, Rewrite my selection. That feature generates several potential rewrites, from which the user can choose.
The authors report on findings of the tool’s evaluation, centered on assessments by 19 adults with dyslexia, most of whom were frequent email composers. These people used LaMPost in a brief test session to work on email messages on topics of their own choosing.
The development of LaMPost illustrates some major advantages of generative AI as a supportive technology. The researchers were able to use few-shot prompting to tailor the functionality of the tool. Here, input from users is combined with desired responses to a few example inputs, as a way of shaping how the tool responds to the user. The mere fact this is possible sharply distinguishes generative AI from conventional software, where such tailoring requires programming. LaMPost also capitalizes on the remarkable ability of generative-AI language models to process content from a wide variety of topics, and their ability to manipulate language in many ways, with no need for specific programming. For example, such systems can provide useful responses to user inputs like “make the language less formal.”
However, in early 2022, user experiences also exposed serious weaknesses in generative AI. LaMPost’s suggestions sometimes included so-called hallucinations, material that looks superficially appropriate in the context, but is irrelevant or worse, and is unrelated to the user’s intent. For example, one rewrite suggestion included material about a patio, while the original draft email had nothing about patios. The paper also discusses other features that people with dyslexia would like to have, that are not included in LaMPost, including a facility able to compose an email draft from an outline. In pilot exploration, AI did not achieve useful elaboration of an outline without introducing unacceptable hallucinations.
If generative AI does enter the inclusive technology space, how will users feel about it? To approach this question, the authors presented LaMPost to their test users in one of two ways: either highlighting the role of AI in the tool, or not mentioning AI at all. The results showed no clear difference in user assessments based on the presentation they received. Not surprisingly, though, some users who received the no-AI framing guessed that LaMPost was based on AI, so the comparison may not be very informative.
The authors conclude that LaMPost, and the generative AI on which it is based, is not yet an effective writing support tool for people with dyslexia. Despite this pessimistic conclusion, this paper looks at significant aspects of the future of inclusive technology. The challenge of hallucinations, which greatly affected LaMPost, has not been eliminated, but has been reduced in more recent generative-AI systems. The authors’ use of few-shot prompting was constrained by the small size of prompts their system supported. More recent systems can accommodate far longer prompts, greatly increasing the effectiveness of the few-shot prompting techniques. So, it seems very likely that LaMPost will not be a deadend in the development of inclusive technology, but an early indication of what is possible.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment