Computer scientists and artificial intelligence investigators working for Google Research collaborated with specialists in communication sciences and disorders at the University College of London and at Massachusetts General Hospital’s Institute of Health Professions to investigate incorporating chatbot technologies – such as OpenAI’S ChatGPT – into Speech Generating Devices for AAC technology users. They focused on 3 selected communicative tasks: extending brief replies, incorporating biographical information into replies, and requesting assistance. Procedurally, an AAC user selects the task type, produces telegraphically truncated text input, and indicates response variability range. AI algorithms process these user inputs to generate 4 grammatically fuller candidate responses to choose from. For users, the upsides are reduced text generation demands and rapid generation of multiple candidate responses; the downsides are uncertain relationships between provided inputs and desired outputs semantically, stylistically, and/or pragmatically.
To study user responses to a prototype interface of this sort, the investigators recruited 12 AAC users to try it out and provide feedback. Each of the 3 scenarios, such as ‘Extend Reply’, is represented by a Speech Macro window on a computer screen showing the communication context (e.g., what a communication partner has just said), provides a text entry field for the AAC user’s brief input, and presents a slider controlling desired response variability range. For example, when the communication partner’s question was “Hey, how is it going?” and the brief typed response was “OK”, and the ‘variability slider’ was set to a moderate level, the AI algorithms produced the following roster of four candidate responses: “It’s going pretty OK” / “I’m doing well, thanks” / “I’m pretty good” / “Not bad, how are you”.
The 12 subjects represented a disparate group, ranging in age from 25-74. They used devices ranging from Tobii Dynavox to Proloquo4Text on iPads to Assistive App Combinations on Android tablets; their input methods ranged from eye gaze to switch scanning to direct selection. On evaluation, the subjects felt AI systems have the potential to save them time, physical exertion, and cognitive effort in production, but they considered it important that the options generated reflect their communication style and preferences. Most participants found the 3 macros usually provided ‘very’ or ‘extremely’ useful suggestions, but not always: occasionally all four candidate responses were lacking. The users agreed that the ‘request help’ macro was especially helpful in carrying out daily tasks, gaining accessibility, and managing medical issues – important matters for them.
For further reading: S. Valencia, R. Cave, K. Kallarackal et al., 2023, “The less I type, the better”: How AI Language Models can Enhance or Impede Communication for AAC Users. Proceedings of CHI’23, article #830, 1-14. https://doi.org/10.1145/3544548.3581560