Thijs Roumen: “why'd you say that? Human-AI relationships for AAC users” February 5th 2026
Abstract: Accessibility is to date the best use case of AI, but (how) should we go about using AI to advance accessibility? Thijs explores this question through a lens of his own work in speech accessibility. The core questions is what role AI should/could play in supporting augmentative and alternative communication (AAC). Together with his student Tobias Weinberg, he developed interfaces to support the creation of humorous interjections through autocomplete. This AI-humor turned out to be a powerful form of backchanneling for participants in the study--it communicates presence rather than elaborate complex humor. In follow-up work, they ran workshops to dive deeper into the micro-culture of backchanneling as developed in AAC communities. Finally, to really find out what role AI should play, they conducted a longitudinal experiment where they trained a personalized AI on 7 months of Tobi's speech. In this talk Thijs discusses the advantages of accurate auto-complete and the costs in terms of privacy, agency, and identity of such ultra-personal AI.