AI Institute proposal

These notes are my reflections on yesterday’s conversation about the AI institute. I’ll make jottings first, and then come back and clean it up.

• I am motivated by Venu’s challenge: If we are proposing something which is sufficiently well-developed that a startup could be building it, it will not be funded by NSF. We need to tackle a big challenge.
• The big challenge I propose is to model and participate in subjectivity. This is the biggest challenge in CS, the one question at the heart of the field: What does it mean to be human? Can we create machines which act like people?
• The key insight I propose to follow is that we are not aiming for subjectivity as regarded by itself, but for participation in intersubjectivity. Our goal is not an AI system which feels itself to be conscious, but one which other participants in the learning environment regard as helpful, responsible, co-present.
• David’s proposal to create an AI system which “is there with you as a partner” and which “goes through the course with you” is very appealing. However, it also brings tremendous responsibility. We are rightly angry when somebody tries to establish personal intimacy under false pretences: when they turn out to be a paid shill rather than acting out of their own self-interest; when they establish rapport by pretending to be struggling with the same issues. The AI system is not, in fact struggling to keep up in class; it has nothing at stake. Have you interacted with an AI chatbot pretending to be human on a customer service line (even introducing artificial delays and playing the sounds of typing keys). It feels manipulative; similar dynamics could become outright abusive in an educational context.
• So we can’t create an AI agent who is also-struggling, a helpful friend who joins your study group, who sees you for who you are, affirms your identities and helps you feel less alone. But perhaps we can create a system which helps shape the social world of the learning environment to allow more of these kinds of interactions to occur. Of course, cultivating the sociocultural learning environment is already a central pedagogical responsibility of educators. However, so little effort is currently made to consider the sociocultural learning environment (and critical engagement with the broader discourses within which any learning environment is framed) in postsecondary learning environments that it feels premature to suggest that a tool is needed to support the task. Perhaps this system ought to be regarded as a professional development tool for teachers. One can read the impulse to “solve” sociocultural problems with technical tools as a way of avoiding responsibility for engaging with them.
• Too often, educational interventions focus exclusively on cognitive framings: they want to figure out which content to present when, how to taxonomize misconceptions, or which just-in-time insight will make the difference in solving a problem. This research is valuable, but if we are trying to focus on diversity, inclusion, marginalization, and justice, this framing is headed in the wrong direction. There is nothing wrong with the cognition of marginalized students. The problem is rooted in situated dynamics: disciplinary identity authorship, stereotype threat, the production of learning opportunities, implicit and explicit assumptions about available resources, prior experiences. These dynamics manifest in microaggressions which can be devastating to those who experience them and may go entirely unnoticed (or worse, regarded as harmless) by those who initiate them.
• In this institute, I propose that we center questions which are urgent for practitioners and at the frontier of understanding for both AI and Learning Sciences. (Pasteur’s quadrant.) Modeling and participating in intersubjectivity meets these criteria. Subjectivity is tricky to work with because it defines its own terms. That is, subjectivity cannot be fully observed or represented from the outside, using a priori criteria or features. Computer Science has plenty of tools for working with such phenomena. There are many models (e.g. hidden markov models) which estimate unknown parameters based on observations which may have a complex and indirect relationship with what is estimated. And unsupervised techniques (e.g. natural language processing word embeddings) can be used to characterize subjective meanings (ranging in scale from the meanings of words to the identities from which we act within figured worlds. There are so many beautiful and specific theories about sociocultural learning (e.g. Vygotsky, Bakhtin, Piaget) which have never been empirically tested in full social context. This institute would be an opportunity to do so.