Research |

I am currently a PhD researcher in the Social and Responsible Computing Group with Carlos Castillo at Pompeu Fabra University, where I am investigating 1) How to support safe and fair human-AI interactions with LLMs, and 2) How to respond to and direct AI in the classroom.

Prior to this, at Georgia Tech, I studied adult learning and online education with Ashok Goel and the Design Intelligence Lab, and I completed a Master's Project with David Joyner on AI tutors that support self-reliance.

Publications

Building Self-Reliance in an Era of Accessible Intelligent Tutors

Dinesh Ayyappan, David A Joyner

Tools for Thought Workshop at CHI (2025)

We worked with high school teachers to build a custom LLM interaction for teaching rhetorical analysis. We designed two different chat interactions with an AI tutor, deployed this with 71 students over two weeks, and report on insights from student behavior and teacher interviews.

As Generative AI tools become increasingly capable of automating cognitively demanding tasks, it is critical to design tools that augment---rather than replace---human thinking. To this end, we propose that tools for thought should make explicit the roles and levels of intervention in human-AI partnerships, especially in educational contexts. Through an exploratory study of 71 high school students using an AI writing tutor with explicitly-labeled modes of assistance, we investigate how structured interfaces influence learners' interactions with intelligent tutors. Our findings suggest that interface design may affect different aspects of self-regulation and student perceptions of AI assistance. We discuss implications for designing educational AI systems that support metacognitive development while preserving student agency. This work contributes to understanding how to balance AI support with the development of independent thinking skills in educational contexts.

PDF

A Constructivist Framing of Wheel Spinning: Identifying Unproductive Behaviors with Sequence Analysis

John Kos, Dinesh Ayyappan, Ashok Goel

Generative Intelligence and Intelligent Tutoring Systems (ITS) (2024)

This research looks at unproductive patterns of behavior in an open-ended, online, simulation with no correct answers. We find a few patterns that precede quitting with some predictive potential for future work.

Despite the increasing prevalence of inquiry-based learning (IBL) platforms, IBL research still has many open questions in constructivist contexts, such as when to offer help. In this paper, we explore wheel spinning in an open-domain inquiry-based modeling platform. We establish why closed-domain conceptions of wheel spinning do not work well in open domains, and we postulate key features of a working characterization of wheel spinning for an open-domain IBL platform. We also present a comparative quasi-experimental study of two groups of learners: While both groups worked with the same publicly available IBL platform called VERA, one group used VERA in a large graduate class at Georgia Tech and the other group was composed of globally-distributed self-directed learners (SDLs) who accessed VERA through various public websites. We use sequential pattern mining on the results in order to determine what wheel spinning looks like on the VERA platform and how to best identify it. Generalizing, we apply this research to the quitting behaviors of SDLs. Finally, modeling both user groups’ action sequences into Markov chains, we identify the stage at which an IBL platform may offer proactive feedback to the learners.

PDF DOI

Research Collaborations

Interested in collaborating? Please get in touch.