Harsh
Kumar.
I am a PhD candidate in Computer Science at the University of Toronto, advised by Ashton Anderson. I study how everyday use of LLMs shapes people’s learning, creativity, and wellbeing. I use those findings to design better evaluations and post-training methods for AI systems.
I study what AI interaction leaves behind in people.
Evaluations today largely capture what models can do, and increasingly what humans and AI accomplish together. My work asks about a third, harder-to-measure layer: the lasting effects of these interactions on human capacities in everyday use. As LLM use scales, even small effects on learning, creativity, and wellbeing may accumulate into broader social consequences. I study these questions through large-scale randomized experiments.
My dissertation, Human Capacities in the Age of AI: Designing and Evaluating LLM Systems for Learning, Creativity, and Wellbeing, will be defended in Fall 2026. Along the way, I spent two summers with the Computational Social Science group at Microsoft Research NYC, working with Dan Goldstein, Jake Hofman, and David Rothschild on how math can be taught with LLMs, including a randomized trial with 1,800 learners. I also spent a summer as an Applied Scientist Intern at Microsoft with Shamsi Iqbal, studying how LLMs can support organizational sensemaking.
My work has been covered by Time, The Economist, The New York Times, Forbes, and Science News.
I am on the 2026–27 academic and industry job market, for roles in human–AI interaction and LLM evaluation/post-training. CV · Email
News
- Jul 2026 — I am attending ICML 2026 in Seoul to present our position paper, We Need Large Language Models Optimized For Our Well-Being. Say hi if you are around!
- Apr 2026 — Two papers at CHI 2026: When AI Gives Advice (on AI vs. human responses to wellbeing advice-seeking) and LLM Use and Critical Thinking under Time Constraints (led by Jiayin Zhi, with Mina Lee), the latter covered by Science News.
- Apr 2025 — Math Education with LLMs was accepted to AIED 2025.
- Mar 2025 — Human Creativity in the Age of LLMs received an Honorable Mention award at CHI 2025. I also discussed the work on the Teaching, Learning, and Everything Else podcast, and it was covered by TechXplore.
- Feb 2025 — LLM Agents for Improving Engagement with Behavior Change Interventions was accepted to CSCW 2025.
Research Themes & Representative Papers
Hundreds of millions of people now use AI in ways that are difficult to measure but consequential for what they learn, create, and rely on. My dissertation examines three of these domains. Read more about my research here · See the full list of papers here
Creativity & Cognition
LLM assistance can boost creative performance in the moment, but may reduce unassisted creative ability and the diversity of what people produce.
Human Creativity in the Age of LLMs
CHI 2025 ● Honorable Mention ● PDF ● Code
Randomized experiments with 1,100 participants on divergent and convergent thinking, and what repeated LLM use does to unassisted creativity.
Human Thinking under Plural LLM Assistance
under review ● Code
Controlled experiments on how systematically varied LLM personas affect mathematical problem solving and open-ended writing.
LLM Use and Critical Thinking under Time Constraints
CHI 2026 ● Project Website
How the timing of LLM access affects critical thinking under time pressure; led by Jiayin Zhi.
Learning
Does help from an LLM build understanding that persists once the model is gone? I also design interventions that target outcomes beyond retention, such as learners’ self-confidence and interest.
Math Education with LLMs: Peril or Promise?
AIED 2025
A randomized trial of LLM tutoring with 1,800 learners; featured in Microsoft’s New Future of Work report.
Guiding Students in Using LLMs in Supported Learning Environments
CSCW 2024
Two classroom interventions investigating how to help students use LLMs for learning.
Supporting Self-Reflection at Scale with LLMs
Learning@Scale 2024
Randomized field experiments in classrooms on LLM-supported self-reflection.
Wellbeing & Mental Health
People increasingly bring personal struggles to LLMs.
When AI Gives Advice
CHI 2026 ● Code
Compares AI and human responses to real online advice-seeking, measuring perceived helpfulness, empathy, safety, and other qualities of support.
LLM Agents for Improving Engagement with Behavior Change Interventions
CSCW 2025
Includes a four-week longitudinal study comparing reflective and informational agents for digital mindfulness.
Using Adaptive Bandit Experiments to Increase and Investigate Engagement in Mental Health
IAAI 2024
Field experiments with Mental Health America’s text-messaging program, reaching ~10,000 users.
Position: We Need Large Language Models Optimized For Our Well-Being
ICML 2026 ● PDF
The position statement for this agenda: optimize models for wellbeing, not short-horizon preferences.
Diagnosing and Repairing Persona Collapse in LLM Advice
request for preprint, under review
Fine-tuning open models, with probes and LLM-judge evaluations, to detect and repair a failure mode in LLM advice.
Grants & Competitions
● DARPA AI Tools for Adult Learning — US$250,000 award for QuickTA, an LLM-based learning tool that has supported 1,500+ students. I co-led the project with John Stamper and Norman Bier (2023)
● XPRIZE Digital Learning Challenge — US$500,000 grand prize; I led development of the contextual-bandit personalization system used for large-scale adaptive classroom experiments (2023)
I have also helped write grants with my mentors, including Florea.ai, a John Templeton Foundation–funded project on wellbeing (PIs: Louis Tay, Karina Vold, Ashton Anderson), where I lead the technical side as the senior graduate student.
Contact
I am always happy to chat — email is the best way to reach me.
Website template adapted from Abhraneel Sarma.