Aakriti Kumar

I'm currently a Postdoctoral Researcher at the Kellogg School of Management and the Northwestern Institute on Complex Systems (NICO), working with Dr. Matt Groh, where I build tools to support human communication and design AI evaluations that support human metacognition. I run large-scale lab, digital, and field experiments in real-world settings, combining behavioral data with computational and qualitative analysis. My research spans organizational behavior, AI, and human–computer interaction.

I earned my PhD in Cognitive Science from UC Irvine, advised by Dr. Mark Steyvers, along with an MS in Statistics from UCI and a B.Tech. from IIT Madras. My PhD research explored how people infer what a human or an AI collaborator knows, how good they are at a task, and when to rely on their advice. Along the way, I've also worked in industry at Honda Research Institute and Motional (a robotaxi startup), and collaborated with researchers at Google Research India.

Selected Publications

Annotated conversation showing empathic and misattuned responses to someone who lost their job

Boosting people's empathic communication skills with LLM role playing games

Aakriti Kumar, Nalin Poungpeth, Diyi Yang, Erina Farrell, Bruce Lambert, Matthew Groh

Under review

paper / Lend an Ear / code

An RCT (N=968) testing whether AI coaching and practice improve how people support each other.

Inter-rater reliability (weighted Cohen's kappa) of experts, LLMs, and crowds across empathy subcomponents

When large language models are reliable for judging empathic communication

Aakriti Kumar, Nalin Poungpeth, Diyi Yang, Erina Farrell, Bruce Lambert, Matthew Groh

Nature Machine Intelligence, 2026

paper / code

When can LLM judgments of empathy stand in for expert and crowd judgments?

Diagram of how people estimate an LLM's confidence from its answers and explanations

What large language models know and what people think they know

Mark Steyvers, Heliodoro Tejeda, Aakriti Kumar, Sheer Karny, Xinyue Hu, Lukas Mayer, Padhraic Smyth

Nature Machine Intelligence, 2024

paper

The calibration gap: people systematically overestimate LLM confidence.

Screenshot of the AI-generated or Real experiment interface

Characterizing photorealism and artifacts in diffusion model-generated images

Narges Kamali, Kento Nakamura, Aakriti Kumar, Andreas Chatzimparmpas, Jessica Hullman, Matthew Groh

CHI 2025

paper / Detect Fakes!

How photorealistic are diffusion models, and what artifacts give them away?

Hierarchical generative model relating population ability, self and other ability, task difficulty, and knowledge signals

Differentiating mental models of self and others: A hierarchical framework for knowledge assessment

Aakriti Kumar, Padhraic Smyth, Mark Steyvers

Psychological Review, 2023

paper

A hierarchical Bayesian framework for how people build distinct mental models of their own knowledge and others'.

Table of real-world AI-assisted decision-making examples mapped to three collaboration challenges

Three challenges for AI-assisted decision-making

Mark Steyvers, Aakriti Kumar

Perspectives on Psychological Science, 2024

paper

Complementarity, mental models of AI, and interaction design as the core challenges for human–AI decision-making.

Selected Publications

News