Stephanie C.Y. Chan

Stephanie C.Y. Chan

Staff Research Scientist, Google DeepMind

Google Scholar LinkedIn @scychan_brains

What I work on

01Future impacts of AI

I'm starting a sub-team within Post-AGI Research at DeepMind, using social science and other methods to understand the potential impacts of AI on society.

02AI to empower humans

AI can be used to augment human potential and human capital, if we do it responsibly and inclusively. In that spirit, I've co-led and contributed to teams working to strengthen the information ecosystem and to support education (launches with Gemini and YouTube).

03Science of AI

I use carefully designed experiments to understand modern AI systems, much as neuroscientists and cognitive scientists try to understand the brain. A lot of my work focuses on emergent behaviors and in-context learning.

In my past lives, I've done research in robotics, human cognition, and astrophysics. I've also performed brain surgeries, built laser-optics rigs, and pitched to VCs.

Selected recent papers

Future impacts of AI

Architecting Trust in Artificial Epistemic Agents2026
From AGI to ASI2026
Positive Alignment: Artificial Intelligence for Human Flourishing2026

AI to empower humans

Backstory: exploring the context of online images2025
LearnLM: Improving Gemini for Learning2024
Towards responsible development of generative AI for education2024

Science of AI

Predictability Shapes Adaptation: An Evolutionary Perspective on Modes of Learning in Transformers2026
The emergence of sparse attention: impact of data distribution and benefits of repetitionOral · NeurIPS2025
How do language models learn facts? Dynamics, curricula and hallucinationsOral · COLM2025
Strategy Coopetition Explains the Emergence and Transience of In-Context LearningOral · ICMLtied #3 of 12,107 submissions2025
Many-Shot In-Context LearningSpotlight · NeurIPS2024
What needs to go right for an induction head?Spotlight · ICML2024
The Transient Nature of Emergent In-Context Learning in Transformers2023
Oral · NeurIPS2022most-downloaded ML paper on arXiv on release dateData Distributional Properties Drive Emergent In-Context Learning in Transformers

Other highlights

Uncovering Competency Gaps in Large Language Models and Their Benchmarks2026
Genie: Generative Interactive EnvironmentsBest Paper · ICML2024

Invited lecturer at Stanford, Harvard, and other universities. Invited keynote speaker at conferences and workshops. Previously co-led Women@DeepMind.

Full list of papers on Google Scholar. I'm so grateful for all the amazing students, mentors, and collaborators that I get to work with, without whom none of this work is possible.

Writing

What happens after we solve continual learning?
LLMs are not just statistical pattern matchers
Compute as Money
Reversing the loop on imposter syndrome