top of page

HUMAN-CENTERED AI | AI FOR RESEARCH | HUMAN-AI INTERACTION

IMG_8821_edited_edited.jpg

2025 NUS YOUNG
FELLOWSHIP PROGRAMME

A ONE-MONTH FULLY FUNDED PROGRAM THEMED "AI AND PHD RESEARCH"

ROLE

Young Fellow and fellowship grant recipient

DESCRIPTION

I was nominated as a 2025 Young Fellow among the top 10% of global applicants, receiving the fellowship grant of SGD 600. This June, I was also invited to the National University of Singapore to explore the university, having the opportunity to participate in lectures from renowned AI scholars, lab visits, and workshops. I also participated in the research competition themed "AI and PhD research," collaborating with my teammates from India, Indonesia, Vietnam, and Thailand.

TEAMMATES

Aadarsh Ramachandiran (IIT), Dewi Aulia Maharani (UI), Nguyen Thanh Thuy Tu (VNU), Shine Min Kha (CMKL)

TIMELINE

Jun 2025 - Jul 2025

LOCATION

National University of Singapore, Singapore

RESEARCH METHODS

Literature Review

TARGET USERS

Researchers using AI tools

PROGRAM HIGHLIGHTS

During the fellowship, I was invited to the National University of Singapore to stay for a week. I was able to engage in lectures by leading AI scholars, workshops, and cutting-edge lab visits. Moreover, I was able to network with other fellows from around the world to engage in team activities.

NUS_YFP_2025–671_edited.jpg

PROJECT INTRODUCTION

This project examines the current landscape of AI in scientific research. By using sakana.ai as an interesting case study, we propose a collaborative model where humans and AI work as co-scientists rather than AI serving merely as a tool or completely replacing human researchers.​

 

Furthermore, we address the question: "Humans can do research alone. AI can too. But should they?" Here, we demonstrate the limitations of fully automated research systems and propose modifications based on Human-Centered AI (HCAI) principles.

3-MINUTE THESIS (3MT) PRESENTATION VIDEO

PROBLEM STATEMENT

Overreliance on AI can lead to
unreliable, non-transparent results
and sideline essential human insight.

This could lead to making unfair baseline comparisons, critical evaluation errors, and misleading findings. Most importantly, this ignores human creativity and ethical responsibility in a discourse overwhelmingly favoring AI replacement.

sakana.ai

Sakana.ai is a comprehensive system that can automate the entire research lifecycle.

The system is capable of:​

  • Generating novel research ideas

  • Writing code and executing experiments

  • Writing and publishing complete scientific papers

  • Operating without human intervention throughout the research process​

 
The system has demonstrated its capabilities by producing a paper on "Compositional Regularization: Unexpected Obstacles in Enhancing Neural Network Generalization" that was accepted at ICLR 2025.

Cong, L., Smith, J., & Doe, A. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv. https://arxiv.org/abs/2408.06292
Yamada, Y., Lee, H., & Patel, R. (2024). The AI Scientist-v2: Workshop-Level Automated Scientific Discovery. In Proceedings of the ICLR 2025 Workshop on AI for Science. Retrieved from https://pub.sakana.ai/ai-scientist-v2/paper/paper.pdf
Foerster Lab, & Sakana.ai. (2025). Towards Fully Automated Open-Ended Scientific Discovery. Sakana.ai. Retrieved from https://sakana.ai/ai-scientist/
Anonymous. (2025). Evaluating Sakana’s AI Scientist for Autonomous Research. arXiv. https://arxiv.org/abs/2502.14297v3
Lippl, S., & Stachenfeld, K. (2025). A Theoretical Analysis of Compositional Generalization in Neural Networks: A Necessary and Sufficient Condition [Poster session]. In ICLR 2025 Posters. https://openreview.net/forum?id=FPBce2P1er
Lippl, S., & Stachenfeld, K. (2025). When Does Compositional Structure Yield Compositional Generalization? A Kernel Theory [Poster session]. In ICLR 2025 Posters (Revised). https://openreview.net/forum?id=FPBce2P1er-revised

OUR THESIS

Human-AI collaboration will lead the future of research processes

"The real potential of AI in science

isn't in replacing or assisting humans,

but in collaborating together."

RESEARCH ON HUMAN-CENTERED AI

As a Human-AI Interaction researcher interested in human-centered AI, I was in charge of conducting a brief literature review on the topic.

Ben Schneiderman on Human-Centered Artificial Intelligence (HCAI)

  • High levels of human control AND high levels of automation are possible.

  • “Empowering, not Replacing.”

  • Governance structures for HCAI are needed.

MIT Slone School of Management

  • AI is better at forecasting demand and diagnosing medical issues.

  • Combinations work when humans and AI do what they each do best.

Complementarity in Human-AI Collaboration:

Concept, Sources, and Evidence

  • “Complementary” AI models designed for teamwork can be created.

  • Humans could be trained to build awareness to use unique contextual information.

When combinations of humans and AI are useful:

A systematic review and meta-analysis

  • Overall performance of human-AI collaboration is better than working separately, when humans are good at deciding when to trust their own judgements rather than those of the algorithms​

Cong, L., Smith, J., & Doe, A. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv. https://arxiv.org/abs/2408.06292
Yamada, Y., Lee, H., & Patel, R. (2024). The AI Scientist-v2: Workshop-Level Automated Scientific Discovery. In Proceedings of the ICLR 2025 Workshop on AI for Science. Retrieved from https://pub.sakana.ai/ai-scientist-v2/paper/paper.pdf
Foerster Lab, & Sakana.ai. (2025). Towards Fully Automated Open-Ended Scientific Discovery. Sakana.ai. Retrieved from https://sakana.ai/ai-scientist/
Anonymous. (2025). Evaluating Sakana’s AI Scientist for Autonomous Research. arXiv. https://arxiv.org/abs/2502.14297v3
Lippl, S., & Stachenfeld, K. (2025). A Theoretical Analysis of Compositional Generalization in Neural Networks: A Necessary and Sufficient Condition [Poster session]. In ICLR 2025 Posters. https://openreview.net/forum?id=FPBce2P1er
Lippl, S., & Stachenfeld, K. (2025). When Does Compositional Structure Yield Compositional Generalization? A Kernel Theory [Poster session]. In ICLR 2025 Posters (Revised). https://openreview.net/forum?id=FPBce2P1er-revised

POSTER PRESENTATION

Redefining Research: Human + AI as Co-Scientists

1.png

Impact

​This project contributes a new framework showing how humans and AI can work together as research partners to achieve much more than either could accomplish alone. By examining sakana.ai's limitations, our research reveals why fully automated AI research fails, and proposes a better model that combines human creativity and judgment with AI's data processing power. Overall, this project shows that this partnership makes research faster, more reliable, and accessible to smaller teams while keeping humans in control of important decisions.

My Key Takeaways

  • We need to focus on developing ways in which humans and AI can coexist. I've seen people who fear how AI will replace human beings in the near future. However, I think this will not be true, as long as we find ways to collaborate with AI together.

​​

  •  Instead of seeing AI as an agent of dehumanization, I see AI as a key to recovering humanity. Now that the overall productivity is boosted with AI, I think we can take the time and effort to take care of those who were marginalized and do research on how we can help them better with new technology.

​

  • Rather than feeling threatened by AI capabilities, researchers can learn to highlight the unique and irreplaceable contributions humans bring to scientific discovery. I think this includes creativity, ethical judgment, contextual understanding, and the ability to ask meaningful questions.

COPYRIGHT @ HEEYOUNG (EMILY) GHANG 2025

bottom of page