Responsible AI & Human-AI Interaction
Incoming Research Scientist at Apple
Previously: Princeton CS PhD, Yale S&DS BSc, MSR FATE & TTIC Intern
https://sunniesuhyoung.github.io/
| Website | https://sunniesuhyoung.github.io |
| Names | Sunnie (pronounced as sunny☀️), Suh Young, 서영 |
| Pronouns | she/her/hers |
📢 research call 📢
please boost! we're running a series of small group design workshops for community governance on fedi. if you are a user, admin, mod, dev, organizer - or have thought quite a bit about governance on fedi - and a legal adult in your locale, join us!
each workshop, scheduled based on peoples' availability, is:
🌐 2 hrs on Zoom
👥 w/ 6-10 people
💲 comes w/ $60 USD per person
details: https://dsmw.cs.princeton.edu/
sign up: https://princetonsurvey.az1.qualtrics.com/jfe/form/SV_e3SdG4y6wocMPie
questions: frictance@princeton.edu
Check out Indu Panigrahi’s LBW at #CHI2025: “Interactivity x Explainability: Toward Understanding How Interactivity Can Improve Computer Vision Explanations.”
🔗 Project Page: https://ind1010.github.io/interactive_XAI
📄 Extended Abstract: https://arxiv.org/abs/2504.10745
@jenn @qveraliao @PrincetonCS @hci @citp
"Fostering Appropriate Reliance on LLMs" received an Honorable Mention at #CHI2025
This work is also the last chapter of my dissertation, so the recognition feels more special🏅🎓😊
🎉 to the team!!!
This work is with @jenn, @qveraliao, Tania Lombrozo, and Olga Russakovsky, a fantastic collaboration between @PrincetonCS, @hci, @citp, Princeton Psychology, and Microsoft Research FATE 🤝
A few of us will be at #CHI2025 in person!
📄 https://arxiv.org/abs/2502.08554
7/7
Large language models (LLMs) can produce erroneous responses that sound fluent and convincing, raising the risk that users will rely on these responses as if they were correct. Mitigating such overreliance is a key challenge. Through a think-aloud study in which participants use an LLM-infused application to answer objective questions, we identify several features of LLM responses that shape users' reliance: explanations (supporting details for answers), inconsistencies in explanations, and sources. Through a large-scale, pre-registered, controlled experiment (N=308), we isolate and study the effects of these features on users' reliance, accuracy, and other measures. We find that the presence of explanations increases reliance on both correct and incorrect responses. However, we observe less reliance on incorrect responses when sources are provided or when explanations exhibit inconsistencies. We discuss the implications of these findings for fostering appropriate reliance on LLMs.
We provide an in-depth discussion of our findings, bringing together the AI, HCI, and Psychology literature. We also contribute nuanced insights on people's interpretation of explanations from LLMs (e.g., are they faithful or not?), source clicking behavior, and more.
6/7
On a more optimistic note, we observe less overreliance on incorrect LLM responses when accurate and relevant #sources are provided or when explanations from LLMs exhibit #inconsistencies (i.e., sets of statements that cannot be true at the same time).
5/7
We find that the presence of #explanations increases reliance on both correct and incorrect LLM responses.
This isn't surprising, given prior HCI/Psychology work on other types of explanations, but raises the question whether LLMs should provide explanations by default 🤔
4/7
Then, through a pre-registered, controlled experiment (N=308), we isolate and study the effects of these features on users' reliance, accuracy, and other measures, such as confidence, source clicking, task time, evaluation of LLM responses, and asking of follow-up questions.
3/7
Princeton HCI
sohyeon hwang
Parastoo Abtahi