Jennifer Chien, PhD

AI start-up Anthropic settles landmark copyright suit for $1.5bn Case will compensate authors but could raise costs of training large language models

This is big news. www.ft.com/content/96b5...

here’s some music to supplement your perusing #unknowingrabbit #ai #fooled

Do you feel like you can tell when something is AI or not? Here’s an example that has fooled the masses, a prime example for measuring cognitive effects of AI outputs on users and be thinking about ways to easily differentiate AI-produced products

Read more here:
dl.acm.org/doi/10.1145/...

on the paypal page under "use this donation for" "Queer in AI + oSTEM Grad School Application Program" should be default selected, but if its not just select that and your donation will go to the scholarship

I know that lots of very important causes are asking for funding right now but I wanted to personally endorse this one, please donate anything you can!

#Donate today: www.paypal.com/donate/?host...

@queerinai.com @ostem.org

A great opportunity for phd students!
#trustworthyai #fellowship
cra.org/cra-and-micr...

It is always rewarding to chat with students about my research! So thankful for the opportunity ☺️
#GenAIUCSD25 #genai #research

I will be talking about this work (and some other in progress work) at the #GenAISummit at UCSD CSE tomorrow!

genaisummit2025.ucsd.edu

We conclude with our proposal: latent-value modeling for determining trustworthiness. We differentiate between the user perspective and the LLM perspective for latent values. Our contribution poses as a reflexive tool and opportunity for various stakeholders to refine their contributed values.
(4/4)

We discuss extant approaches to value alignment, such as direct implementation of human values and behavioral approaches (e.g., RLHF) and the significant challenges they face for determining trust. (3/4)

Along the way, we discuss some alternative means to managing stochasticity (e.g., eliminating stochasticity and representing stochasticity to users), before proposing value alignment for trustworthiness.
(2/4)

📄 How do we think about trusting systems that exhibit stochasticity (*cough* LLMs)? In this work, we explicate the tension stochasticity poses with traditional black-box approaches to trust and propose causal modeling of latent values to directly determine trustworthiness

arxiv.org/pdf/2501.16461

In case you missed it, I am at the conference all week and on the job market, so happy to chat all things research (e.g. RAI, ethics, user agency, content moderation, human-centered AI) #NeurIPS2024 @neuripsconf.bsky.social

Interested in hearing more? Stop by tomorrow and let's chat~ #contentmoderation #participatorydesign #dataannotation

In this work, we collab w/ lips.social to answer the following:
How is community moderation (CM) different from algorithmic moderation? What value is added through tagging content differently? What kinds of people/groups need to be involved in tagging vs moderating? How does CM limit expression?

Planning your #NeurIPS2024 itinerary?
Stop by the #QueerinAI poster session from 6:30-8pm PST tomorrow (Tuesday) to hear about some of my ongoing work: "Diversifying Data Annotation: Measuring Differences in Community-Driven Content Moderation"