NLP folks are migrating to CHI. And so are ML practices. Just reviewed a paper that bolded their highest scoring intervention — no stats.
Remote
Zana Buçinca
@zbucinca@hci.social
mastodon 4.7.3PhD Candidate @Harvard; Human-AI Interaction
135 Followers
100 Following
16 Posts
Joined November 09, 2022
New study on Lab in the Wild: Given descriptions of people, can you recommend exercises that suit their constraints, goals, preferences, and needs?
Compare yourself to other test takers!
http://ai.labinthewild.org/exercise-recommendation-sets-optimal/?v=v0&PLATFORM=litw
Someone just responded to an email I sent in December 2022.
I am the first person from Kosovo to do a PhD at Harvard, but my most impressive statement remains “I went to school and was friends with Dua Lipa”.
Spending a weekend reading tens of papers in another field, just to add one single sentence to my own: Decades of research in [this field] have found [X, Y, Z].
NLP researchers be like “old school” work and cite a 2023 paper.
Anticipating harms of AI systems is a critical yet challenging task. We introduce AHA!, a generative framework that given a description of an AI system generates examples of possible harms. To generate examples of harm, we systematically consider stakeholders and AI problematic behaviors and narrate those via vignettes. These vignettes are then fed to the crowds and an LLM for surfacing specific harms that can occur to different stakeholders.
(1/2)
There’s more time to submit to TREW!
Due to popular demand, we’ve extended the submission deadline to March 1st.
Replying to
Our approach optimizes different objectives by adaptively providing decision support to humans —- the right type of assistance, to the right person, at the right time.
For example, when optimizing accuracy, we may withhold AI support when the AI is uncertain so the human does not overrely on it. When seeking to optimize human skills, we may show only partial support so people cognitively engage with the AI support.
Replying to
Or envisioning futures in which we seek to enhance collaboration and relatedness in the workplace, such dynamic assistance may even at appropriate times advise decision-makers to seek insights from and consult other colleagues for complex decision-making scenarios.
Replying to
Here, we instantiated our approach with human-AI accuracy and human learning as two objectives.
Our results consistently show that people interacting with policies optimized for accuracy achieve significantly better accuracy and even human-AI complementarity in decision-making. Our results further indicate that human learning is more difficult to optimize than accuracy, with participants who interacted with learning-optimized policies showing significant learning improvement only at times.
Replying to
Our research demonstrates (offline) RL to be a promising approach to model dynamics of human-AI decision-making, leading to policies that may optimize various human-centric objectives and provide novel insights about the AI-assisted decision-making space. Our work emphasizes the importance of considering human-centric objectives along with accuracy while also opening up the novel research challenge of devising human-AI interaction techniques and explanations that support such objectives.
By now, I must have heard every joke there could ever exist when CS or HCI people struggle to make a projector work.
“How many CS people does it take …”
“Talk about more HCI research ..”
“Maybe we need an AI …” …