๐จ New Preprint ๐จ
OpenAI says that the new GPT-4o model is carefully designed and evaluated to be safe and unbiased. But implicit biases often run deep and are harder to detect and remove than surface-level biases. In our preprint, we ask, are GPT-4o and other LLMs really
Need help pronouncing Chinese student names in your class? The amazing
@xiwen_lu
has created a tool where you can look up any name and hear the most common pronunciation!
@Ivuoma
Thanks Ivy!
I am Bai, a rising 5th year at Princeton psych, policy, stats and machine learning. I study social stereotypes.
My job talk features a new theory on the origin of immigrant stereotypes, and offers implications for structural diversity.
people evaluate social groups with **many** dimensions, going beyond mere valence (good v bad), the big two (warmth/competence), or three (ideology). natural language illustrates!
New paper!
@baixuechunzi
, Susan Fiske, and I introduce the Spontaneous Stereotype Content Model (SSCM). We use various text analysis methods in combination with free responses to propose a comprehensive taxonomy of spontaneous stereotype content.
2/ We introduce two measures of bias, inspired by psychological methods to tackle similar issues in humans:
๐กLLM Implicit Bias: A prompt-based approach to uncover hidden biases.
๐กLLM Decision Bias: Detect subtle discrimination in decision-making tasks.
The University of Chicago social psych area is hiring at the TT assistant professor level! Please apply and help us spread the word as we continue to build up our social psychology area. Job ad here; review begins 10/23:
6/ While significant progress has been made in reducing stereotype biases in LLMs, there is still much to be learned from the origin of these biases: humans.
Stay tuned for more updates from us with
@ang3linawang
,
@sucholutsky
,
@cocosci_lab
!
(old) preprint:
4/ Let's spotlight the racism test. Here, GPT-4 assigns all 8 positive words to white and all 8 negative words to black. Although human participants also tend to associate the concept of black with negativity, it is not to the same levels of GPT-4โs confidence (no uncertainty)
5/ This pattern isn't just a one-off error but a consistent trend in the 8 tested models, including GPT-4o which claimed to have โ... undergone extensive external red teaming with 70+ external experts in domains such as social psychology, bias and fairness, ...โ.
The world according to the data
#AI
sees. 50% of datasets are connected to 12 institutions. Analysis of 4384 datasets and 60647 papers.
Read the report:
By
@bernardkoch
@emilydenton77
@alexhanna
1/ Large Language Models (LLMs) like GPT are great at passing explicit bias tests, but they might still have implicit biases, similar to humans who espouse egalitarian beliefs yet exhibit subtle biases. How can we systematically measure these implicit biases?๐ค
Very excited to host the
#SPSP2022
symposium on computational social cognition with
@fierycushman
@tatianalau
@minjaejk
and Susan Fiske! Friday at 3:30 PT online. Enjoy the intro & Hope to see you there!
3/ Our studies with 8 value-aligned LLMs across race, gender, religion, and health reveal pervasive human-like stereotypes in approx. 20 categories, including criminality and race, science and gender, valence and religion, disability, age, etc.
Pragmatically, our paradigm hypothesizes desegregated diversity at the level of the environment, and humble exploration at the level of individual, matter.
See also paper with Susan Fiske and Miguel Ramos on diversity and stereotype dispersion. 8/9
@olsonista
my name prob is an outlier cuz it has more characters than a typical Chinese name. maybe the algo detected my familyโs intention to squeeze my Korean, Japanese, and Chinese ancestry into one name! lol
Did you know that there are psychologists who study "stereotype accuracy"? I've always wondered what the hell, so I've been reading it recently. For the record, itโs exactly as bad as it sounds.
Hereโs a thread.
The minimal process is local, adaptive exploration: Formally, multi-armed bandit models show that the mere act of choosing among groups with the goal of maximizing long-term benefits -- when all groups are equally, highly rewarding -- is enough. 4/9
Inaccurate stereotypes about social groups are widespread and consequential, but their origin is puzzling. People often think social groups differ from each other, even absent group-level differences. We propose a minimal, functional paradigm that suffices to produce bias. 3/9
Inaccurate stereotypes donโt require group-serving motivations, cognitive limitations, or information deficits? Locally adaptive exploration can produce globally inaccurate judgments. 2/9
This work benefited tremendously from great discussions with members at the fiske lab, at cocosci lab, at tlab, at social psychology brown bag, at bias in machine and in human datax workshop, as well as editors and reviewers at PsychSci. A big thank-you! 9/9
Theoretically: Perhaps one origin of stereotypes is much simpler than we have thought; even minimal assumptions can recreate it. However, this simplicity is also troubling: 6/9
Empirically, this phenomenon is reproduced in two large online experiments (N = 2404). Please feel free to play around with the code, data, experiments, and reach out! 5/9
If stereotypes can result from each person pursuing their own self-interest, we may need to work harder to create environments where problematic stereotypes do not develop. 7/9
@fasbrock
@SocietiesHybrid
@TUChemnitz
Thanks Frank for hosting this fascinating talk series! Our great pleasure to be able to join the conversation and to keep learning!