Causal analysis of social bias in CLIP

Carina I Hausladen, Manuel Knott, Pietro Perona, Colin Camerer

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: CLIP, Social biases, Causal datasets, Confounds, Valence words, Social psychology, Stereotypes

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We studied social biases in CLIP using causal and correlational datasets, highlighting the importance of accounting for confounds when evaluating bias.

Abstract: We propose the first experimental study to causally measure bias in social perception in the latent space of multi-modal models. Previous studies compute correlations between a model's social judgments and protected attributes, such as race, age, and gender, using observational wild-collected human-annotated datasets, such as FairFace. In order to establish causal links between protected attributes and algorithmic bias, we use a synthetic dataset of face images instead, CausalFace, where both legally protected attributes and potential confound attributes, such as facial expression, lighting, and pose, are controlled independently and systematically, and thus allow an experimental exploration, which lets us reach causal conclusions. Our analysis is based on measuring cosine similarities between images and word prompts, including valence words drawn from the two leading social psychology theories elucidating human stereotypes: The ABC Model and the Stereotype Content Model. We find that non-protected attributes are powerful confounds and profoundly influence social perception, injecting variability in measurements whose size is comparable to that induced by legally protected attributes. Clear intersecting biases of race, gender, and age only emerge when these unprotected attributes are controlled for, which is only possible using CausalFace. FairFace does not permit a similar level of insight due to spurious correlations introduced by uncontrolled attributes and a lack of specific annotations.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7341