Papers by Aylin Caliskan
VIGNETTE: Socially Grounded Bias Evaluation for Vision-Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing studies on VLM bias focus on portrait-style images and gender-occupation associations . existing studies ignore broader and more complex social stereotypes and their implied harm . |
| Approach: | They propose a large-scale VQA benchmark for evaluating bias in vision-language models . they use a question-answering framework that spans factuality, perception, stereotyping, and decision making . |
| Outcome: | The proposed framework examines bias in vision-language models using 30M+ images . findings reveal subtle, multifaceted, and surprising stereotypical patterns . |
Low Frequency Names Exhibit Bias and Overfitting in Contextualizing Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Infrequent names are less similar to initial representations, and are more self-similar, suggesting that models rely on less context-informed representations of uncommon and minority names. |
| Approach: | They use a dataset of U.S. first names with labels based on predominant gender and racial group to examine effect of training corpus frequency on tokenization, contextualization, similarity to initial representation, and bias. |
| Outcome: | The results show that infrequent names are less similar to initial representations and have a Spearman’s rho between frequency and self-similarity as low as .763 . |
Talent or Luck? Evaluating Attribution Bias in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing studies on social biases in large language models focus on surface-level associations or isolated stereotypes. |
| Approach: | They propose a cognitively grounded bias evaluation framework to capture demographic biases across three contexts: single-actor, actor–actor and actor–observer. |
| Outcome: | The proposed framework captures comparative and perspective-driven biases overlooked in previous work. |
Pre-trained Speech Processing Models Contain Human-Like Biases that Propagate to Speech Emotion Recognition (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work has established that a person’s demographics and speech style affect how well speech processing models perform for them. |
| Approach: | They propose a method to detect bias in pre-trained models by using word embedding association tests in natural language processing to quantify bias in models' representations of different concepts. |
| Outcome: | The proposed method detects bias in pre-trained models and can have real-world effects. |
BiasDora: Exploring Hidden Biased Associations in Vision-Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies on social biases focus on a limited set of documented associations, such as gender-profession or race-crime. |
| Approach: | They propose to examine hidden, implicit bias associations across 9 bias dimensions by probing VLMs to uncover hidden, unexamined associations. |
| Outcome: | The proposed methods reveal that biases vary in negativity, toxicity, and extremity. |
Global Gallery: The Fine Art of Painting Culture Portraits through Multilingual Instruction Tuning (2024.naacl-long)
Copied to clipboard
| Challenge: | This study examines the ability of Large Language Models to encapsulate cultural nuances across diverse linguistic landscapes. |
| Approach: | They examine the efficacy of language-specific instruction tuning and the impact of pretraining on dominant language data in Large Language Models. |
| Outcome: | The findings highlight a nuanced landscape, with inconsistencies and biases, particularly in non-Western cultures. |
Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior studies have reported that large language models (LLMs) are also susceptible to human-like cognitive biases, but the extent to which LLMs selectively reason toward identity-congruent conclusions remains unexplored. |
| Approach: | They investigate whether assigning 8 personas across 4 political and socio-demographic attributes induces motivated reasoning in LLMs. |
| Outcome: | The proposed model is assigned 8 personas across 4 political and socio-demographic attributes and shows that they have 9% reduced veracity discernment compared to models without persona. |
Contrastive Visual Semantic Pretraining Magnifies the Semantics of Natural Language Representations (2022.acl-long)
Copied to clipboard
| Challenge: | Large-scale "natural language supervision" using image captions has enabled the first "zero-shot" AI image classifiers, which allow users to create their own image classes using natural language, yet outperform supervised models on common language-and-image tasks. |
| Approach: | They compare the geometry and semantic properties of contextualized English language representations formed by GPT-2 and CLIP, a zero-shot multimodal image classifier which adapts the GPT2 architecture to encode image captions. |
| Outcome: | The proposed classifier outperforms GPT-2 on word-level semantic intrinsic evaluation tasks and achieves a new corpus-based state of the art for the RG65 evaluation. |
Intrinsic Bias is Predicted by Pretraining Data and Correlates with Downstream Performance in Vision-Language Encoders (2025.naacl-long)
Copied to clipboard
| Challenge: | Recent work has found that vision-language models trained under the Contrastive Language Image Pre-training framework contain intrinsic social biases, but how these biase relates to downstream performance has been unclear. |
| Approach: | They present the largest comprehensive analysis to-date of how upstream pre-training factors and downstream performance of CLIP models relate to their intrinsic biases. |
| Outcome: | The proposed model performance analysis shows that the choice of pre-training dataset is the most significant upstream predictor of bias, whereas architectural variations have minimal impact. |
‘Person’ == Light-skinned, Western Man, and Sexualization of Women of Color: Stereotypes in Stable Diffusion (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Using CLIP-cosine similarity for zero-shot classification of images, we chronicle results from 136 prompts (50 results/prompt) of front-facing images of faces from 6 different continents, 27 countries and 3 genders. |
| Approach: | They use CLIP-cosine similarity for zero-shot classification of images generated by CLIP based Stable Diffusion v2.1 verified by manual examination to determine what gender and nationality/continental identity is assigned to ‘a person’. |
| Outcome: | The results show that the image generator Stable Diffusion displays gender and nationality/continental identity in the absence of such information. |
Biases Propagate in Encoder-based Vision-Language Models: A Systematic Analysis From Intrinsic Measures to Zero-shot Retrieval Outcomes (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing encoder-based vision-language models (VLMs) contain intrinsic biases that manifest in biased outputs. |
| Approach: | They propose a framework to measure intrinsic bias propagation by correlating intrinsic bias with extrinsic bias in zero-shot text-to-image and image-totext retrieval. |
| Outcome: | The proposed framework shows that larger/better-performing models exhibit greater bias propagation, raising concerns given the trend towards increasingly complex AI models. |
ValNorm Quantifies Semantics to Reveal Consistent Valence Biases Across Languages and Over Centuries (2021.emnlp-main)
Copied to clipboard
| Challenge: | Word embeddings learn implicit biases from word co-occurrence statistics . valNorm is a new intrinsic evaluation task and method to quantify affect in word embedded word sets . |
| Approach: | They propose a method to quantify valence dimension of affect in human-rated word sets . they apply ValNorm to embeddings from seven languages and 200 years of text . |
| Outcome: | The proposed method achieves a high accuracy in quantifying the valence of non-discriminatory, non-social group word sets. |