Aylin Caliskan

Federico Bianchi, Pratyusha Kalluri, Esin Durmus et al. · stanford

Machine learning models that convert user-written text descriptions into images are now widely available online and used by millions of users to generate millions of images a day. We investigate the potential for these models to amplify dangerous and complex stereotypes. We find a broad range of ordinary prompts produce stereotypes, including prompts simply mentioning traits, descriptors, occupations, or objects. For example, we find cases of prompting for basic traits or social roles resulting in images reinforcing whiteness as ideal, prompting for occupations resulting in amplification of racial and gender disparities, and prompting for objects resulting in reification of American norms. Stereotypes are present regardless of whether prompts explicitly mention identity and demographic language or avoid such language. Moreover, stereotypes persist despite mitigation strategies; neither user attempts to counter stereotypes by requesting images with specific counter-stereotypes nor institutional attempts to add system ``guardrails'' have prevented the perpetuation of stereotypes. Our analysis justifies concerns regarding the impacts of today's models, presenting striking exemplars, and connecting these findings with deep insights into harms drawn from social scientific and humanist disciplines. This work contributes to the effort to shed light on the uniquely complex biases in language-vision models and demonstrates the ways that the mass deployment of text-to-image generation models results in mass dissemination of stereotypes and resulting harms.

Aylin Caliskan, Pimparkar Parth Ajay, Tessa Charlesworth et al.

The statistical regularities in language corpora encode well-known social biases into word embeddings. Here, we focus on gender to provide a comprehensive analysis of group-based biases in widely-used static English word embeddings trained on internet corpora (GloVe 2014, fastText 2017). Using the Single-Category Word Embedding Association Test, we demonstrate the widespread prevalence of gender biases that also show differences in: (1) frequencies of words associated with men versus women; (b) part-of-speech tags in gender-associated words; (c) semantic categories in gender-associated words; and (d) valence, arousal, and dominance in gender-associated words. First, in terms of word frequency: we find that, of the 1,000 most frequent words in the vocabulary, 77% are more associated with men than women, providing direct evidence of a masculine default in the everyday language of the English-speaking world. Second, turning to parts-of-speech: the top male-associated words are typically verbs (e.g., fight, overpower) while the top female-associated words are typically adjectives and adverbs (e.g., giving, emotionally). Gender biases in embeddings also permeate parts-of-speech. Third, for semantic categories: bottom-up, cluster analyses of the top 1,000 words associated with each gender. The top male-associated concepts include roles and domains of big tech, engineering, religion, sports, and violence; in contrast, the top female-associated concepts are less focused on roles, including, instead, female-specific slurs and sexual content, as well as appearance and kitchen terms. Fourth, using human ratings of word valence, arousal, and dominance from a ~20,000 word lexicon, we find that male-associated words are higher on arousal and dominance, while female-associated words are higher on valence.

Katelyn X. Mei, Sonia Fereidooni, Aylin Caliskan · cambridge

The rapid deployment of artificial intelligence (AI) models demands a thorough investigation of biases and risks inherent in these models to understand their impact on individuals and society. This study extends the focus of bias evaluation in extant work by examining bias against social stigmas on a large scale. It focuses on 93 stigmatized groups in the United States, including a wide range of conditions related to disease, disability, drug use, mental illness, religion, sexuality, socioeconomic status, and other relevant factors. We investigate bias against these groups in English pre-trained Masked Language Models (MLMs) and their downstream sentiment classification tasks. To evaluate the presence of bias against 93 stigmatized conditions, we identify 29 non-stigmatized conditions to conduct a comparative analysis. Building upon a psychology scale of social rejection, the Social Distance Scale, we prompt six MLMs: RoBERTa-base, RoBERTa-large, XLNet-large, BERTweet-base, BERTweet-large, and DistilBERT. We use human annotations to analyze the predicted words from these models, with which we measure the extent of bias against stigmatized groups. When prompts include stigmatized conditions, the probability of MLMs predicting negative words is approximately 20 percent higher than when prompts have non-stigmatized conditions. In the sentiment classification tasks, when sentences include stigmatized conditions related to diseases, disability, education, and mental illness, they are more likely to be classified as negative. We also observe a strong correlation between bias in MLMs and their downstream sentiment classifiers (r =0.79). The evidence indicates that MLMs and their downstream sentiment classification tasks exhibit biases against socially stigmatized groups.

Robert Wolfe, Mahzarin R. Banaji, Aylin Caliskan

We examine the state-of-the-art multimodal "visual semantic" model CLIP ("Contrastive Language Image Pretraining") for the rule of hypodescent, or one-drop rule, whereby multiracial people are more likely to be assigned a racial or ethnic label corresponding to a minority or disadvantaged racial or ethnic group than to the equivalent majority or advantaged group. A face morphing experiment grounded in psychological research demonstrating hypodescent indicates that, at the midway point of 1,000 series of morphed images, CLIP associates 69.7% of Black-White female images with a Black text label over a White text label, and similarly prefers Latina (75.8%) and Asian (89.1%) text labels at the midway point for Latina-White female and Asian-White female morphs, reflecting hypodescent. Additionally, assessment of the underlying cosine similarities in the model reveals that association with White is correlated with association with "person," with Pearson's rho as high as 0.82 over a 21,000-image morph series, indicating that a White person corresponds to the default representation of a person in CLIP. Finally, we show that the stereotype-congruent pleasantness association of an image correlates with association with the Black text label in CLIP, with Pearson's rho = 0.48 for 21,000 Black-White multiracial male images, and rho = 0.41 for Black-White multiracial female images. CLIP is trained on English-language text gathered using data collected from an American website (Wikipedia), and our findings demonstrate that CLIP embeds the values of American racial hierarchy, reflecting the implicit and explicit beliefs that are present in human minds. We contextualize these findings within the history and psychology of hypodescent. Overall, the data suggests that AI supervised using natural language will, unless checked, learn biases that reflect racial hierarchies.

Yiyou Sun, Xinyang Han, Weichen Zhang et al.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 subfields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP-relevant impact.

Chahat Raj, Anjishnu Mukherjee, Aylin Caliskan et al.

Large Language Models (LLMs) perpetuate social biases, reflecting prejudices in their training data and reinforcing societal stereotypes and inequalities. Our work explores the potential of the Contact Hypothesis, a concept from social psychology for debiasing LLMs. We simulate various forms of social contact through LLM prompting to measure their influence on the model's biases, mirroring how intergroup interactions can reduce prejudices in social contexts. We create a dataset of 108,000 prompts following a principled approach replicating social contact to measure biases in three LLMs (LLaMA 2, Tulu, and NousHermes) across 13 social bias dimensions. We propose a unique debiasing technique, Social Contact Debiasing (SCD), that instruction-tunes these models with unbiased responses to prompts. Our research demonstrates that LLM responses exhibit social biases when subject to contact probing, but more importantly, these biases can be significantly reduced by up to 40% in 1 epoch of instruction tuning LLaMA 2 following our SCD strategy. Our code and data are available at https://github.com/chahatraj/breakingbias.

Chahat Raj, Anjishnu Mukherjee, Aylin Caliskan et al.

Existing works examining Vision-Language Models (VLMs) for social biases predominantly focus on a limited set of documented bias associations, such as gender:profession or race:crime. This narrow scope often overlooks a vast range of unexamined implicit associations, restricting the identification and, hence, mitigation of such biases. We address this gap by probing VLMs to (1) uncover hidden, implicit associations across 9 bias dimensions. We systematically explore diverse input and output modalities and (2) demonstrate how biased associations vary in their negativity, toxicity, and extremity. Our work (3) identifies subtle and extreme biases that are typically not recognized by existing methodologies. We make the Dataset of retrieved associations, (Dora), publicly available here https://github.com/chahatraj/BiasDora.

Robert Wolfe, Yiwei Yang, Bill Howe et al.

Nine language-vision AI models trained on web scrapes with the Contrastive Language-Image Pretraining (CLIP) objective are evaluated for evidence of a bias studied by psychologists: the sexual objectification of girls and women, which occurs when a person's human characteristics, such as emotions, are disregarded and the person is treated as a body. We replicate three experiments in psychology quantifying sexual objectification and show that the phenomena persist in AI. A first experiment uses standardized images of women from the Sexual OBjectification and EMotion Database, and finds that human characteristics are disassociated from images of objectified women: the model's recognition of emotional state is mediated by whether the subject is fully or partially clothed. Embedding association tests (EATs) return significant effect sizes for both anger (d >0.80) and sadness (d >0.50), associating images of fully clothed subjects with emotions. GRAD-CAM saliency maps highlight that CLIP gets distracted from emotional expressions in objectified images. A second experiment measures the effect in a representative application: an automatic image captioner (Antarctic Captions) includes words denoting emotion less than 50% as often for images of partially clothed women than for images of fully clothed women. A third experiment finds that images of female professionals (scientists, doctors, executives) are likely to be associated with sexual descriptions relative to images of male professionals. A fourth experiment shows that a prompt of "a [age] year old girl" generates sexualized images (as determined by an NSFW classifier) up to 73% of the time for VQGAN-CLIP and Stable Diffusion; the corresponding rate for boys never surpasses 9%. The evidence indicates that language-vision AI models trained on web scrapes learn biases of sexual objectification, which propagate to downstream applications.

Robert Wolfe, Aylin Caliskan

Three state-of-the-art language-and-image AI models, CLIP, SLIP, and BLIP, are evaluated for evidence of a bias previously observed in social and experimental psychology: equating American identity with being White. Embedding association tests (EATs) using standardized images of self-identified Asian, Black, Latina/o, and White individuals from the Chicago Face Database (CFD) reveal that White individuals are more associated with collective in-group words than are Asian, Black, or Latina/o individuals. In assessments of three core aspects of American identity reported by social psychologists, single-category EATs reveal that images of White individuals are more associated with patriotism and with being born in America, but that, consistent with prior findings in psychology, White individuals are associated with being less likely to treat people of all races and backgrounds equally. Three downstream machine learning tasks demonstrate biases associating American with White. In a visual question answering task using BLIP, 97% of White individuals are identified as American, compared to only 3% of Asian individuals. When asked in what state the individual depicted lives in, the model responds China 53% of the time for Asian individuals, but always with an American state for White individuals. In an image captioning task, BLIP remarks upon the race of Asian individuals as much as 36% of the time, but never remarks upon race for White individuals. Finally, provided with an initialization image from the CFD and the text "an American person," a synthetic image generator (VQGAN) using the text-based guidance of CLIP lightens the skin tone of individuals of all races (by 35% for Black individuals, based on pixel brightness). The results indicate that biases equating American identity with being White are learned by language-and-image AI, and propagate to downstream applications of such models.

Inyoung Cheong, Aylin Caliskan, Tadayoshi Kohno

Our interdisciplinary study investigates how effectively U.S. laws confront the challenges posed by Generative AI to human values. Through an analysis of diverse hypothetical scenarios crafted during an expert workshop, we have identified notable gaps and uncertainties within the existing legal framework regarding the protection of fundamental values, such as privacy, autonomy, dignity, diversity, equity, and physical/mental well-being. Constitutional and civil rights, it appears, may not provide sufficient protection against AI-generated discriminatory outputs. Furthermore, even if we exclude the liability shield provided by Section 230, proving causation for defamation and product liability claims is a challenging endeavor due to the intricate and opaque nature of AI systems. To address the unique and unforeseeable threats posed by Generative AI, we advocate for legal frameworks that evolve to recognize new threats and provide proactive, auditable guidelines to industry stakeholders. Addressing these issues requires deep interdisciplinary collaborations to identify harms, values, and mitigation strategies.

Robert Wolfe, Aylin Caliskan

We evaluate the state-of-the-art multimodal "visual semantic" model CLIP ("Contrastive Language Image Pretraining") for biases related to the marking of age, gender, and race or ethnicity. Given the option to label an image as "a photo of a person" or to select a label denoting race or ethnicity, CLIP chooses the "person" label 47.9% of the time for White individuals, compared with 5.0% or less for individuals who are Black, East Asian, Southeast Asian, Indian, or Latino or Hispanic. The model is more likely to rank the unmarked "person" label higher than labels denoting gender for Male individuals (26.7% of the time) vs. Female individuals (15.2% of the time). Age affects whether an individual is marked by the model: Female individuals under the age of 20 are more likely than Male individuals to be marked with a gender label, but less likely to be marked with an age label, while Female individuals over the age of 40 are more likely to be marked based on age than Male individuals. We also examine the self-similarity (mean pairwise cosine similarity) for each social group, where higher self-similarity denotes greater attention directed by CLIP to the shared characteristics (age, race, or gender) of the social group. As age increases, the self-similarity of representations of Female individuals increases at a higher rate than for Male individuals, with the disparity most pronounced at the "more than 70" age range. All ten of the most self-similar social groups are individuals under the age of 10 or over the age of 70, and six of the ten are Female individuals. Existing biases of self-similarity and markedness between Male and Female gender groups are further exacerbated when the groups compared are individuals who are White and Male and individuals who are Black and Female. Results indicate that CLIP reflects the biases of the language and society which produced its training data.

Shiva Omrani Sabbaghi, Robert Wolfe, Aylin Caliskan

Language models are trained on large-scale corpora that embed implicit biases documented in psychology. Valence associations (pleasantness/unpleasantness) of social groups determine the biased attitudes towards groups and concepts in social cognition. Building on this established literature, we quantify how social groups are valenced in English language models using a sentence template that provides an intersectional context. We study biases related to age, education, gender, height, intelligence, literacy, race, religion, sex, sexual orientation, social class, and weight. We present a concept projection approach to capture the valence subspace through contextualized word embeddings of language models. Adapting the projection-based approach to embedding association tests that quantify bias, we find that language models exhibit the most biased attitudes against gender identity, social class, and sexual orientation signals in language. We find that the largest and better-performing model that we study is also more biased as it effectively captures bias embedded in sociocultural data. We validate the bias evaluation method by overperforming on an intrinsic valence evaluation task. The approach enables us to measure complex intersectional biases as they are known to manifest in the outputs and applications of language models that perpetuate historical biases. Moreover, our approach contributes to design justice as it studies the associations of groups underrepresented in language such as transgender and homosexual individuals.

Shiva Omrani Sabbaghi, Aylin Caliskan

Does the grammatical gender of a language interfere when measuring the semantic gender information captured by its word embeddings? A number of anomalous gender bias measurements in the embeddings of gendered languages suggest this possibility. We demonstrate that word embeddings learn the association between a noun and its grammatical gender in grammatically gendered languages, which can skew social gender bias measurements. Consequently, word embedding post-processing methods are introduced to quantify, disentangle, and evaluate grammatical gender signals. The evaluation is performed on five gendered languages from the Germanic, Romance, and Slavic branches of the Indo-European language family. Our method reduces the strength of grammatical gender signals, which is measured in terms of effect size (Cohen's d), by a significant average of d = 1.3 for French, German, and Italian, and d = 0.56 for Polish and Spanish. Once grammatical gender is disentangled, the association between over 90% of 10,000 inanimate nouns and their assigned grammatical gender weakens, and cross-lingual bias results from the Word Embedding Association Test (WEAT) become more congruent with country-level implicit bias measurements. The results further suggest that disentangling grammatical gender signals from word embeddings may lead to improvement in semantic machine learning tasks.

Robert Wolfe, Aylin Caliskan

VAST, the Valence-Assessing Semantics Test, is a novel intrinsic evaluation task for contextualized word embeddings (CWEs). VAST uses valence, the association of a word with pleasantness, to measure the correspondence of word-level LM semantics with widely used human judgments, and examines the effects of contextualization, tokenization, and LM-specific geometry. Because prior research has found that CWEs from GPT-2 perform poorly on other intrinsic evaluations, we select GPT-2 as our primary subject, and include results showing that VAST is useful for 7 other LMs, and can be used in 7 languages. GPT-2 results show that the semantics of a word incorporate the semantics of context in layers closer to model output, such that VAST scores diverge between our contextual settings, ranging from Pearson's rho of .55 to .77 in layer 11. We also show that multiply tokenized words are not semantically encoded until layer 8, where they achieve Pearson's rho of .46, indicating the presence of an encoding process for multiply tokenized words which differs from that of singly tokenized words, for which rho is highest in layer 0. We find that a few neurons with values having greater magnitude than the rest mask word-level semantics in GPT-2's top layer, but that word-level semantics can be recovered by nullifying non-semantic principal components: Pearson's rho in the top layer improves from .32 to .76. After isolating semantics, we show the utility of VAST for understanding LM semantics via improvements over related work on four word similarity tasks, with a score of .50 on SimLex-999, better than the previous best of .45 for GPT-2. Finally, we show that 8 of 10 WEAT bias tests, which compare differences in word embedding associations between groups of words, exhibit more stereotype-congruent biases after isolating semantics, indicating that non-semantic structures in LMs also mask biases.

Robert Wolfe, Aylin Caliskan

We examine the effects of contrastive visual semantic pretraining by comparing 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 GPT-2 architecture to encode image captions. We find that contrastive visual semantic pretraining significantly mitigates the anisotropy found in contextualized word embeddings from GPT-2, such that the intra-layer self-similarity (mean pairwise cosine similarity) of CLIP word embeddings is under .25 in all layers, compared to greater than .95 in the top layer of GPT-2. CLIP word embeddings outperform GPT-2 on word-level semantic intrinsic evaluation tasks, and achieve a new corpus-based state of the art for the RG65 evaluation, at .88. CLIP also forms fine-grained semantic representations of sentences, and obtains Spearman's rho = .73 on the SemEval-2017 Semantic Textual Similarity Benchmark with no fine-tuning, compared to no greater than rho = .45 in any layer of GPT-2. Finally, intra-layer self-similarity of CLIP sentence embeddings decreases as the layer index increases, finishing at .25 in the top layer, while the self-similarity of GPT-2 sentence embeddings formed using the EOS token increases layer-over-layer and never falls below .97. Our results indicate that high anisotropy is not an inevitable consequence of contextualization, and that visual semantic pretraining is beneficial not only for ordering visual representations, but also for encoding useful semantic representations of language, both on the word level and the sentence level.

Sourojit Ghosh, Pranav Narayanan Venkit, Sanjana Gautam et al.

Our research investigates the impact of Generative Artificial Intelligence (GAI) models, specifically text-to-image generators (T2Is), on the representation of non-Western cultures, with a focus on Indian contexts. Despite the transformative potential of T2Is in content creation, concerns have arisen regarding biases that may lead to misrepresentations and marginalizations. Through a community-centered approach and grounded theory analysis of 5 focus groups from diverse Indian subcultures, we explore how T2I outputs to English prompts depict Indian culture and its subcultures, uncovering novel representational harms such as exoticism and cultural misappropriation. These findings highlight the urgent need for inclusive and culturally sensitive T2I systems. We propose design guidelines informed by a sociotechnical perspective, aiming to address these issues and contribute to the development of more equitable and representative GAI technologies globally. Our work also underscores the necessity of adopting a community-centered approach to comprehend the sociotechnical dynamics of these models, complementing existing work in this space while identifying and addressing the potential negative repercussions and harms that may arise when these models are deployed on a global scale.

Isaac Slaughter, Craig Greenberg, Reva Schwartz et al. · uw

Previous work has established that a person's demographics and speech style affect how well speech processing models perform for them. But where does this bias come from? In this work, we present the Speech Embedding Association Test (SpEAT), a method for detecting bias in one type of model used for many speech tasks: pre-trained models. The SpEAT is inspired by word embedding association tests in natural language processing, which quantify intrinsic bias in a model's representations of different concepts, such as race or valence (something's pleasantness or unpleasantness) and capture the extent to which a model trained on large-scale socio-cultural data has learned human-like biases. Using the SpEAT, we test for six types of bias in 16 English speech models (including 4 models also trained on multilingual data), which come from the wav2vec 2.0, HuBERT, WavLM, and Whisper model families. We find that 14 or more models reveal positive valence (pleasantness) associations with abled people over disabled people, with European-Americans over African-Americans, with females over males, with U.S. accented speakers over non-U.S. accented speakers, and with younger people over older people. Beyond establishing that pre-trained speech models contain these biases, we also show that they can have real world effects. We compare biases found in pre-trained models to biases in downstream models adapted to the task of Speech Emotion Recognition (SER) and find that in 66 of the 96 tests performed (69%), the group that is more associated with positive valence as indicated by the SpEAT also tends to be predicted as speaking with higher valence by the downstream model. Our work provides evidence that, like text and image-based models, pre-trained speech based-models frequently learn human-like biases. Our work also shows that bias found in pre-trained models can propagate to the downstream task of SER.

Sourojit Ghosh, Aylin Caliskan

We study stereotypes embedded within one of the most popular text-to-image generators: Stable Diffusion. We examine what stereotypes of gender and nationality/continental identity does Stable Diffusion display in the absence of such information i.e. what gender and nationality/continental identity is assigned to `a person', or to `a person from Asia'. Using vision-language model CLIP's cosine similarity to compare images generated by CLIP-based Stable Diffusion v2.1 verified by manual examination, we chronicle results from 136 prompts (50 results/prompt) of front-facing images of persons from 6 different continents, 27 nationalities and 3 genders. We observe how Stable Diffusion outputs of `a person' without any additional gender/nationality information correspond closest to images of men and least with persons of nonbinary gender, and to persons from Europe/North America over Africa/Asia, pointing towards Stable Diffusion having a concerning representation of personhood to be a European/North American man. We also show continental stereotypes and resultant harms e.g. a person from Oceania is deemed to be Australian/New Zealander over Papua New Guinean, pointing to the erasure of Indigenous Oceanic peoples, who form a majority over descendants of colonizers both in Papua New Guinea and in Oceania overall. Finally, we unexpectedly observe a pattern of oversexualization of women, specifically Latin American, Mexican, Indian and Egyptian women relative to other nationalities, measured through an NSFW detector. This demonstrates how Stable Diffusion perpetuates Western fetishization of women of color through objectification in media, which if left unchecked will amplify this stereotypical representation. Image datasets are made publicly available.

Kyra Wilson, Aylin Caliskan

Artificial intelligence (AI) hiring tools have revolutionized resume screening, and large language models (LLMs) have the potential to do the same. However, given the biases which are embedded within LLMs, it is unclear whether they can be used in this scenario without disadvantaging groups based on their protected attributes. In this work, we investigate the possibilities of using LLMs in a resume screening setting via a document retrieval framework that simulates job candidate selection. Using that framework, we then perform a resume audit study to determine whether a selection of Massive Text Embedding (MTE) models are biased in resume screening scenarios. We simulate this for nine occupations, using a collection of over 500 publicly available resumes and 500 job descriptions. We find that the MTEs are biased, significantly favoring White-associated names in 85.1\% of cases and female-associated names in only 11.1\% of cases, with a minority of cases showing no statistically significant differences. Further analyses show that Black males are disadvantaged in up to 100\% of cases, replicating real-world patterns of bias in employment settings, and validate three hypotheses of intersectionality. We also find an impact of document length as well as the corpus frequency of names in the selection of resumes. These findings have implications for widely used AI tools that are automating employment, fairness, and tech policy.

Gandalf Nicolas, Aylin Caliskan

This study introduces a taxonomy of stereotype content in contemporary large language models (LLMs). We prompt ChatGPT 3.5, Llama 3, and Mixtral 8x7B, three powerful and widely used LLMs, for the characteristics associated with 87 social categories (e.g., gender, race, occupations). We identify 14 stereotype dimensions (e.g., Morality, Ability, Health, Beliefs, Emotions), accounting for ~90% of LLM stereotype associations. Warmth and Competence facets were the most frequent content, but all other dimensions were significantly prevalent. Stereotypes were more positive in LLMs (vs. humans), but there was significant variability across categories and dimensions. Finally, the taxonomy predicted the LLMs' internal evaluations of social categories (e.g., how positively/negatively the categories were represented), supporting the relevance of a multidimensional taxonomy for characterizing LLM stereotypes. Our findings suggest that high-dimensional human stereotypes are reflected in LLMs and must be considered in AI auditing and debiasing to minimize unidentified harms from reliance in low-dimensional views of bias in LLMs.

Saloni Dash, Amélie Reymond, Emma S. Spiro et al. · uw

Reasoning in humans is prone to biases due to underlying motivations like identity protection, that undermine rational decision-making and judgment. This motivated reasoning at a collective level can be detrimental to society when debating critical issues such as human-driven climate change or vaccine safety, and can further aggravate political polarization. Prior studies have reported that large language models (LLMs) are also susceptible to human-like cognitive biases, however, the extent to which LLMs selectively reason toward identity-congruent conclusions remains largely unexplored. Here, we investigate whether assigning 8 personas across 4 political and socio-demographic attributes induces motivated reasoning in LLMs. Testing 8 LLMs (open source and proprietary) across two reasoning tasks from human-subject studies -- veracity discernment of misinformation headlines and evaluation of numeric scientific evidence -- we find that persona-assigned LLMs have up to 9% reduced veracity discernment relative to models without personas. Political personas specifically, are up to 90% more likely to correctly evaluate scientific evidence on gun control when the ground truth is congruent with their induced political identity. Prompt-based debiasing methods are largely ineffective at mitigating these effects. Taken together, our empirical findings are the first to suggest that persona-assigned LLMs exhibit human-like motivated reasoning that is hard to mitigate through conventional debiasing prompts -- raising concerns of exacerbating identity-congruent reasoning in both LLMs and humans.

Harshita Chopra, Kshitish Ghate, Aylin Caliskan et al.

Large Language Model (LLM) agents are increasingly deployed in settings where they interact with a wide variety of people, including users who are unclear, impatient, or reluctant to share information. However, collecting real interaction data at scale remains expensive. The field has turned to LLM-based user simulators as stand-ins, but these simulators inherit the behavior of their underlying models: cooperative and homogeneous. As a result, agents that appear strong in simulation often fail under the unseen, diverse communication patterns of real users. To narrow this gap, we introduce Persona Policies (PPol), a plug-and-play control layer that induces realistic behavioral variation in user simulators while preserving the original task goals. Rather than hand-crafting personas, we cast persona generation as an LLM-driven evolutionary program search that optimizes a Python generator to discover behaviors and translate them into task-preserving roleplay policies. Candidate generators are guided by a multi-objective fitness score combining human-likeness with broad coverage of human behavioral patterns. Once optimized, the generator produces a diverse population of human-like personas for any task in the domain. Across tau^2-bench retail and airline domains, evolved PPol programs yield 33-62% absolute gains in fitness score over the baseline simulator. In a blinded evaluation, annotators rated PPol-conditioned users as human 80.4% of the time, close to real human traces and nearly twice as frequently as baseline simulators. Agents trained with PPol are more robust to challenging, out-of-distribution behaviors, improving task success by +17% relative to training only on existing simulated interactions. This offers a novel approach to strengthen simulator-based evaluation and training without changing tasks or rewards.

Dean Light, Michael Theologitis, Kshitish Ghate et al.

Humans intuitively solve complex problems by flexibly shifting among reasoning modes: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems. Current LLM agents lack this flexibility, as their scaffolds hard-code such reasoning decisions in advance. These scaffolds are effective when their prescribed structure matches the task, but brittle when solving the task requires adapting the structure of reasoning itself. We introduce Deep Reasoning -- an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning. Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction. We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across more controlled reasoning threads. We evaluate it against state-of-the-art scaffolding methods across four hard benchmarks: multi-hop reasoning, long-chain question answering, long-context aggregation, and deep research-style information seeking. DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average. DOLORES distributes cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucinations. This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings. These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.

Kshitish Ghate, Andy Liu, Devansh Jain et al. · allen-ai, cmu

As large language models (LLMs) are deployed globally, creating pluralistic systems that can accommodate the diverse preferences and values of users worldwide becomes essential. We introduce EVALUESTEER, a benchmark to measure LLMs' and reward models' (RMs) steerability towards users' value and stylistic preference profiles grounded in psychology and human-LLM interaction literature. To address the gap in existing datasets that do not support controlled evaluations of RM steering, we synthetically generated 165,888 preference pairs -- systematically varying pairs along 4 value dimensions (traditional, secular-rational, survival, and self-expression) and 4 style dimensions (verbosity, readability, confidence, and warmth). We use EVALUESTEER to evaluate whether, given a user profile and a pair of candidate value-laden and style-laden responses, LLMs and RMs are able to select the output that aligns with the user's preferences. We evaluate six open-source and proprietary LLMs and RMs under eleven systematic prompting conditions and six preference comparison scenarios. Notably, our results show that, when given the user's full profile of values and stylistic preferences, the best models achieve <75% accuracy at choosing the correct response, in contrast to >99% accuracy when only relevant style and value preferences are provided. EVALUESTEER thus highlights the limitations of current RMs at identifying and adapting to relevant user profile information, and provides a challenging testbed for developing RMs that can be steered towards diverse human values and preferences.

Edwin Dauber, Aylin Caliskan, Richard Harang et al.

Program authorship attribution has implications for the privacy of programmers who wish to contribute code anonymously. While previous work has shown that complete files that are individually authored can be attributed, we show here for the first time that accounts belonging to open source contributors containing short, incomplete, and typically uncompilable fragments can also be effectively attributed. We propose a technique for authorship attribution of contributor accounts containing small source code samples, such as those that can be obtained from version control systems or other direct comparison of sequential versions. We show that while application of previous methods to individual small source code samples yields an accuracy of about 73% for 106 programmers as a baseline, by ensembling and averaging the classification probabilities of a sufficiently large set of samples belonging to the same author we achieve 99% accuracy for assigning the set of samples to the correct author. Through these results, we demonstrate that attribution is an important threat to privacy for programmers even in real-world collaborative environments such as GitHub. Additionally, we propose the use of calibration curves to identify samples by unknown and previously unencountered authors in the open world setting. We show that we can also use these calibration curves in the case that we do not have linking information and thus are forced to classify individual samples directly. This is because the calibration curves allow us to identify which samples are more likely to have been correctly attributed. Using such a curve can help an analyst choose a cut-off point which will prevent most misclassifications, at the cost of causing the rejection of some of the more dubious correct attributions.

Anna-Maria Gueorguieva, Aylin Caliskan

Large language models (LLMs) have been shown to exhibit social bias, however, bias towards non-protected stigmatized identities remain understudied. Furthermore, what social features of stigmas are associated with bias in LLM outputs is unknown. From psychology literature, it has been shown that stigmas contain six shared social features: aesthetics, concealability, course, disruptiveness, origin, and peril. In this study, we investigate if human and LLM ratings of the features of stigmas, along with prompt style and type of stigma, have effect on bias towards stigmatized groups in LLM outputs. We measure bias against 93 stigmatized groups across three widely used LLMs (Granite 3.0-8B, Llama-3.1-8B, Mistral-7B) using SocialStigmaQA, a benchmark that includes 37 social scenarios about stigmatized identities; for example deciding wether to recommend them for an internship. We find that stigmas rated by humans to be highly perilous (e.g., being a gang member or having HIV) have the most biased outputs from SocialStigmaQA prompts (60% of outputs from all models) while sociodemographic stigmas (e.g. Asian-American or old age) have the least amount of biased outputs (11%). We test if the amount of biased outputs could be decreased by using guardrail models, models meant to identify harmful input, using each LLM's respective guardrail model (Granite Guardian 3.0, Llama Guard 3.0, Mistral Moderation API). We find that bias decreases significantly by 10.4%, 1.4%, and 7.8%, respectively. However, we show that features with significant effect on bias remain unchanged post-mitigation and that guardrail models often fail to recognize the intent of bias in prompts. This work has implications for using LLMs in scenarios involving stigmatized groups and we suggest future work towards improving guardrail models for bias mitigation.

Yongxu Sun, Michael Saxon, Ian Yang et al.

Text-to-Video (T2V) generators such as Sora raise concerns about whether generated content reflects societal bias. We extend embedding-association tests from words and images to video by introducing the Video Embedding Association Test (VEAT) and Single-Category VEAT (SC-VEAT). We validate these methods by reproducing the direction and magnitude of associations from widely used baselines, including Implicit Association Test (IAT) scenarios and OASIS image categories. We then quantify race (African American vs. European American) and gender (women vs. men) associations with valence (pleasant vs. unpleasant) across 17 occupations and 7 awards. Sora videos associate European Americans and women more with pleasantness (both d>0.8). Effect sizes correlate with real-world demographic distributions: percent men and White in occupations (r=0.93, r=0.83) and percent male and non-Black among award recipients (r=0.88, r=0.99). Applying explicit debiasing prompts generally reduces effect-size magnitudes, but can backfire: two Black-associated occupations (janitor, postal service) become more Black-associated after debiasing. Together, these results reveal that easily accessible T2V generators can actually amplify representational harms if not rigorously evaluated and responsibly deployed.

Kshitish Ghate, Isaac Slaughter, Kyra Wilson et al. · uw

While recent work has found that vision-language models trained under the Contrastive Language Image Pre-training (CLIP) framework contain intrinsic social biases, the extent to which different upstream pre-training features of the framework relate to these biases, and hence how intrinsic bias and downstream performance are connected has been unclear. In this work, we present the largest comprehensive analysis to-date of how the upstream pre-training factors and downstream performance of CLIP models relate to their intrinsic biases. Studying 131 unique CLIP models, trained on 26 datasets, using 55 architectures, and in a variety of sizes, we evaluate bias in each model using 26 well-established unimodal and cross-modal principled Embedding Association Tests. We find that the choice of pre-training dataset is the most significant upstream predictor of bias, whereas architectural variations have minimal impact. Additionally, datasets curated using sophisticated filtering techniques aimed at enhancing downstream model performance tend to be associated with higher levels of intrinsic bias. Finally, we observe that intrinsic bias is often significantly correlated with downstream performance ($0.3 \leq r \leq 0.8$), suggesting that models optimized for performance inadvertently learn to amplify representational biases. Comparisons between unimodal and cross-modal association tests reveal that social group bias depends heavily on the modality. Our findings imply that more sophisticated strategies are needed to address intrinsic model bias for vision-language models across the entire model development pipeline.

Chahat Raj, Bowen Wei, Aylin Caliskan et al.

While bias in large language models (LLMs) is well-studied, similar concerns in vision-language models (VLMs) have received comparatively less attention. Existing VLM bias studies often focus on portrait-style images and gender-occupation associations, overlooking broader and more complex social stereotypes and their implied harm. This work introduces VIGNETTE, a large-scale VQA benchmark with 30M+ images for evaluating bias in VLMs through a question-answering framework spanning four directions: factuality, perception, stereotyping, and decision making. Beyond narrowly-centered studies, we assess how VLMs interpret identities in contextualized settings, revealing how models make trait and capability assumptions and exhibit patterns of discrimination. Drawing from social psychology, we examine how VLMs connect visual identity cues to trait and role-based inferences, encoding social hierarchies, through biased selections. Our findings uncover subtle, multifaceted, and surprising stereotypical patterns, offering insights into how VLMs construct social meaning from inputs.

Yuyang Jiang, Longjie Guo, Yuchen Wu et al.

Large language model (LLM)-powered chatbots are increasingly used for opinion exploration. Prior research examined how LLMs alter user views, yet little work extended beyond one-way influence to address how user input can affect LLM responses and how such bi-directional influence manifests throughout the multi-turn conversations. This study investigates this dynamic through 50 controversial-topic discussions with participants (N=266) across three conditions: static statements, standard chatbot, and personalized chatbot. Results show that human opinions barely shifted, while LLM outputs changed more substantially, narrowing the gap between human and LLM stance. Personalization amplified these shifts in both directions compared to the standard setting. Analysis of multi-turn conversations further revealed that exchanges involving participants' personal stories were most likely to trigger stance changes for both humans and LLMs. Our work highlights the risk of over-alignment in human-LLM interaction and the need for careful design of personalized chatbots to more thoughtfully and stably align with users.

Kyra Wilson, Mattea Sim, Anna-Maria Gueorguieva et al.

In this study, we conduct a resume-screening experiment (N=528) where people collaborate with simulated AI models exhibiting race-based preferences (bias) to evaluate candidates for 16 high and low status occupations. Simulated AI bias approximates factual and counterfactual estimates of racial bias in real-world AI systems. We investigate people's preferences for White, Black, Hispanic, and Asian candidates (represented through names and affinity groups on quality-controlled resumes) across 1,526 scenarios and measure their unconscious associations between race and status using implicit association tests (IATs), which predict discriminatory hiring decisions but have not been investigated in human-AI collaboration. When making decisions without AI or with AI that exhibits no race-based preferences, people select all candidates at equal rates. However, when interacting with AI favoring a particular group, people also favor those candidates up to 90% of the time, indicating a significant behavioral shift. The likelihood of selecting candidates whose identities do not align with common race-status stereotypes can increase by 13% if people complete an IAT before conducting resume screening. Finally, even if people think AI recommendations are low quality or not important, their decisions are still vulnerable to AI bias under certain circumstances. This work has implications for people's autonomy in AI-HITL scenarios, AI and work, design and evaluation of AI hiring systems, and strategies for mitigating bias in collaborative decision-making tasks. In particular, organizational and regulatory policy should acknowledge the complex nature of AI-HITL decision making when implementing these systems, educating people who use them, and determining which are subject to oversight.

Kyra Wilson, Sourojit Ghosh, Aylin Caliskan

Text-to-image generators (T2Is) are liable to produce images that perpetuate social stereotypes, especially in regards to race or skin tone. We use a comprehensive set of 93 stigmatized identities to determine that three versions of Stable Diffusion (v1.5, v2.1, and XL) systematically associate stigmatized identities with certain skin tones in generated images. We find that SD XL produces skin tones that are 13.53% darker and 23.76% less red (both of which indicate higher likelihood of societal discrimination) than previous models and perpetuate societal stereotypes associating people of color with stigmatized identities. SD XL also shows approximately 30% less variability in skin tones when compared to previous models and 18.89-56.06% compared to human face datasets. Measuring variability through metrics which directly correspond to human perception suggest a similar pattern, where SD XL shows the least amount of variability in skin tones of people with stigmatized identities and depicts most (60.29%) stigmatized identities as being less diverse than non-stigmatized identities. Finally, SD shows more homogenization of skin tones of racial and ethnic identities compared to other stigmatized or non-stigmatized identities, reinforcing incorrect equivalence of biologically-determined skin tone and socially-constructed racial and ethnic identity. Because SD XL is the largest and most complex model and users prefer its generations compared to other models examined in this study, these findings have implications for the dynamics of bias amplification in T2Is, increasing representational harms and challenges generating diverse images depicting people with stigmatized identities.

Chahat Raj, Mahika Banerjee, Aylin Caliskan et al.

When a student fails an exam, do we tend to blame their effort or the test's difficulty? Attribution, defined as how reasons are assigned to event outcomes, shapes perceptions, reinforces stereotypes, and influences decisions. Attribution Theory in social psychology explains how humans assign responsibility for events using implicit cognition, attributing causes to internal (e.g., effort, ability) or external (e.g., task difficulty, luck) factors. LLMs' attribution of event outcomes based on demographics carries important fairness implications. Most works exploring social biases in LLMs focus on surface-level associations or isolated stereotypes. This work proposes a cognitively grounded bias evaluation framework to identify how models' reasoning disparities channelize biases toward demographic groups.

Steven A. Lehr, Aylin Caliskan, Suneragiri Liyanage et al.

How good a research scientist is ChatGPT? We systematically probed the capabilities of GPT-3.5 and GPT-4 across four central components of the scientific process: as a Research Librarian, Research Ethicist, Data Generator, and Novel Data Predictor, using psychological science as a testing field. In Study 1 (Research Librarian), unlike human researchers, GPT-3.5 and GPT-4 hallucinated, authoritatively generating fictional references 36.0% and 5.4% of the time, respectively, although GPT-4 exhibited an evolving capacity to acknowledge its fictions. In Study 2 (Research Ethicist), GPT-4 (though not GPT-3.5) proved capable of detecting violations like p-hacking in fictional research protocols, correcting 88.6% of blatantly presented issues, and 72.6% of subtly presented issues. In Study 3 (Data Generator), both models consistently replicated patterns of cultural bias previously discovered in large language corpora, indicating that ChatGPT can simulate known results, an antecedent to usefulness for both data generation and skills like hypothesis generation. Contrastingly, in Study 4 (Novel Data Predictor), neither model was successful at predicting new results absent in their training data, and neither appeared to leverage substantially new information when predicting more versus less novel outcomes. Together, these results suggest that GPT is a flawed but rapidly improving librarian, a decent research ethicist already, capable of data generation in simple domains with known characteristics but poor at predicting novel patterns of empirical data to aid future experimentation.

Sourojit Ghosh, Aylin Caliskan

In this multicultural age, language translation is one of the most performed tasks, and it is becoming increasingly AI-moderated and automated. As a novel AI system, ChatGPT claims to be proficient in such translation tasks and in this paper, we put that claim to the test. Specifically, we examine ChatGPT's accuracy in translating between English and languages that exclusively use gender-neutral pronouns. We center this study around Bengali, the 7$^{th}$ most spoken language globally, but also generalize our findings across five other languages: Farsi, Malay, Tagalog, Thai, and Turkish. We find that ChatGPT perpetuates gender defaults and stereotypes assigned to certain occupations (e.g. man = doctor, woman = nurse) or actions (e.g. woman = cook, man = go to work), as it converts gender-neutral pronouns in languages to `he' or `she'. We also observe ChatGPT completely failing to translate the English gender-neutral pronoun `they' into equivalent gender-neutral pronouns in other languages, as it produces translations that are incoherent and incorrect. While it does respect and provide appropriately gender-marked versions of Bengali words when prompted with gender information in English, ChatGPT appears to confer a higher respect to men than to women in the same occupation. We conclude that ChatGPT exhibits the same gender biases which have been demonstrated for tools like Google Translate or MS Translator, as we provide recommendations for a human centered approach for future designers of AIs that perform language translation to better accommodate such low-resource languages.

Robert Wolfe, Aylin Caliskan

We use a dataset of U.S. first names with labels based on predominant gender and racial group to examine the effect of training corpus frequency on tokenization, contextualization, similarity to initial representation, and bias in BERT, GPT-2, T5, and XLNet. We show that predominantly female and non-white names are less frequent in the training corpora of these four language models. We find that infrequent names are more self-similar across contexts, with Spearman's r between frequency and self-similarity as low as -.763. Infrequent names are also less similar to initial representation, with Spearman's r between frequency and linear centered kernel alignment (CKA) similarity to initial representation as high as .702. Moreover, we find Spearman's r between racial bias and name frequency in BERT of .492, indicating that lower-frequency minority group names are more associated with unpleasantness. Representations of infrequent names undergo more processing, but are more self-similar, indicating that models rely on less context-informed representations of uncommon and minority names which are overfit to a lower number of observed contexts.

Ryan Steed, Aylin Caliskan

Recent advances in machine learning leverage massive datasets of unlabeled images from the web to learn general-purpose image representations for tasks from image classification to face recognition. But do unsupervised computer vision models automatically learn implicit patterns and embed social biases that could have harmful downstream effects? We develop a novel method for quantifying biased associations between representations of social concepts and attributes in images. We find that state-of-the-art unsupervised models trained on ImageNet, a popular benchmark image dataset curated from internet images, automatically learn racial, gender, and intersectional biases. We replicate 8 previously documented human biases from social psychology, from the innocuous, as with insects and flowers, to the potentially harmful, as with race and gender. Our results closely match three hypotheses about intersectional bias from social psychology. For the first time in unsupervised computer vision, we also quantify implicit human biases about weight, disabilities, and several ethnicities. When compared with statistical patterns in online image datasets, our findings suggest that machine learning models can automatically learn bias from the way people are stereotypically portrayed on the web.

Akshat Pandey, Aylin Caliskan

Ridehailing applications that collect mobility data from individuals to inform smart city planning predict each trip's fare pricing with automated algorithms that rely on artificial intelligence (AI). This type of AI algorithm, namely a price discrimination algorithm, is widely used in the industry's black box systems for dynamic individualized pricing. Lacking transparency, studying such AI systems for fairness and disparate impact has not been possible without access to data used in generating the outcomes of price discrimination algorithms. Recently, in an effort to enhance transparency in city planning, the city of Chicago regulation mandated that transportation providers publish anonymized data on ridehailing. As a result, we present the first large-scale measurement of the disparate impact of price discrimination algorithms used by ridehailing applications. The application of random effects models from the meta-analysis literature combines the city-level effects of AI bias on fare pricing from census tract attributes, aggregated from the American Community Survey. An analysis of 100 million ridehailing samples from the city of Chicago indicates a significant disparate impact in fare pricing of neighborhoods due to AI bias learned from ridehailing utilization patterns associated with demographic attributes. Neighborhoods with larger non-white populations, higher poverty levels, younger residents, and high education levels are significantly associated with higher fare prices, with combined effect sizes, measured in Cohen's d, of -0.32, -0.28, 0.69, and 0.24 for each demographic, respectively. Further, our methods hold promise for identifying and addressing the sources of disparate impact in AI algorithms learning from datasets that contain U.S. geolocations.

Wei Guo, Aylin Caliskan

With the starting point that implicit human biases are reflected in the statistical regularities of language, it is possible to measure biases in English static word embeddings. State-of-the-art neural language models generate dynamic word embeddings dependent on the context in which the word appears. Current methods measure pre-defined social and intersectional biases that appear in particular contexts defined by sentence templates. Dispensing with templates, we introduce the Contextualized Embedding Association Test (CEAT), that can summarize the magnitude of overall bias in neural language models by incorporating a random-effects model. Experiments on social and intersectional biases show that CEAT finds evidence of all tested biases and provides comprehensive information on the variance of effect magnitudes of the same bias in different contexts. All the models trained on English corpora that we study contain biased representations. Furthermore, we develop two methods, Intersectional Bias Detection (IBD) and Emergent Intersectional Bias Detection (EIBD), to automatically identify the intersectional biases and emergent intersectional biases from static word embeddings in addition to measuring them in contextualized word embeddings. We present the first algorithmic bias detection findings on how intersectional group members are strongly associated with unique emergent biases that do not overlap with the biases of their constituent minority identities. IBD and EIBD achieve high accuracy when detecting the intersectional and emergent biases of African American females and Mexican American females. Our results indicate that biases at the intersection of race and gender associated with members of multiple minority groups, such as African American females and Mexican American females, have the highest magnitude across all neural language models.

Autumn Toney-Wails, Aylin Caliskan

Word embeddings learn implicit biases from linguistic regularities captured by word co-occurrence statistics. By extending methods that quantify human-like biases in word embeddings, we introduceValNorm, a novel intrinsic evaluation task and method to quantify the valence dimension of affect in human-rated word sets from social psychology. We apply ValNorm on static word embeddings from seven languages (Chinese, English, German, Polish, Portuguese, Spanish, and Turkish) and from historical English text spanning 200 years. ValNorm achieves consistently high accuracy in quantifying the valence of non-discriminatory, non-social group word sets. Specifically, ValNorm achieves a Pearson correlation of r=0.88 for human judgment scores of valence for 399 words collected to establish pleasantness norms in English. In contrast, we measure gender stereotypes using the same set of word embeddings and find that social biases vary across languages. Our results indicate that valence associations of non-discriminatory, non-social group words represent widely-shared associations, in seven languages and over 200 years.

Autumn Toney, Akshat Pandey, Wei Guo et al.

This paper considers the problem of automatically characterizing overall attitudes and biases that may be associated with emerging information operations via artificial intelligence. Accurate analysis of these emerging topics usually requires laborious, manual analysis by experts to annotate millions of tweets to identify biases in new topics. We introduce extensions of the Word Embedding Association Test from Caliskan et al. to a new domain (Caliskan, 2017). Our practical and unsupervised method is used to quantify biases promoted in information operations. We validate our method using known information operation-related tweets from Twitter's Transparency Report. We perform a case study on the COVID-19 pandemic to evaluate our method's performance on non-labeled Twitter data, demonstrating its usability in emerging domains.

Ryan Steed, Aylin Caliskan

Research in social psychology has shown that people's biased, subjective judgments about another's personality based solely on their appearance are not predictive of their actual personality traits. But researchers and companies often utilize computer vision models to predict similarly subjective personality attributes such as "employability." We seek to determine whether state-of-the-art, black box face processing technology can learn human-like appearance biases. With features extracted with FaceNet, a widely used face recognition framework, we train a transfer learning model on human subjects' first impressions of personality traits in other faces as measured by social psychologists. We find that features extracted with FaceNet can be used to predict human appearance bias scores for deliberately manipulated faces but not for randomly generated faces scored by humans. Additionally, in contrast to work with human biases in social psychology, the model does not find a significant signal correlating politicians' vote shares with perceived competence bias. With Local Interpretable Model-Agnostic Explanations (LIME), we provide several explanations for this discrepancy. Our results suggest that some signals of appearance bias documented in social psychology are not embedded by the machine learning techniques we investigate. We shed light on the ways in which appearance bias could be embedded in face processing technology and cast further doubt on the practice of predicting subjective traits based on appearances.

Aylin Caliskan, Joanna J. Bryson, Arvind Narayanan

Artificial intelligence and machine learning are in a period of astounding growth. However, there are concerns that these technologies may be used, either with or without intention, to perpetuate the prejudice and unfairness that unfortunately characterizes many human institutions. Here we show for the first time that human-like semantic biases result from the application of standard machine learning to ordinary language---the same sort of language humans are exposed to every day. We replicate a spectrum of standard human biases as exposed by the Implicit Association Test and other well-known psychological studies. We replicate these using a widely used, purely statistical machine-learning model---namely, the GloVe word embedding---trained on a corpus of text from the Web. Our results indicate that language itself contains recoverable and accurate imprints of our historic biases, whether these are morally neutral as towards insects or flowers, problematic as towards race or gender, or even simply veridical, reflecting the {\em status quo} for the distribution of gender with respect to careers or first names. These regularities are captured by machine learning along with the rest of semantics. In addition to our empirical findings concerning language, we also contribute new methods for evaluating bias in text, the Word Embedding Association Test (WEAT) and the Word Embedding Factual Association Test (WEFAT). Our results have implications not only for AI and machine learning, but also for the fields of psychology, sociology, and human ethics, since they raise the possibility that mere exposure to everyday language can account for the biases we replicate here.

Aylin Caliskan, Fabian Yamaguchi, Edwin Dauber et al.

The ability to identify authors of computer programs based on their coding style is a direct threat to the privacy and anonymity of programmers. While recent work found that source code can be attributed to authors with high accuracy, attribution of executable binaries appears to be much more difficult. Many distinguishing features present in source code, e.g. variable names, are removed in the compilation process, and compiler optimization may alter the structure of a program, further obscuring features that are known to be useful in determining authorship. We examine programmer de-anonymization from the standpoint of machine learning, using a novel set of features that include ones obtained by decompiling the executable binary to source code. We adapt a powerful set of techniques from the domain of source code authorship attribution along with stylistic representations embedded in assembly, resulting in successful de-anonymization of a large set of programmers. We evaluate our approach on data from the Google Code Jam, obtaining attribution accuracy of up to 96% with 100 and 83% with 600 candidate programmers. We present an executable binary authorship attribution approach, for the first time, that is robust to basic obfuscations, a range of compiler optimization settings, and binaries that have been stripped of their symbol tables. We perform programmer de-anonymization using both obfuscated binaries, and real-world code found "in the wild" in single-author GitHub repositories and the recently leaked Nulled.IO hacker forum. We show that programmers who would like to remain anonymous need to take extreme countermeasures to protect their privacy.