Search for: All records

While advances in fairness and alignment have helped mitigate overt biases exhibited by large language models (LLMs) when explicitly prompted, we hypothesize that these models may still exhibit implicit biases when simulating human behavior. To test this hypothesis, we propose a technique to systematically uncover such biases across a broad range of sociodemographic categories by assessing decision-making disparities among agents with LLM-generated, sociodemographically-informed personas. Using our technique, we tested six LLMs across three sociodemographic groups and four decision-making scenarios. Our results show that state-of-the-art LLMs exhibit significant sociodemographic disparities in nearly all simulations, with more advanced models exhibiting greater implicit biases despite reducing explicit biases. Furthermore, when comparing our findings to real-world disparities reported in empirical studies, we find that the biases we uncovered are directionally aligned but markedly amplified. This directional alignment highlights the utility of our technique in uncovering systematic biases in LLMs rather than random variations; moreover, the presence and amplification of implicit biases emphasizes the need for novel strategies to address these biases. 

How do practitioners who develop consumer AI products scope, motivate, and conduct privacy work? Respecting pri- vacy is a key principle for developing ethical, human-centered AI systems, but we cannot hope to better support practitioners without answers to that question. We interviewed 35 industry AI practitioners to bridge that gap. We found that practitioners viewed privacy as actions taken against pre-defined intrusions that can be exacerbated by the capabilities and requirements of AI, but few were aware of AI-specific privacy intrusions documented in prior literature. We found that their privacy work was rigidly defined and situated, guided by compliance with privacy regulations and policies, and generally demoti- vated beyond meeting minimum requirements. Finally, we found that the methods, tools, and resources they used in their privacy work generally did not help address the unique pri- vacy risks introduced or exacerbated by their use of AI in their products. Collectively, these findings reveal the need and opportunity to create tools, resources, and support structures to improve practitioners’ awareness of AI-specific privacy risks, motivations to do AI privacy work, and ability to ad- dress privacy harms introduced or exacerbated by their use of AI in consumer products. 

Modern advances in AI have increased employer interest in tracking workers’ biometric signals — e.g., their brainwaves and facial expressions — to evaluate and make predictions about their performance and productivity. These technologies afford managers information about internal emotional and physiological states that were previously accessible only to individual workers, raising new concerns around worker privacy and autonomy. Yet, the research literature on the impact of AI-powered biometric work monitoring (AI-BWM) technologies on workers remains fragmented across disciplines and industry sectors, limiting our understanding of its impacts on workers at large. In this paper, we sytematically review 129 papers, spanning varied disciplines and industry sectors, that discuss and analyze the impact of AI-powered biometric monitoring technologies in occupational settings. We situate this literature across a process model that spans the development, deployment, and usage phases of these technologies. We further draw on Shelby et al.’s Taxonomy of Socio-technical Harms in AI systems to systematize the harms experienced by workers across the three phases of our process model. We find that the development, deployment, and sustained use of AI-powered biometric work monitoring technologies put workers at risk of a number of the socio-technical harms specified by Shelby et al.: e.g., by forcing workers to exert additional emotional labor to avoid flagging unreliable affect monitoring systems, or through the use of these data to make inferences about productivity. Our research contributes to the field of critical AI studies by highlighting the potential for a cascade of harms to occur when the impact of these technologies on workers is not considered at all phases of our process model. 

Privacy is a key principle for developing ethical AI technologies, but how does including AI technologies in products and services change privacy risks? We constructed a taxonomy of AI privacy risks by an- alyzing 321 documented AI privacy incidents. We codifed how the unique capabilities and requirements of AI technologies described in those incidents generated new privacy risks, exacerbated known ones, or otherwise did not meaningfully alter the risk. We present 12 high-level privacy risks that AI technologies either newly created (e.g., exposure risks from deepfake pornography) or exacerbated (e.g., surveillance risks from collecting training data). One upshot of our work is that incorporating AI technologies into a product can alter the privacy risks it entails. Yet, current approaches to privacy-preserving AI/ML (e.g., federated learning, diferential pri- vacy, checklists) only address a subset of the privacy risks arising from the capabilities and data requirements of AI.