Searcharxiv⌕ Search

arXiv subjects

Wesley Hanwen Deng

Publications and source records attributed to Wesley Hanwen Deng.

At least 19 recordsLinked to original sources

A Crowd-Audited Atlas of Implicit Bias in LLM-Powered Agents

Prior work has shown that large language models (LLMs) express implicit biases in high-publicity, one-off decision scenarios such as hiring and intelligence assessment. However, as people increasingly employ autonomous, LLM-powered AI agents that make decisions on their behalf in doing day-to-day tasks, it is of critical importance that we understand how implicit biases, based on sociodemographic assumptions, emerge across a broad milieu of workaday decision scenarios. We used the 17 domains of daily activity in the American Time Use Survey (ATUS) to generate an atlas of 506 decision scenarios in which LLMs --- endowed with personas representing different races, gender identities, and political ideologies --- expressed significant implicit bias. Next, in a crowdsourced audit, 500 participants made 8,010 binary choices about which implicitly biased decision scenarios most deserved further review. We found that participants prioritized decision scenarios that had broad relevance (e.g., about food choices), were high stakes (e.g., about health), and where all members of a simulated sociodemographic group made unanimous decisions (because it was a clear indication of stereotyping). They were less swayed by the magnitude of an observed disparity or the specific ATUS domain the decision scenario represented. When examining the rationale behind their prioritization choices, we found that participants often interpreted disparities as traits of demographic groups and judged whether those traits seemed plausible, often based on their own internalized stereotypes. We discuss implications for building personalized AI agents, digital twins, social simulations, and for using crowdsourced bias audits.

cs.CY↗

Right Advice, Wrong Reasons: How Decision and Reasoning Alignment Shape Reliance on AI Advisors

When people seek advice from AI on social situations, the system may recommend a decision they fully agree with but justify it through reasoning they disagree with. Conversely, it may offer reasoning they agree with, and yet reach a decision they reject. We distinguish decision alignment, whether an advisor recommends the option favored by a user's measured values, from reason alignment, whether its stated reasons align with the user's values. In this work, we ask how these two forms of alignment shape people's preferences for AI advisors and their willingness to use them in different roles. We conducted a within-subjects study (N=121) in which participants evaluated three AI advisors varying in decision and reason alignment across eight everyday moral dilemmas. Participants then chose which advisors they would consult, allow to draft messages for review, or allow to act on their behalf in new tasks. When seeking advice, participants most often preferred an advisor that reasoned through their prioritized values while recommending the opposing decision (40%). However, when participants had to pick advisors for new tasks, their preferences shifted toward advisors that also recommended decisions aligned with their values, rather than those that only shared their reasons. Also, as participants considered tasks that gave AI advisors greater autonomy, they became more likely to opt out of using any advisor, from 12% when consulting to 38% when allowing an advisor to act on their behalf. Our findings suggest that decision and reason alignment shape advisor preferences differently across contexts of AI use, and that alignment with a user's values does not necessarily imply willingness to delegate.

cs.HC↗

MM-SCALE: Evaluating Evidence-Grounded Moral Judgment in Vision-Language Models

Vision-Language Models increasingly make socially consequential judgments from image-text inputs, yet current evaluations often stop at verdict-level accuracy: whether the final label is human-aligned, not whether it follows from the correct evidence. We introduce MM-SCALE (Multimodal Moral SCALE), a benchmark for evidence-grounded moral judgment built around a design choice: each image is paired with multiple action scenarios, so models must compare how different actions interact with the same visual context rather than score images in isolation. MM-SCALE contains 8,444 image contexts and 21,977 action scenarios, each annotated with a 5-point moral acceptability rating and a modality-grounding label indicating whether the human judgment relied on text, image, or both. Three tasks evaluate whether models (i) assign calibrated scalar moral scores, (ii) preserve human orderings within a shared visual context, and (iii) ground rationales in the evidence source humans found critical to judgment. Across models, aggregate metrics overstate alignment: NDCG@5 remains high across models, while within-image pairwise accuracy remains only 0.53-0.62 even on scenario pairs whose human mean ratings differ by at least one point. CoT has inconsistent effects on calibration and does not reliably improve within-image ordering. Evidence grounding evaluations further show that models often describe visual context without making it critical for the verdict. These results reveal an evidence-grounding failure in current VLMs and position MM-SCALE as an evaluation testbed for moral judgment under shared image contexts. This dataset includes potentially harmful and sensitive visual content. Images are intended solely for evaluation.

cs.CV↗

Critical or Compliant? The Double-Edged Sword of Reasoning in Chain-of-Thought Explanations

Explanations are often promoted as tools for transparency, but they can also foster confirmation bias; users may assume reasoning is correct whenever outputs appear acceptable. We study this double-edged role of Chain-of-Thought (CoT) explanations in multimodal moral scenarios by systematically perturbing reasoning chains and manipulating delivery tones. Specifically, we analyze reasoning errors in vision language models (VLMs) and how they impact user trust and the ability to detect errors. Our findings reveal two key effects: (1) users often equate trust with outcome agreement, sustaining reliance even when reasoning is flawed, and (2) the confident tone suppresses error detection while maintaining reliance, showing that delivery styles can override correctness. These results highlight how CoT explanations can simultaneously clarify and mislead, underscoring the need for NLP systems to provide explanations that encourage scrutiny and critical thinking rather than blind trust. All code will be released publicly.

cs.CL↗

Who Does What in AI Auditing? Designing Human-AI Collaboration for Auditing Generative AI

AI auditing increasingly incorporates AI agents to expand the scale and breadth of audit coverage, yet little is known about how auditing work should be divided without displacing human judgment. We introduce Human-Agent Audit Collaboration (HAAC), a workflow and system for structuring human-AI collaboration in AI auditing. Drawing on prior work and formative consultations with AI auditing practitioners, HAAC specifies how agents can support exploration, assessment, reporting, and review while preserving human oversight where contextual judgment is critical. We instantiate HAAC for conversational shopping agents and evaluate it through two studies. With 71 auditors, AI assistance increased attack success and broadened exploration, while also shaping later attacks and increasing auditors' reliance on AI-generated assessments and reports. Interviews with Responsible AI practitioners showed that actionable audits require visibility into coverage, reproducible attack trajectories, and evaluation of the auditing agents themselves. Our findings identify design considerations for effective and accountable human-AI auditing.

cs.HC↗

Locating Translation as a Craft in the Age of AI

Rapid development of Large Language Models (LLMs) and similar automated approaches for translation tasks is increasingly affecting the landscape of translation technologies. As concerns about the outsourcing of translator work to these automated translation tools grow, it is increasingly crucial to gather insights from the translation community directly. To this end, we conduct an interview study with 19 professional translators working across 11 languages and 11 domains to understand their perspectives, experiences, and concerns with using translation technologies in their work. We find that translators are cautious when incorporating new tools into their workflow, with several expressing concerns that machine translation (MT) and LLMs are infringing on the necessary human aspects and verification processes of translation. Importantly, translators are worried that these tools have potential for harmful downstream effects due to compromising the human aspects of translation work. These findings demonstrate the need to develop translation technologies that directly serve translators' needs rather than replacing human translation. This can be done by focusing more on the assistive tools that emphasize the uncertain, social, and ultimately human character of translation, rather than automation.

cs.CY↗

Use and Effects of LLMs in Peer Review: A Randomized Experiment and Survey at ICML 2026

LLMs are rapidly reshaping peer review, making it important to understand how reviewers use them in practice and how different LLM-use policies affect review outcomes. We investigate these questions through a randomized experiment and an anonymous post-survey at ICML 2026, a major machine learning conference involving over 24,000 papers and 17,000 reviewers. Reviewers were assigned to either a conservative policy prohibiting all LLM use or a permissive policy allowing limited assistance, with randomization among a subset of main-track papers and reviewers. Policy assignment had near-zero effects on final paper decisions, paper scores, and reviewer confidence, although reviews under the permissive policy were 5.5-7% longer. Post-survey responses (N=1,486) revealed diverse attitudes toward LLMs and substantial noncompliance: 22.5% of conservative-policy reviewers reported using an LLM despite the prohibition, and 36.5% of permissive-policy reviewers reported at least one explicitly disallowed use. We discuss implications for future peer-review policy and tool design.

cs.HC↗

PersonaTeaming: Supporting Persona-Driven Red-Teaming for Generative AI

Recent developments in AI safety research have called for red-teaming methods that effectively surface potential risks posed by generative AI models, with growing emphasis on how red-teamers' backgrounds and perspectives shape their strategies and the risks they uncover. While automated red-teaming approaches promise to complement human red-teaming through larger-scale exploration, existing automated approaches do not account for human identities and rarely incorporate human inputs. In this work, we explore persona-driven red-teaming to advance both automated red-teaming and human-AI collaboration. We first develop PersonaTeaming Workflow, which incorporates personas into the adversarial prompt generation process to explore a wider spectrum of adversarial strategies. Compared to RainbowPlus, a state-of-the-art automated red-teaming method, PersonaTeaming Workflow achieves higher attack success rates while maintaining prompt diversity. However, since automated personas only approximate real human perspectives, we further instantiate PersonaTeaming Workflow as PersonaTeaming Playground, a user-facing interface that enables red-teamers to author their own personas and collaborate with AI to mutate and refine prompts. In a user study with 11 industry practitioners, we found that PersonaTeaming Playground enabled diverse red-teaming strategies and outputs that practitioners perceived as useful, and that AI-generated suggestions in the PersonaTeaming Playground encouraged out-of-the-box thinking even when practitioners did not follow them strictly. Together, our work advances both automated and human-in-the-loop approaches to red-teaming, while shedding light on interaction patterns and design insights for supporting human-AI collaboration in generative AI red-teaming.

cs.HC↗

Cognitive Chain-of-Thought (CoCoT): Structured Multimodal Reasoning about Social Situations

Chain-of-Thought (CoT) prompting helps models think step by step. But naive CoT breaks down in visually grounded social tasks, where models must perceive, understand, and judge all at once; bridging perception with norm-grounded reasoning. Recent work has introduced structured reasoning for multi-turn agent planning and visual QA, decomposing tasks into sequential sub-goals. To extend this to single-shot multimodal social reasoning, we introduce Cognitive Chain-of-Thought (CoCoT), a reasoning framework that structures vision-language-model (VLM) reasoning through three cognitively inspired stages: Perception (extract grounded facts), Situation (infer situations), and Norm (applying social norms). Evaluation across multiple distinct tasks such as multimodal intent disambiguation, multimodal theory of mind, social commonsense reasoning, and safety instruction following, shows consistent improvements (5.9% to 4.6% on average). We further explore the utility of CoCoT for improving models' reasoning through training and show that supervised fine-tuning on CoCoT-structured traces yields 5-6% improvements without explicit CoCoT prompting at inference, demonstrating that models internalize the structured reasoning pattern rather than merely following instructions. We show that structuring model reasoning through cognitively grounded stages enhances interpretability and social alignment, laying the groundwork for more reliable multimodal systems.

cs.CL↗

"Death by a thousand taxonomies?": AI Risk Classification In Practice

The harms in which AI is implicated range in nature and scope from unsafe user interactions through to the societal-wide consequences of AI adoption. Classification of the diverse risks of AI is foundational to AI governance: regulators, technology firms, and policymakers need structured accounts of risk upon which to act. Researchers and practitioners have accordingly developed many Sociotechnical Outcome Taxonomies (SOT). This paper presents an empirical study of SOT development and use, drawing on 25 interviews with researchers and practitioners across industry, academia, civil society, and government. We find SOT are weakly integrated into AI governance processes, and identify two features of SOT design and use that explain why. First, the design choices through which SOT produce structured representations of the complex problem space of AI risks tend to be invisible to downstream taxonomy users. Those users treat the resulting categories as exhaustive accounts of risk rather than as interpretive aids. Second, SOT typically enumerate harms without linking them to decision points or actors implicated in their occurrence, leaving accountability difficult to assign. We close with design recommendations for SOT developers and users, and argue realising the potential of SOT requires governance infrastructure that does not yet exist.

cs.CY↗

What We Know about Responsible AI Practices in Industry: A Half Decade of Empirical Research

Responsible AI (RAI) has become a central concern for technology companies, regulators, and the public. How industry practitioners interpret, implement, and sustain RAI work directly shapes the design and deployment of AI systems. As empirical scholarship examining RAI practices in industry has rapidly expanded, findings are dispersed across studies that focus on different roles, organizational contexts, and interventions. This work synthesizes current knowledge through a literature review of 161 empirical studies spanning six years, each engaging industry practitioners via interviews, surveys, workshops, ethnographies, and other methods. Our synthesis reveals both meaningful progress and persistent challenges in industry RAI practice. Practitioner awareness has increased, RAI activities have become more professionalized, and interventions such as toolkits and guidelines are more widely adopted. At the same time, practitioners continue to face substantial barriers, including limited training, uneven organizational support, and a lack of interventions tailored to day-to-day work practices. By consolidating and organizing these findings, we provide a more complete account of industry RAI than any single study to date. We conclude by discussing implications for RAI researchers, practitioners seeking to adopt effective practices, and policymakers aiming to ground governance efforts in the realities of industry contexts.

cs.HC↗

Vipera: Blending Visual and LLM-Driven Guidance for Systematic Auditing of Text-to-Image Generative AI

Despite their increasing capabilities, text-to-image generative AI systems are known to produce biased, offensive, and otherwise problematic outputs. While recent advancements have supported testing and auditing of generative AI, existing auditing methods still face challenges in supporting effectively explore the vast space of AI-generated outputs in a structured way. To address this gap, we conducted formative studies with five AI auditors and synthesized five design goals for supporting systematic AI audits. Based on these insights, we developed Vipera, an interactive auditing interface that employs multiple visual cues including a scene graph to facilitate image sensemaking and inspire auditors to explore and hierarchically organize the auditing criteria. Additionally, Vipera leverages LLM-powered suggestions to facilitate exploration of unexplored auditing directions. Through a controlled experiment with 24 participants experienced in AI auditing, we demonstrate Vipera's effectiveness in helping auditors navigate large AI output spaces and organize their analyses while engaging with diverse criteria.

cs.HC↗

Seeing Twice: How Side-by-Side T2I Comparison Changes Auditing Strategies

While generative AI systems have gained popularity in diverse applications, their potential to produce harmful outputs limits their trustworthiness and utility. A small but growing line of research has explored tools and processes to better engage non-AI expert users in auditing generative AI systems. In this work, we present the design and evaluation of MIRAGE, a web-based tool exploring a "contrast-first" workflow that allows users to pick up to four different text-to-image (T2I) models, view their images side-by-side, and provide feedback on model performance on a single screen. In our user study with fifteen participants, we used four predefined models for consistency, with only a single model initially being shown. We found that most participants shifted from analyzing individual images to general model output patterns once the side-by-side step appeared with all four models; several participants coined persistent "model personalities" (e.g., cartoonish, saturated) that helped them form expectations about how each model would behave on future prompts. Bilingual participants also surfaced a language-fidelity gap, as English prompts produced more accurate images than Portuguese or Chinese, an issue often overlooked when dealing with a single model. These findings suggest that simple comparative interfaces can accelerate bias discovery and reshape how people think about generative models.

cs.HC↗

PersonaTeaming: Exploring How Introducing Personas Can Improve Automated AI Red-Teaming

Recent developments in AI governance and safety research have called for red-teaming methods that can effectively surface potential risks posed by AI models. Many of these calls have emphasized how the identities and backgrounds of red-teamers can shape their red-teaming strategies, and thus the kinds of risks they are likely to uncover. While automated red-teaming approaches promise to complement human red-teaming by enabling larger-scale exploration of model behavior, current approaches do not consider the role of identity. As an initial step towards incorporating people's background and identities in automated red-teaming, we develop and evaluate a novel method, PersonaTeaming, that introduces personas in the adversarial prompt generation process to explore a wider spectrum of adversarial strategies. In particular, we first introduce a methodology for mutating prompts based on either "red-teaming expert" personas or "regular AI user" personas. We then develop a dynamic persona-generating algorithm that automatically generates various persona types adaptive to different seed prompts. In addition, we develop a set of new metrics to explicitly measure the "mutation distance" to complement existing diversity measurements of adversarial prompts. Our experiments show promising improvements (up to 144.1%) in the attack success rates of adversarial prompts through persona mutation, while maintaining prompt diversity, compared to RainbowPlus, a state-of-the-art automated red-teaming method. We discuss the strengths and limitations of different persona types and mutation methods, shedding light on future opportunities to explore complementarities between automated and human red-teaming approaches.

cs.AI↗

"I Don't Think RAI Applies to My Model'' -- Engaging Non-champions with Sticky Stories for Responsible AI Work

Responsible AI (RAI) tools -- checklists, templates, and governance processes -- often engage RAI champions, individuals intrinsically motivated to advocate ethical practices, but fail to reach non-champions, who frequently dismiss them as bureaucratic tasks. To explore this gap, we shadowed meetings and interviewed data scientists at an organization, finding that practitioners perceived RAI as irrelevant to their work. Building on these insights and theoretical foundations, we derived design principles for engaging non-champions, and introduced sticky stories -- narratives of unexpected ML harms designed to be concrete, severe, surprising, diverse, and relevant, unlike widely circulated media to which practitioners are desensitized. Using a compound AI system, we generated and evaluated sticky stories through human and LLM assessments at scale, confirming they embodied the intended qualities. In a study with 29 practitioners, we found that, compared to regular stories, sticky stories significantly increased time spent on harm identification, broadened the range of harms recognized, and fostered deeper reflection.

cs.HC↗

Why (not) use AI? Analyzing People's Reasoning and Conditions for AI Acceptability

In recent years, there has been a growing recognition of the need to incorporate lay-people's input into the governance and acceptability assessment of AI usage. However, how and why people judge acceptability of different AI use cases remains under-explored, despite it being crucial towards understanding and addressing potential sources of disagreement. In this work, we investigate the demographic and reasoning factors that influence people's judgments about AI's development via a survey administered to demographically diverse participants (N=197). As a way to probe into these decision factors as well as inherent variations of perceptions across use cases, we consider ten distinct labor-replacement (e.g., Lawyer AI) and personal health (e.g., Digital Medical Advice AI) AI use cases. We explore the relationships between participants' judgments and their rationales such as reasoning approaches (cost-benefit reasoning vs. rule-based). Our empirical findings reveal a number of factors that influence acceptance. We find lower acceptance of labor-replacement usage over personal health, significant influence of demographics factors such as gender, employment, education, and AI literacy level, and prevalence of rule-based reasoning for unacceptable use cases. Moreover, we observe unified reasoning type (e.g., cost-benefit reasoning) leading to higher agreement. Based on these findings, we discuss the key implications towards understanding and mitigating disagreements on the acceptability of AI use cases to collaboratively build consensus.

cs.CY↗

WeAudit: Scaffolding User Auditors and AI Practitioners in Auditing Generative AI

There has been growing interest from both practitioners and researchers in engaging end users in AI auditing, to draw upon users' unique knowledge and lived experiences. However, we know little about how to effectively scaffold end users in auditing in ways that can generate actionable insights for AI practitioners. Through formative studies with both users and AI practitioners, we first identified a set of design goals to support user-engaged AI auditing. We then developed WeAudit, a workflow and system that supports end users in auditing AI both individually and collectively. We evaluated WeAudit through a three-week user study with user auditors and interviews with industry Generative AI practitioners. Our findings offer insights into how WeAudit supports users in noticing and reflecting upon potential AI harms and in articulating their findings in ways that industry practitioners can act upon. Based on our observations and feedback from both users and practitioners, we identify several opportunities to better support user engagement in AI auditing processes. We discuss implications for future research to support effective and responsible user engagement in AI auditing and red-teaming.

cs.HC↗

MIRAGE: Multi-model Interface for Reviewing and Auditing Generative Text-to-Image AI

While generative AI systems have gained popularity in diverse applications, their potential to produce harmful outputs limits their trustworthiness and usability in different applications. Recent years have seen growing interest in engaging diverse AI users in auditing generative AI that might impact their lives. To this end, we propose MIRAGE as a web-based tool where AI users can compare outputs from multiple AI text-to-image (T2I) models by auditing AI-generated images, and report their findings in a structured way. We used MIRAGE to conduct a preliminary user study with five participants and found that MIRAGE users could leverage their own lived experiences and identities to surface previously unnoticed details around harmful biases when reviewing multiple T2I models' outputs compared to reviewing only one.

cs.HC↗