The design of generative AI and computer vision systems is often guided by technical in-lab evaluations that can differ substantially from real-world uses. This misalignment, at best, can lead to inefficiencies and, at worst, cause unintended harms in unforeseen contexts. Humans of Generative AI (HuG) recenters attention on the people who use and are affected by these systems. We invite talks and posters from human-centric research that inform the design or evaluation of generative AI and computer vision systems. Through this workshop, we will encourage and develop cross-disciplinary collaboration between computer vision and human-centric researchers, two often-disconnected communities.

Invited Speakers

Keynotes

Juliana Castro Varón

Juliana Castro Varón

Senior Design Editor of AI Initiatives at The New York Times

Juliana Castro Varón is the Senior Design Editor at the Artificial Intelligence Initiatives team at The New York Times and the founder of the bilingual open-access publishing house Cita Press.

Juliana has received fellowships and support from Fulbright, the Mellon Foundation and Harvard, where she is an affiliate at the Berkman Klein Center for Internet & Society. Juliana also contributes humor and cartoons to The New Yorker magazine.

Rishi Bommasani

Rishi Bommasani

Senior Research Scholar at Stanford HAI

Rishi Bommasani is a senior research scholar at Stanford's Institute for Human-Centered Artificial Intelligence. He researches the societal and economic impact of AI. His research has received several recognitions at ML conferences and has been covered by the New York Times, Nature, Science, the Washington Post, and Wall Street Journal.

His research shapes public policy: he is the lead author of the California Report on Frontier AI Policy that led to the first US laws on frontier AI, an independent expert chair of the EU AI Act Code of Practice that clarifies the first comprehensive worldwide laws on frontier AI, and an author of the International Scientific Report on the Safety of Advanced AI.

Panelists

The panel will feature keynote speakers Rishi Bommasani and Juliana Castro Varón alongside our set of invited panelists.

Tiana Oreglia

Tiana Oreglia

Concept Artist and Artist Advocate

Tiana Oreglia is a freelance concept artist, comic artist, and illustrator based in and from San Francisco.

Despite having a degree in animation from Sheridan College, she's worked predominantly in the games industry for the past six years, while also working in animation and editorial illustration. She's worked on titles such as Voodoo Detective as the lead character designer and with Valve on their latest Team Fortress 2 comic as a penciler, inker, and colorist. For the past couple years, she's been advocating for artists to legislators in California as new technologies threaten to displace workers.

Elissa M. Redmiles

Elissa M. Redmiles

Assistant Professor at Georgetown University

Dr. Redmiles uses computational, economic, and social science methods to understand users' security, privacy, and online safety-related decision-making processes. Her work specifically investigates inequalities that arise in these processes in order to ultimately design systems that facilitate safety equitably across users. Dr. Redmiles' current projects focus on security, privacy and safety in digital labor, digital intimacy, digitally-mediated offline interactions, and medical data donation; building transparency tools for privacy enhancing technologies such as differential privacy; and measuring biases in and ethics of AI-based technologies. Her research has received multiple paper recognitions at USENIX Security, ACM CCS, ACM CHI, ACM CSCW, and ACM EAAMO and has been featured in popular press publications such as the New York Times, Wall Street Journal, Scientific American, Rolling Stone, Wired, and Forbes.

More about Elissa

Dr. Elissa M. Redmiles is an Assistant Professor at Georgetown University in the Computer Science Department and a Faculty Associate at the Berkman Klein Center for Internet & Society at Harvard University. She was previously a faculty member at the Max Planck Institute for Software Systems and has additionally served as a consultant and researcher at multiple institutions, including Microsoft Research, Facebook, the World Bank, the Center for Democracy and Technology, and the Partnership on AI. Dr. Redmiles received her B.S. (Cum Laude), M.S. and Ph.D. in Computer Science -- with a concentration in Survey Methodology -- from the University of Maryland.

Moderator

David A. Forsyth

David A. Forsyth

Fulton-Watson-Copp Chair in Computer Science, University of Illinois at Urbana-Champaign

David A. Forsyth is the Fulton-Watson-Copp Chair in Computer Science at the University of Illinois at Urbana-Champaign. He moved there from UC Berkeley, where he was also a full professor. He has occupied the Fulton-Watson-Copp Chair in Computer Science at the University of Illinois since 2014 and has published over 170 papers on computer vision, computer graphics, and machine learning.

More about David

He has served as program co-chair for IEEE Computer Vision and Pattern Recognition in 2000, 2011, 2018, and 2021; general co-chair for CVPR 2006 and 2015 and ICCV 2019; and program co-chair for the European Conference on Computer Vision 2008. He is a regular member of the program committee of all major international conferences on computer vision and now regularly serves as Senior Advisor to the Program Chairs at major computer vision conferences. He has served six years on the SIGGRAPH program committee and is a regular reviewer for that conference.

He has received best paper awards at the International Conference on Computer Vision and at the European Conference on Computer Vision. He received an IEEE Technical Achievement Award in 2005, the Mark Everingham Prize in 2024, and the PAMI-TC Significant Researcher Award in 2025. He became an IEEE Fellow in 2009 and an ACM Fellow in 2014. His textbook, Computer Vision: A Modern Approach, joint with J. Ponce and published by Prentice Hall, was widely adopted as a course text. Further textbooks, Probability and Statistics for Computer Science and Applied Machine Learning, are currently in print. He has served two terms as Editor in Chief of IEEE TPAMI, serves on a number of scientific advisory boards, and has an active practice as an expert witness.

Schedule

The workshop will be held on June 4, 2026 from 8:30a-1:00p in Room 710. Times are listed in Denver local time (MDT).

8:30a
Welcome
Topic 1: Human-Centric Findings That Inform the Technical Design of AI Systems
8:45a
Lightning Talks 1
Talk details
  1. Training-Time Attribution and Use-Time Valuation in Generative AI Creative Work Lu Xian
  2. Nonverbal Cues in the Loop: Towards Discomfort Detection and Consent-Aware Design in Video Dating Ratna Kandala, Niva Manchanda, Akshata Kishore Moharir
  3. A Continuum from Naturalistic to Synthetic Traffic Data David Kuehn
  4. The Past and Future of Nude Images in Research Princessa Cintaqia, Arshia Arya, Deepak Kumar, Elissa M. Redmiles, Allison McDonald, Lucy Qin
  5. From Policy to Design: Uncovering Socioemotional Risks in Generative AI Systems for Education Emmanuel Adeloju, Lindsey McCaleb, Rebekah Jongewaard, Nicole Oster, Punya Mishra
  6. Caught in a Mafia Romance: How Users Explore Intimate Narratives with Chatbots Julia Kieserman, Cat Mai, Sara Lignell, Lucy Qin, Athanasios Andreou, Damon McCoy, Rosanna Bellini
9:30a
Keynote: Juliana Castro Varón
9:55a
Break
Panel Discussion: Humans of Generative AI
10:05a
Panel Discussion
10:50a
Break
Topic 2: Technical Designs That Consider Unattended Needs of Users
11:00a
Lightning Talks 2
Talk details
  1. Overlooked but Ubiquitous: Benchmarking Disability Bias in T2I Models Sophia Lichtenberg, Judith Masthoff, Albert Gatt
  2. Structured Listening: Codifying Human-Meaningful Voice Signals to Ground Generative AI Reasoning Rachel Manzelli, Carter Huffman
  3. The Evolutionary Editor: StackGP Red-Teaming for Visual Bias in Newsroom AI Ruchika Gupta, Aarav Gupta
  4. When 95% Accuracy Isn't Good Enough: Bridging the Gap Between Forensic Biometric Facial Recognition Benchmarks and Real-World Human Impact Justin Norman
  5. ``Please Reach Out'': Evaluating LLM Responses to Depression and Anxiety Symptoms with the MH-Disclosure Dataset Micah Benson, Samishti Bhatia, Zhengyang Shan, Afitab Iyigun, Taijah Chavis, Anthony J. Rosellini, Aaron Mueller
  6. Tamper resistant Machine Unlearning Algorithms for Open-weight AI Models Amro Abdalla, Ismail Shaheen, Dan DeGenaro, Rupayan Mallick, Bogdan Raita, Sarah Adel Bargal
11:45a
Keynote: Rishi Bommasani
Poster Session
12:10p
Poster Session
12:55p-1:00p
Closing
1:30p
Unofficial group lunch!

Accepted Works

Human-Centric Findings Informing AI System Design

  • Training-Time Attribution and Use-Time Valuation in Generative AI Creative Work Lu Xian
    Abstract

    Existing CV approaches to consent, credit, and compensation largely operationalize creative contribution as training-time source attribution: who contributed data, whose style influenced generation, and whose work should be protected from extraction. Drawing on empirical findings from AI-assisted design practice, we show that in professional creative work, AI-generated outputs often remain provisional until designers make them usable through aesthetic judgment, selective adoption, and collaborative review. This reveals a mismatch between attribution-centered technical accounts of contribution and the downstream labor through which outputs acquire use value in practice. We therefore propose complementing training-time attribution with use-time valuation, an account of the human work that turns generated candidates into client deliverables, brand artifacts, or accessible interfaces. This further suggests that metrics used to recognize contribution, and, by extension, to support compensation, should also consider workflow-centric ones that capture whether and how generated outputs acquire value in real production contexts.

  • Nonverbal Cues in the Loop: Towards Discomfort Detection and Consent-Aware Design in Video Dating Ratna Kandala, Niva Manchanda, Akshata Kishore Moharir
    Abstract

    Online dating has become the dominant pathway to romantic relationships, yet current platforms strip the nonverbal cues: gaze, facial expression, body posture, response timing, that humans rely on to signal comfort, disinterest, and consent. This creates a communication gap with disproportionate safety consequences for women. We propose a fairness-first Computer Vision (CV) research agenda organized around real-time discomfort detection, engagement asymmetry modeling, consent-aware interaction design, and longitudinal interaction summarization. Responsible deployment demands purpose-built datasets with disaggregated fairness evaluation and architectural commitments to on-device processing.

  • A Continuum from Naturalistic to Synthetic Traffic Data David Kuehn
    Abstract

    Both naturalistic and synthetic data are important for advancing highway transportation safety and reliability. Many researchers create synthetic data to supplement the distribution of naturalistic data or to enhance data privacy. Synthetic data are created from naturalistic examples. Naturalistic data all have some level of processing. Rather than consider naturalistic and synthetic data are discrete categories, researchers can benefit from considering naturalistic and synthetic data as a continuum with tradeoffs across the continuum on cost, usability and fit-for-purpose.

  • The Past and Future of Nude Images in Research Princessa Cintaqia, Arshia Arya, Deepak Kumar, Elissa M. Redmiles, Allison McDonald, Lucy Qin
    Abstract

    In order to train, test, and evaluate nudity detection models, machine learning researchers typically rely on nude images scraped from the Internet. While this practice is common for assembling datasets for general image-recognition tasks, the scraping of nude images raises unique ethical concerns. Our research finds that this content is collected and, in some cases, subsequently distributed by researchers without consent, leading to potential misuse and exacerbating harm against the subjects depicted. We argue that the distribution of nonconsensually collected nude images by researchers perpetuates image-based sexual abuse and that the machine learning community should stop the nonconsensual use of nude images in research. To characterize the scope and nature of this problem, we conducted a systematic review of papers published in computing venues that collect and use nude images. Our results paint a grim reality: norms around the usage of nude images are sparse, leading to a litany of problematic practices like distributing and publishing nude images with uncensored faces, and intentionally collecting and sharing abusive content. We also propose a future participatorily-governed dataset that could enable research with nude images while respecting the dignity of all people.

  • From Policy to Design: Uncovering Socioemotional Risks in Generative AI Systems for Education Emmanuel Adeloju, Lindsey McCaleb, Rebekah Jongewaard, Nicole Oster, Punya Mishra
    Abstract

    Much of the discourse on Generative AI (GenAI) safety centers on concerns such as bias in outputs, data privacy, misinformation, and hallucinations. These are real problems that are getting the attention of the technical community. What has yet to receive adequate attention is a parallel category of socioemotional harms that are structurally rooted in deliberate system design decisions, particularly those that affect K-12 students. These are harms identified by K-12 stakeholders and policymakers and are recognized as major concerns that schools and institutions need to proactively and reactively guard against. We therefore present the findings from a systematic analysis of 48 institutional AI guidance and policy documents from K-12 educational context spanning 33 US states and territories, 9 independent organizations, and 6 other countries. We evaluated the documents through a five-category framework, including (1) Anthropomorphization and Mental Model distortion, (2) Developmental Vulnerability, (3) Emotional Attachment and Parasocial Relationships, (4) Engineered Manipulation and Supernormal Stimuli, and (5) Social Isolation and Relationship Displacement. Our results revealed two key concerns and design problems. First, Emotional Attachment and Parasocial Relationships was the least addressed category across the entire corpus, despite well-documented cases of adolescent harms from AI companions. In contrast, Engineered Manipulation and Supernormal Stimuli and Developmental Vulnerability were the most frequently addressed categories across the corpus, reflecting growing awareness and concern about the deliberate design of AI systems to ensure students' stickiness. The polarity from these findings reveals struggling attempts of policymakers at grappling with the effects of GenAI and presents an opportunity for AI developers to align technical designs with consideration of educational standards and requirements. We propose a design-centric socioemotional (re)evaluation of “engagement-first” ethos in GenAI development and the alignment of system architecture with the nuanced needs of K-12 students, who are particularly vulnerable.

  • Caught in a Mafia Romance: How Users Explore Intimate Narratives with Chatbots Julia Kieserman, Cat Mai, Sara Lignell, Lucy Qin, Athanasios Andreou, Damon McCoy, Rosanna Bellini
    Abstract

    AI chatbots, built using large language models, are increasingly integrated into society and mimic the patterns of human text exchanges. While previous research has raised concerns that humans may form romantic attachment to chatbots, the range of AI-mediated interactions that people wish to create for themselves or others with chatbots remains poorly understood, particularly given the fast evolving landscape of chatbots. We provide an empirical study of Character.AI (cAI), a popular chatbot platform that enables users to design and share character-based bots, and synthesize this with an analysis of Reddit posts from cAI users. Contrary to popular narratives, we identify that users want to: (1) engage in intimate roleplay with young adult, masculine-presenting characters that place users in a position of inferior power in well-defined scenarios and (2) immerse themselves in boundless, fantasy settings. We further find that users problematize both the excessive and insufficient sexualized content in such interactions which warrants novel digital-safety features.

Technical Designs for Unattended User Needs

  • Overlooked but Ubiquitous: Benchmarking Disability Bias in T2I Models Sophia Lichtenberg, Judith Masthoff, Albert Gatt
    Abstract

    Text-to-image (T2I) models exhibit social biases and prior work has focused on gender, skin tone, and cultural representation within occupational context. However, disability remains systematically underexplored. Current evaluation practices often fail to align with sociologically grounded definitions of stereotyping, limiting assessment of representational harms toward people with disabilities (PWD).
    We created INCLUsive Disability Evaluation (INCLUDE-BENCH), the first large-scale benchmark for evaluating disability-related bias in T2I models. INCLUDE-BENCH comprises 119K generated images based on context-rich and intersectional prompt design. We evaluated 15 open-source and 2 closed models and our findings show that: (1) mobility-impaired and default disability prompts predominantly yield wheelchair depictions across all models; (2) disability-conditioned generations consistently exhibit less diversity and (3) stereotypical portrayals demonstrate stronger disability–text alignment and (4) introduce Stereotype Content Model (SCM) Score, demonstrating that T2I models reflect real-world stereotypical associations.

  • Structured Listening: Codifying Human-Meaningful Voice Signals to Ground Generative AI Reasoning Rachel Manzelli, Carter Huffman
    Abstract

    Generative AI systems are increasingly tasked with understanding voice conversations, yet the representations they consume systematically discard the dimensions of speech that humans most rely on: emotion, intent, behavioral dynamics, and participant roles. Transcription-first pipelines lose paralinguistic signals at the moment of conversion. Reasoning directly over raw audio is computationally intractable at scale, and the compression required to make it feasible discards the same signals. When a language model reasons over what remains, it is reasoning over a shadow of the conversation. This paper presents the Ensemble Listening Model (ELM) as a response to this structural gap. Rather than asking a generative model to interpret raw conversational audio, the ELM first produces a rich, structured representation explicitly designed around what humans care about when they listen: clip-level signals such as emotion, and tonality; detections such as manipulation tactics or interruption patterns; participant-level inferences about roles and intents; and conversation-level summaries. These structured outputs then serve as context for a downstream generative model, grounding its reasoning in human-meaningful terms. The ELM achieves this through a hierarchical ensemble of small, specialized models that operate orders of magnitude more efficiently than monolithic foundation models while matching or exceeding their accuracy. We argue that this points toward a broader design principle: the human-meaningful structure latent in conversational data should be made explicit before it reaches a generative model, not inferred from tokens after the fact.

  • The Evolutionary Editor: StackGP Red-Teaming for Visual Bias in Newsroom AI Ruchika Gupta, Aarav Gupta
    Abstract

    As generative artificial intelligence (GAI) is integrated into journalistic workflows for visual content creation, news organizations face a critical ethical vulnerability: the unexamined injection of systemic biases and stereotypes into public discourse. While media guidelines emphasize neutrality and credibility, there is a distinct lack of technical evaluation frameworks capable of auditing multimodal "black-box" models before deployment. Current auditing methods rely on manual, human-driven prompt testing, which is insufficient for uncovering the deep, non-linear biases embedded within massive latent spaces.

    This paper proposes a novel, automated evaluation framework for newsrooms utilizing Stack-based Genetic Programming (StackGP) to actively "red-team" visual generative models. By treating bias detection as a dynamic optimization problem, our system architecture iteratively evolves adversarial news prompts to expose hidden demographic stereotypes, emotional manipulation, and visual homogenization in AI outputs. The framework translates the societal mandate of journalistic integrity into a concrete technical assessment: the evolutionary algorithm mutates benign news prompts (e.g., "a corporate CEO") and uses visual bias classifiers as the fitness function, systematically hunting for prompts that force the GAI into biased visual generation. This proposed architecture provides impacted stakeholders with a robust, scalable tool to quantify and mitigate under-considered representational harms. By shifting from manual oversight to evolutionary auditing, we offer a technical standard that empowers newsrooms to maintain editorial credibility and audience trust in an era of automated content production.

  • When 95% Accuracy Isn't Good Enough: Bridging the Gap Between Forensic Biometric Facial Recognition Benchmarks and Real-World Human Impact Justin Norman
    Abstract

    Biometric facial recognition systems are increasingly deployed in high-stakes forensic contexts, yet their reported accuracies, often exceeding 95%, are derived from evaluations conducted under favorable laboratory conditions that fail to represent the complexity of real-world use cases. This disconnect between lab performance and reality has serious consequences for civil liberties, as institutions and the public increasingly place confidence in systems whose capabilities in applied settings remain poorly characterized.

    I present a body of work that addresses this gap. First, I introduce a publicly available, model-agnostic forensic evaluation framework modeled on eyewitness identification procedures, in which probe images are matched against perceptually similar decoys rather than easily distinguishable alternatives. Evaluating leading architectures (FaceNet, ArcFace) across more than 200,000 images under controlled forensic conditions, I find that previously reported accuracies drop by 10-30 percentage points relative to standard benchmarks.

    Second, I investigate how generative AI interacts with forensic biometric systems at multiple points in the recognition pipeline. I evaluate whether neural super-resolution, deblurring, and head pose correction can improve recognition of degraded forensic images, and find that generative models frequently hallucinate facial features, producing visually convincing but identity-altered outputs that degrade rather than improve recognition. I additionally examine the growing use of synthetic face datasets in FRT model training and evaluation, identifying the risks of diversity washing, in which statistical diversity in generated faces creates false confidence in system fairness, and consent circumvention, in which synthetic data decouples biometric information from the individuals it represents, complicating regulatory enforcement.

    Drawing on these findings, I propose a governance framework for responsible biometric deployment that addresses institutional oversight, deployment-readiness criteria, safeguards for ethical implementation, and investment in the multidisciplinary expertise required to navigate these technical, legal, and ethical challenges.

  • ``Please Reach Out'': Evaluating LLM Responses to Depression and Anxiety Symptoms with the MH-Disclosure Dataset Micah Benson, Samishti Bhatia, Zhengyang Shan, Afitab Iyigun, Taijah Chavis, Anthony J. Rosellini, Aaron Mueller
    Abstract

    We evaluate how large language models (LLMs) respond when users disclose symptoms of mental health disorders. We propose a manually written dataset based on widely used depression and anxiety diagnostic questionnaires; our dataset, MH-Disclosure, covers 42 symptoms at 4 levels of severity each, with 5 paraphrases per item. We define 3 metrics to analyze LLM responses: awareness of a possible mental health disorder, referrals to professionals, and inclusion of emergency support hotlines. Using the proposed dataset and metrics, we find that models often fail to refer or provide support hotlines appropriately given high-severity prompts, yet also overdiagnose given low-severity prompts. We observe high variance in each metric across models. In follow-up experiments, instructing models to roleplay a fantasy character significantly reduces referral rates. We believe these findings reflect a lack of shared standards among model developers for determining how LLMs should respond to expressions of distress in conversational settings. We call for those (post-)training LLMs to work closely with psychologists to determine the optimal format and severity thresholds for safety filters when interfacing with the mental health of users.

  • Tamper resistant Machine Unlearning Algorithms for Open-weight AI Models Amro Abdalla, Ismail Shaheen, Dan DeGenaro, Rupayan Mallick, Bogdan Raita, Sarah Adel Bargal
    Abstract

    Text-to-image (T2I) models have emerged as powerful generative tools capable of producing highquality images faithful to input prompts. However, their accessibility and adaptability make them vulnerable to malicious fine-tuning, where adversaries adapt pre-trained models to generate harmful or copyrighted content. Methods like DreamBooth [24], LoRA, and Textual Inversion enable this adaptation with minimal resources and without needing to train from scratch. This vulnerability persists even when existing safety mechanisms, such as safety checkers or concept erasure methods, are in place, as they can be bypassed, disabled, or undone through lightweight adaptation techniques. This creates a significant risk: once a model is open-sourced, it becomes difficult to guarantee its continued alignment with safety goals. Current defenses either degrade the model’s generative capabilities or fail to withstand adversarial fine-tuning. While safety checkers and licensing agreements offer a first line of defense, they are not an inherent property of T2I models and are easily circumvented. To enhance the inherent safety of T2I models, concept erasure techniques have been proposed to remove undesirable concepts by modifying the model’s internal representations. Although such techniques can suppress the generation of undesired concepts, they are vulnerable to circumvention. Moreover, as we show in our experiments, simple fine-tuning can reintroduce the erased concepts, undermining the long-term effectiveness of concept erasure methods as a safety mechanism.

    To address the limitations of concept erasure and defend against its circumvention, model immunization has been proposed as a proactive defense against malicious fine-tuning of T2I models. IMMA, for example, introduces a bi-level optimization approach inspired by MAML, aiming to learn poor model initializations that hinder adaptation to undesirable concepts. By simulating the fine-tuning process during immunization, IMMA makes it more difficult for adversaries to reintroduce harmful content through fine-tuning. However, IMMA’s framework significantly compromises the model’s performance on safe concepts, degrading both its generative quality and its ability to be fine-tuned for benign applications. To this end, we propose GIFT—a Gradient-aware Immunization framework to defend T2I diffusion models against malicious Fine-Tuning while preserving their ability to generate safe content. GIFT's immunization objective is composed of two parts: (1) a loss maximization term and (2) a representation noising term. We demonstrate that immunizing a model with GIFT significantly impairs its ability to re-learn harmful content while maintaining generative ability across a wide range of safe concepts. Our evaluation covers several concept categories—including objects, art styles, and NSFW content—and considers multiple fine-tuning strategies, e.g., LoRA and DreamBooth.

Call for Participation

We invite poster and lightning talk submissions that document real-world use & impact of generative AI systems and/or leverage human insights to inform the technical design of generative AI and computer vision systems. We especially encourage interdisciplinary work spanning computer vision, machine learning, HCI, safety, security, ethics, and social science.

Poster and lightning talk submissions have closed.

Deadline extended: April 10, 2026 Friday, April 17, 2026 at 11:59 PM Anywhere on Earth.
Go to OpenReview

Important Dates

Submission Overview

Expand submission details Hide submission details

Topics of Interest

We welcome submissions on the following themes:

Topic 1: Human-Centric Findings That Inform AI Design

As AI systems are adopted in everyday settings, they are often used in ways not anticipated by their designers. We welcome work that documents these real-world practices and their implications, including their use in creative writing, journalism, and the generation of sexual content, as well as their impacts on stakeholders due to the theft of intellectual content, and labor displacement. This can also include, but is not limited to:

Topic 2: Technical Designs for Unattended Real-World Needs

Technical standards, requirements, interfaces, or evaluation protocols may need to be adapted in response to observed real-world uses, unmet needs, or harms. We welcome work that explores how this technical gap may be addressed. This can include system designs that accommodate real-world use cases, novel protections against under-considered harms such as artistic style theft, or discussions of how to translate societal or policy goals into concrete technical processes.

Submission Requirements

Each submission should include:

Submissions should be made through OpenReview.

Presentation

Accepted posters are also expected to give a 3-5 minute lightning talk during the workshop. The final talk length will depend on the number of accepted submissions. At least one author of each accepted submission must attend the workshop to present the poster and lightning talk.

We will publish a list of accepted posters and their abstracts on the workshop website.

Workshop Policy

Submissions do not need to be anonymized. The workshop is non-archival and will not appear in the CVPR proceedings.

Organizers

Jaron Mink

Assistant Professor, Arizona State University

David A. Forsyth

Professor, University of Illinois–Urbana-Champaign

Elissa M. Redmiles

Assistant Professor, Georgetown University

Sarah Adel Bargal

Assistant Professor, Georgetown University

Shawn Shan

Assistant Professor, Dartmouth College

Lucy Qin

Fritz Postdoctoral Fellow, Georgetown University

Anand Bhattad

Assistant Professor, Johns Hopkins University

Shiry Ginosar

Assistant Professor, Toyota Technical Institute at Chicago

Eunice Yiu

Postdoctoral Scholar, University of California, Berkeley

Contact

For questions, email jaron.mink@asu.edu.