This work was done with my amazing colleagues at UC Irvine and Northwestern University–Xinru Tang, Jimin Heo, Dwayne R. Morgan, Darren Gergle, Erik B. Sudderth, and Anne Marie Piper–with financial support from the National Science Foundation.

"It's trained by non-disabled people": Evaluating How Image Quality Affects Product Captioning with Vision-Language Models Vision-Language Models (VLMs) are increasingly used by blind and low-vision (BLV) people to identify and understand products in their everyday lives, such as food, personal care items, and household goods. Despite their prevalence, we lack an empirical understanding of how common image quality issues--such as blur, misframing, and rotation--affect the accuracy of VLM-generated captions and whether the resulting captions meet BLV people's information needs. Based on a survey of 86 BLV participants, we develop an annotated dataset of 1,859 product images from BLV people to systematically evaluate how image quality issues affect VLM-generated captions. While the best VLM achieves 98% accuracy on images with no quality issues, accuracy drops to 75% overall when quality issues are present, worsening considerably as issues compound. We discuss the need for model evaluations that center on disabled people's experiences throughout the process and offer concrete recommendations for HCI and ML researchers to make VLMs more reliable for BLV people.

To learn more:
📝 Paper: arxiv.org/abs/2511.08917
🎙️ Talk: Power, Values, and the Politics of Accessibility (Mon, 13 Apr at 11:51 AM; P1 - Room 112)
📊 Dataset: github.com/Accessibilit...
🔍 Dataset Browser: huggingface.co/spaces/kgarg...

#HCI #AI #accessibility

A research poster titled "It’s trained by non-disabled people": Evaluating How Image Quality Affects Product Captioning with Vision-Language Models published at CHI 2026. It presents two studies on how blind and low-vision (BLV) people use AI for product identification and how image quality affects model performance. The left side reports survey results from 86 BLV users with key results including: 55% use AI mostly at home, many prefer human help in stores when browsing products (49%) or searching for products (55%) due to reliability and speed, and most rely on humans for critical details like allergens while using AI for general identification; 69% say image quality impacts captions. The right side shows a line chart of the accuracies of four models, which drop sharply as image quality worsens (from ~95–98% to ~28–69%). An example shows a blurry cake-mix box that two GPT-4.1 and Gemini 2.5 Flash incorrectly captioned as lasagna and cat food, respectively.

🎉 Our #CHI2026 paper received a 🏅𝗕𝗲𝘀𝘁 𝗣𝗮𝗽𝗲𝗿 𝗛𝗼𝗻𝗼𝗿𝗮𝗯𝗹𝗲 𝗠𝗲𝗻𝘁𝗶𝗼𝗻! 🎉

We study how VLM tools struggle to accurately identify product information for BLV people through a large survey and an evaluation of four VLMs on a dataset of 1,859 images we developed.

If you're at CHI, come see my talk!

A poster advertising a paper talk: What Remotely Matters? Understanding Individual, Team, and Organizational Factors in Remote Work at Scale by Kapil Garg, Diego Gómez-Zará, Elizabeth Gerber, Darren Gergle, Noshir Contractor, and Michael Massimi. The talk is at CSCW on Tuesday, October 21, 2025, in Smatroll. There are 2 columns. The first describes the problem and research questions. Problem: Work is collaborative, but does it need to be in-person? Workplaces seem to say yes, but the discussions about how to make knowledge workers effective focus only on where we should work instead of how we can work better when we’re distributed. Research Questions: 1. What factors do knowledge workers perceive as enablers and barriers of remote or hybrid work activities? 2. What individual, team, and organizational-level differences affect workers’ perceptions of the enablers and barriers of remote work? The second describes key findings from a theoretically-informed survey of 1,526 U.S. knowledge workers: 1. Personal Familiarity Matters Less: Workers didn’t need to know their teammates on a personal level, but did need a strong sense of belonging on the team, which newer teams struggled with 2. Workers and Managers Value Different Factors: Managers value observability; workers value autonomy. 3. Hybrid Teams Struggle More Than Virtual Teams: Managing multiple modalities introduces friction to collaborative work, especially when norms for how to work aren't clear. 4. Reliability of Tools to Support Work: Routine inconveniences when collaborating (e.g., microphone off, missing document access) derail collaboration. 5. Norm Setting for Tightly Coupled Work: Highly collaborative work requires norm setting, which newly-formed teams may lack and larger teams may outgrow.

Every day, another company is pushing for RTO. But do we need to? Our #CSCW2025 paper argues that RTO discussions should move beyond “where should we work” towards “how can we work better when we’re distributed”.

Come to my talk on Tuesday at 2:30pm and read our paper!
dl.acm.org/doi/10.1145/...