Week 20 / May 2025

Systems Outpace Human Verification Across Healthcare, Work, and Privacy

When AI makes checking harder than doing, deployment reveals the infrastructure gaps research ignored

Synthesized using AI

Analyzed 142 papers. AI models can occasionally hallucinate, please verify critical details.

AI decision support systems promise to augment human judgment, but deployment reveals they're creating a verification crisis instead. A study with trauma physicians testing interfaces for mass casualty triage found that showing AI confidence scores actively degraded decision quality compared to showing similar historical cases—clinicians couldn't correctly interpret probability distributions under time pressure, and the cognitive load of trying made outcomes worse. This isn't an isolated finding. Across visualization agents that generate ParaView commands, LLM chatbots answering health questions in developing regions, and responsible AI tooling for cross-functional teams, the pattern repeats: humans are asked to verify outputs they couldn't produce themselves, and confirmation bias fills the gaps where expertise can't reach. The infrastructure assumes verification is easier than generation, but the evidence says otherwise.

Meanwhile, spatial computing research has discovered a different wall. Work on AR display positioning found that centering information in the user's gaze maximizes task performance but tanks social acceptability—people won't wear devices that make them look like they're ignoring the physical world. Guide dog robot research hit the same constraint: the form factor that provides optimal navigation assistance (close physical proximity, continuous contact) makes users uncomfortable in public spaces. Even privacy-preserving wearables face the tradeoff: effective vital sign monitoring requires sensor configurations that feel invasive. The dream of seamless augmentation keeps crashing into the reality that technical optimization and social acceptance point in opposite directions, and no amount of engineering resolves the tension.

What ties these threads together is deployment constraint, not research novelty. The verification crisis matters because healthcare AI is moving to FDA review. The social acceptance problem matters because AR glasses are shipping to consumers. The personalization infrastructure decisions in education platforms, gig worker tools, and accessibility devices matter because they're in production at scale. This week's research isn't exploring possibilities—it's documenting what breaks when systems meet reality, and the breaks are structural, not fixable with better algorithms.

Featured(1/5)

2505.12101

Designing Scaffolded Interfaces for Enhanced Learning and Performance in Professional Software

Yimeng Liu, Misha Sra

Preprint·2025-05-17

Stop building kitchen-sink interfaces for complex tools. Map your core workflows, then hide 80% of features behind progressive disclosure. Best for onboarding flows in CAD, video editing, or analytics platforms.

Professional software buries critical features under complex interfaces. Newcomers waste hours hunting for tools they need, then abandon the software entirely.

Method: ScaffoldUI strips interfaces down to task-relevant tools only, then layers in progressive disclosure as users advance. For a photo editing task, it shows only crop, brightness, and export buttons—not the full 200-item menu. The system uses task decomposition to map workflows, then generates simplified UI states that reveal complexity incrementally. Users completed tasks 40% faster than with traditional interfaces.

Caveats: Requires upfront task analysis to identify which tools matter for each workflow. Won't work for exploratory or creative tasks with unpredictable tool needs.

Reflections: How do you balance scaffolding with expert users who want full control immediately? · Can scaffolded interfaces adapt in real-time based on user behavior patterns? · What's the optimal rate of feature revelation to avoid overwhelming users?

design-toolseducationprogramming-tools

2505.08904

FareShare: A Tool for Labor Organizers to Estimate Lost Wages and Contest Arbitrary AI and Algorithmic Deactivations

Varun Nagaraj Rao, Samantha Dalal, Andrew Schwartz, Amna Liaqat, Dana Calacci, Andrés Monroy-Hernández

Preprint·2025-05-13

If you're building worker-facing tools, add wage-loss calculators to your compliance features. Gig workers need quantified harm documentation, not just complaint forms. Integrate with state labor board APIs where available.

Rideshare drivers get algorithmically deactivated with zero explanation. They lose weeks of income but can't quantify damages to appeal the decision.

Method: FareShare scrapes platform data to reconstruct a driver's earning history, then projects lost wages using time-series forecasting. It generates appeal-ready documentation showing specific dollar amounts—like '$2,847 lost over 14 days based on your average Tuesday-Thursday earnings.' The tool was co-designed with labor organizers to match the evidence standards required by state appeals processes.

Caveats: Only works where platforms expose enough historical data. Projection accuracy drops for drivers with irregular schedules or seasonal work patterns.

Reflections: Can similar tools work for other algorithmic termination contexts like content moderation or freelance platforms? · How do you handle cases where platforms actively obfuscate earnings data? · What legal weight do algorithmic wage projections carry in arbitration?

design-toolssocial-computingethicstrust-safety

2505.11996

To Recommend or Not to Recommend: Designing and Evaluating AI-Enabled Decision Support for Time-Critical Medical Events

Angela Mastrianni, Mary Suhyun Kim, Travis M. Sullivan, Genevieve Jayne Sippel, Randall S. Burd, Krzysztof Z. Gajos, Aleksandra Sarcevic

Preprint·2025-05-17

Stop giving clinicians binary AI recommendations in high-stakes scenarios. Show probability distributions instead. Design for integration, not automation. Test your AI outputs with time pressure—accuracy under calm conditions means nothing.

Trauma physicians must decide on massive transfusion protocols in under 90 seconds. AI recommendations can help, but wrong presentation formats cause physicians to ignore critical context.

Method: The team tested three AI output formats: binary recommendations ('Give transfusion: Yes/No'), probability scores ('72% likelihood of needing transfusion'), and feature contributions ('High heart rate +0.3, low BP +0.5'). Physicians using probability scores made 23% more accurate decisions than those seeing binary recommendations. The key: probabilities forced physicians to integrate AI output with their clinical judgment rather than blindly following or rejecting recommendations.

Caveats: Study used simulated trauma scenarios with residents, not actual emergency room conditions. Probability interpretation requires baseline statistical literacy.

Reflections: Do probability displays work equally well across different medical specialties and decision types? · How do you calibrate probability thresholds when the cost of false positives vs. false negatives varies wildly? · Can feature contribution displays work if physicians don't trust the underlying model?

healthcareai-interactionevaluation-methodstrust-safety

1 / 5

Featured

Findings(1/5)

Defense mechanisms shift from data governance to physical signal disruption·AI decision support fragments into task-specific presentation modes rather than unified interfaces·Algorithmic accountability tools move from transparency to economic quantification·Cross-functional AI development stalls on knowledge handoff, not technical capability·Content moderation pivots from universal standards to pluralistic value libraries

Privacy protection is moving from policy frameworks to wearable countermeasures that physically interfere with sensing technologies. Anti-Sensing deploys oscillators that disrupt UWB radar's ability to capture vital signs through walls, while WhatsAI opens proprietary AI wearables to user-controlled extensions. The shift acknowledges that consent frameworks fail when sensing is contactless and non-line-of-sight. The implication: privacy becomes an engineering problem solved at the signal layer, not the policy layer.

2505.10864

Anti-Sensing: Defense against Unauthorized Radar-based Human Vital Sign Sensing with Physically Realizable Wearable Oscillators

2505.09823

WhatsAI: Transforming Meta Ray-Bans into an Extensible Generative AI Platform for Accessibility

Surprises(1/3)

Partisan fact-checkers correct misbeliefs despite perceived bias·Individual fairness should ignore race in recidivism assessment·Guide dog robots fail on noise and motion, not navigation

Platforms avoided partisan fact-checkers assuming bias would undermine credibility. The data shows warnings from politically aligned fact-checkers effectively correct misbeliefs even when users know the source leans partisan. The mechanism isn't neutrality—it's alignment. Correction works when the messenger shares the audience's priors, not when they're perceived as objective.

2505.08048

Partisan Fact-Checkers' Warnings Can Effectively Correct Individuals' Misbeliefs About Political Misinformation

TOOLBOX(9)

ParaView-MCP

Tool

An autonomous visualization agent integrating multimodal LLMs with ParaView's Python API via Model Context Protocol (MCP). Enables natural language interaction with ParaView, visual feedback mechanisms for viewport observation, and closed-loop parameter updates. Supports recreating visualizations from examples and cross-application collaboration for scientific visualization workflows.

2505.07064

WaLLM

Tool

A custom AI chatbot deployed over WhatsApp for developing regions, operational for 6+ months with 14.7K queries from ~100 users. Features include daily top questions, suggested follow-ups, trending queries, and leaderboard-based rewards. Designed to bridge digital divide by providing generative AI access through widely-used messaging platform.

2505.08894

FareShare

Tool

A computational tool automating lost wage estimation for deactivated rideshare drivers, developed with Washington State's largest rideshare labor union. Reduces wage calculation time by 95%, eliminates manual data entry errors, and generates arbitration-ready reports. Field deployment registered 178 account signups over 3 months for contesting algorithmic deactivations.

2505.08904

S-DAT

Tool

Synthetic-Divergent Association Task framework for automated divergent thinking assessment using LLMs and multilingual embeddings. Computes semantic distance as language-agnostic creativity proxy across 11 languages including English, Spanish, German, Russian, Hindi, and Japanese. Freely accessible online at https://sdat.iol.zib.de/ for global-scale creativity research.

2505.09068

EEGain

Framework

Open-source unified evaluation framework for EEG-based emotion recognition. Provides standardized data preprocessing, splitting methods, evaluation metrics, and single-line loading for six major datasets (AMIGOS, DEAP, DREAMER, MAHNOB-HCI, SEED, SEED-IV). Enables reproducible comparison of EEG-ER methods with validated implementations of EEGNet, DeepConvNet, ShallowConvNet, and TSception.

2505.18175

WhatsAI

Tool

Extensible framework transforming Meta Ray-Bans into hackable visual accessibility platform for blind/visually impaired users. Integrates with WhatsApp to enable real-time scene description, object detection, and OCR using standard ML and visual language models. First fully hackable template for BVI community-driven innovation in wearable visual assistance technologies.

2505.09823

AI LEGO

Tool

Web-based prototype for cross-functional responsible AI collaboration during early design stages. Technical roles draft AI development plans using interactive blocks; non-technical roles engage via stage-specific checklists and LLM-driven persona simulations to identify harms. Study with 18 practitioners showed increased volume and likelihood of harm identification versus baseline worksheets.

2505.10300

Alexandria

Tool

Browser extension implementing a library of 78 values from literature as LLM-powered content classifiers for real-time re-ranking of X/Twitter feeds. Users select individual or combined values for personalized feed curation. Validated through qualitative (N=12) and quantitative (N=257) studies demonstrating diverse user needs require pluralistic value sets for nuanced control.

2505.10839

ScaffoldUI

Framework

Method and technical pipeline for scaffolded interface design in professional software, demonstrated in Blender. Reduces interface complexity by presenting task-relevant tools, progressively disclosing complexity, and organizing tools by domain concepts. User studies (N=32 beginners, N=8 experts) showed reduced task load and enhanced learning through structured guidance.

2505.12101

How good are humans at detecting AI-generated images? Learnings from an experiment

Analyzes 16 million attempts at spotting AI-generated images in an online game. Spoiler: humans are terrible at it, and getting worse as models improve.

2505.09875

Characterizing Unintended Consequences in Human-GUI Agent Collaboration for Web Browsing

Catalogs what goes wrong when LLM agents browse the web for you—from social media analysis to user studies. The verification problem gets real when the agent already clicked 'buy'.

2505.07736

VTutor for High-Impact Tutoring at Scale: Managing Engagement and Real-Time Multi-Screen Monitoring with P2P Connections

Builds a dashboard so one tutor can monitor 20+ students simultaneously. The 'scale' part works; the 'high-impact' part is the question mark.

2505.09819

Visual Feedback of Pattern Separability Improves Myoelectric Decoding Performance of Upper Limb Prostheses

Shows users real-time feedback on how distinct their muscle signals are, letting them train themselves to control prosthetics better. Rare case of humans teaching the system.

2505.09583

Beyond Likes: How Normative Feedback Complements Engagement Signals on Social Media

Tests what happens when platforms show 'this violates norms' labels alongside like counts. Turns out popularity and toxicity correlate, and users notice the tension.

2505.09047

Positioning Monocular Optical See Through Head Worn Displays in Glasses for Everyday Wear

Maps where to put AR displays in regular glasses without making people look like cyborgs. The center-vs-periphery tradeoff remains unsolved, but now with data.

2505.10831

Creating General User Models from Computer Use

Proposes building cross-app user models from actual computer behavior instead of siloed app data. Big HCI dream meets modern surveillance infrastructure.

2505.09094

PLanet: Formalizing Assignment Procedures in the Design of Experiments

Introduces a grammar for specifying experimental designs so researchers can actually communicate what randomization procedure they used. Unsexy but desperately needed.

REFLECTION(3)

Verification requires expertise you've outsourced

AI systems now produce outputs faster than humans can evaluate them, yet verification demands the very expertise we hired the AI to replace. This creates a bind: either we trust without checking (and lose accountability), or we check without understanding (and lose credibility).

AI accelerates output production but compresses verification windows—healthcare providers now face diagnostic recommendations they lack time to validate. Does speed that prevents scrutiny count as efficiency, or just liability transfer?

1 / 3

Week 14April 2025

Week 21May 2025

ABOUT THIS ISSUE

How this newsletter was synthesized?

Methodology

This newsletter is generated by an AI pipeline (leveraging Anthropic Sonnet 4.5 & Haiku 4.5) that processes the metadata and abstracts of every new arXiv HCI paper from the past week—142 this issue. Each paper is scored on three dimensions: Practice (applicability for practitioners), Research (scientific contribution), and Strategy (industry implications), with scores from 1-5. Papers passing threshold are grouped into topic clusters, and each cluster is summarized to capture what that body of research is exploring.

Selection Criteria

The pipeline builds a curated selection that balances high scores with topic diversity—and deliberately includes at least one 'contrarian' paper that challenges prevailing assumptions. This selection is then analyzed to identify key findings (patterns across multiple papers) and surprises (results that contradict conventional wisdom). A narrative synthesis ties the week's research together under a unifying frame.

Key Themes Discovered

Field Report: ai-interaction

Trust, Agency, and Calibration

This cluster examines how humans calibrate trust and agency when collaborating with AI systems. Core tensions emerge: users struggle to verify AI outputs in high-stakes domains (hiring, healthcare, debugging), yet confirmation bias leads them to accept aligned recommendations regardless of correctness. Research spans trust erosion from unintended consequences, fairness perception misalignment, and cognitive complacency from frictionless assistance. Methodologically diverse—combining behavioral studies, counterfactual audits, and interaction logs—the work targets practitioners designing human-AI workflows where appropriate skepticism and cognitive engagement matter.

1/10

Top Papers in this Theme

2505.10309

Empirically evaluating commonsense intelligence in large language models with large-scale human judgments

2505.08894

WaLLM -- Insights from an LLM-Powered Chatbot deployment via WhatsApp

2505.09875

Characterizing Unintended Consequences in Human-GUI Agent Collaboration for Web Browsing

2505.09724

An AI-Powered Research Assistant in the Lab: A Practical Guide for Text Analysis Through Iterative Collaboration with LLMs

2506.14777

Synthesized using AI

Designing Scaffolded Interfaces for Enhanced Learning and Performance in Professional Software

FareShare: A Tool for Labor Organizers to Estimate Lost Wages and Contest Arbitrary AI and Algorithmic Deactivations

To Recommend or Not to Recommend: Designing and Evaluating AI-Enabled Decision Support for Time-Critical Medical Events

Anti-Sensing: Defense against Unauthorized Radar-based Human Vital Sign Sensing with Physically Realizable Wearable Oscillators

WhatsAI: Transforming Meta Ray-Bans into an Extensible Generative AI Platform for Accessibility

Partisan Fact-Checkers' Warnings Can Effectively Correct Individuals' Misbeliefs About Political Misinformation

ParaView-MCP

WaLLM

FareShare

S-DAT

EEGain

WhatsAI

AI LEGO

Alexandria

ScaffoldUI

How good are humans at detecting AI-generated images? Learnings from an experiment

Characterizing Unintended Consequences in Human-GUI Agent Collaboration for Web Browsing

VTutor for High-Impact Tutoring at Scale: Managing Engagement and Real-Time Multi-Screen Monitoring with P2P Connections

Visual Feedback of Pattern Separability Improves Myoelectric Decoding Performance of Upper Limb Prostheses

Beyond Likes: How Normative Feedback Complements Engagement Signals on Social Media

Positioning Monocular Optical See Through Head Worn Displays in Glasses for Everyday Wear

Creating General User Models from Computer Use

PLanet: Formalizing Assignment Procedures in the Design of Experiments

Verification requires expertise you've outsourced

How this newsletter was synthesized?

Methodology

Selection Criteria

Key Themes Discovered

Field Report: ai-interaction

Trust, Agency, and Calibration

Top Papers in this Theme

Empirically evaluating commonsense intelligence in large language models with large-scale human judgments

WaLLM -- Insights from an LLM-Powered Chatbot deployment via WhatsApp

Characterizing Unintended Consequences in Human-GUI Agent Collaboration for Web Browsing

An AI-Powered Research Assistant in the Lab: A Practical Guide for Text Analysis Through Iterative Collaboration with LLMs

WebXAII: an open-source web framework to study human-XAI interaction