Week 2 / January 2025

AI Integration Creates Calibration Problems, Not Just Solutions

From RLHF preference manipulation to automation inequality, deployment reveals messy adaptation work

Synthesized using AI

Analyzed 89 papers. AI models can occasionally hallucinate, please verify critical details.

AI systems are creating new alignment problems faster than they solve old ones. Research on RLHF preference elicitation demonstrates that interface design systematically manipulates user preferences—participants shown identical AI outputs in different framings consistently shift their stated preferences based on presentation order, comparison structure, and visual emphasis. This isn't noise in the data; it's evidence that preference collection is a design problem masquerading as a measurement problem. Supporting work on AI-assisted programming shows developers report major productivity gains while empirical measurements find modest improvements and increased cognitive load during verification, and studies of autonomous vehicle interpretation reveal that multimodal explanations improve trust but users still struggle to evaluate system decisions in real-time. The pattern: AI integration assumes stable human preferences and reliable human verification, but deployment reveals both are artifacts of system design.

Work on automation preferences complicates this further by documenting that what people want automated varies systematically by gender and income. Analysis of time-use data paired with automation desire surveys shows women prioritize different household tasks than men, and lower-income groups show distinct patterns from higher-income groups—contradicting the assumption that time-savings universally drive adoption. Research on urban segregation patterns in mobility data and smart city service adoption disparities reinforces this: technical systems encode and amplify existing inequalities when designers treat 'helpfulness' as universal rather than culturally and economically contingent. The implication isn't that AI systems can't be helpful—it's that helpfulness requires explicit negotiation of whose preferences matter and whose verification capacity the system assumes, not just better algorithms.

Featured(1/5)

2501.06348

Why Automate This? Exploring Correlations between Desire for Robotic Automation, Invested Time and Well-Being

Ruchira Ray, Leona Pang, Sanjana Srivastava, Li Fei-Fei, Samantha Shorey, Roberto Martín-Martín

Preprint·2025-01-10

Stop pitching robot features on time savings alone. Audit your automation roadmap against emotional burden, not task duration. Prioritize tasks users report as unpleasant, even if they're quick.

Teams build robots for tasks users don't want automated. The mismatch: engineers optimize for time savings, but users care about emotional burden.

Method: Cross-referenced BEHAVIOR-1K dataset with American Time Use Survey data to map 50 household tasks against time investment and emotional valence. Found that desire to automate correlates more strongly with negative affect during tasks than with time spent—people want robots for emotionally draining chores, not just long ones. Gender patterns emerged: women prioritized automating care tasks despite lower time investment.

Caveats: Dataset limited to household tasks; workplace automation preferences may follow different patterns around professional identity.

Reflections: Do automation preferences shift when tasks involve social interaction or caregiving relationships? · How do income-based differences in automation desire translate to willingness to pay? · Can real-time affect sensing during tasks predict automation acceptance better than retrospective surveys?

social-computingevaluation-methodssustainability

2501.03603

Composing Data Stories with Meta Relations

Haotian Li, Lu Ying, Leixian Shen, Yun Wang, Yingcai Wu, Huamin Qu

Preprint·2025-01-07

Stop auto-generating linear data stories from correlation matrices. Build authoring tools that let analysts explicitly tag narrative intent—whether they're building contrast, showing causality, or establishing progression. Surface meta relation options as first-class UI elements.

Auto-generated data stories feel generic because tools only see data-level patterns. They miss the narrative logic humans use to connect findings.

Method: Introduced meta relations—seven types including contrast, causality, and progression—that operate above the data layer to structure narrative flow. Built a system where users tag findings with meta relations, then the tool generates story sequences using a constraint satisfaction algorithm. Formative study with 12 participants revealed humans rely on domain knowledge and audience context to select relations, not just statistical patterns in the data.

Caveats: Requires manual tagging of meta relations; doesn't auto-detect narrative intent from data alone.

Reflections: Can LLMs infer meta relations from analyst comments or draft text during exploratory analysis? · How do meta relation choices vary across industries (journalism vs. business intelligence)? · What's the minimum set of meta relations needed for 80% of data storytelling scenarios?

data-visualizationdesign-toolsai-interaction

2501.02684

Towards Decoding Developer Cognition in the Age of AI Assistants

Ebtesam Al Haque, Chris Brown, Thomas D. LaToza, Brittany Johnson

Preprint·2025-01-05

Design AI coding tools with cognitive load budgets in mind. Surface confidence scores for suggestions to reduce validation overhead. Optimize for tasks where generation is harder than verification—boilerplate, not architecture.

Developers report massive productivity gains from AI coding assistants, but empirical studies show modest improvements. The gap suggests we're measuring the wrong thing.

Method: Used functional near-infrared spectroscopy (fNIRS) to measure cognitive load in the prefrontal cortex during AI-assisted coding tasks. Tracked 20 developers across bug-fixing and feature-implementation scenarios with and without AI assistance. Found that AI assistants reduce cognitive load during routine tasks but increase it during debugging AI-generated code—developers spend mental effort validating suggestions rather than writing from scratch.

Caveats: fNIRS measures prefrontal cortex activity; doesn't capture full cognitive picture or long-term learning effects.

Reflections: Does repeated AI assistant use atrophy certain coding skills over time? · Can we detect when developers are blindly accepting suggestions versus critically evaluating them? · What's the cognitive load difference between inline suggestions versus chat-based code generation?

programming-toolsai-interactionevaluation-methods

1 / 5

Featured

Findings(1/5)

Error detection shifts from task models to social signal inference·AI assistance creates a perception-measurement gap in productivity·Automation preference correlates with affect, not time investment·Preference modeling shifts from describing humans to shaping them·Data storytelling moves from data-level relations to meta-narrative structures

Robots are learning to detect their own failures by reading human reactions rather than relying on task-specific models or explicit user reports. Papers 2501.05723 and 2501.14792 demonstrate systems that infer error states from naturally occurring social cues—facial expressions, body language, compensatory movements—enabling flexible, real-time detection across contexts. This inverts the traditional approach of hardcoding failure modes. The implication: robots become context-adaptive not through better world models, but through better human models.

2501.05723

Robot Error Awareness Through Human Reactions: Implementation, Evaluation, and Recommendations

2501.14792

A Wearable Strain-Sensor-Based Shoulder Patch for Fatigue Detection in Bicep Curls

2501.05322

"What's Happening"- A Human-centered Multimodal Interpreter Explaining the Actions of Autonomous Vehicles

Surprises(1/3)

ChatGPT generates accessible code only when explicitly prompted, not by default·Federated learning solves the wrong privacy problem for human sensing·Sleep tracking in high-pressure cultures amplifies the problem it measures

We assumed foundation models trained on vast code repositories would internalize accessibility standards as best practice. Paper 2501.03572 shows ChatGPT's default code generation consistently fails WCAG compliance, only meeting standards when accessibility is explicitly requested. The model learned code patterns, not the values embedded in standards. Accessibility isn't emergent from scale—it requires deliberate constraint.

2501.03572

From Code to Compliance: Assessing ChatGPT's Utility in Designing an Accessible Webpage -- A Case Study

TOOLBOX(7)

Remex

Tool

AI-powered data storytelling tool that suggests meta relations during the analysis stage and applies them for organizing data stories. Enables users to compose findings into stylized narratives by leveraging meta relations beyond dataset-level information. Used in exploratory user study to evaluate AI involvement in storytelling workflows.

2501.03603

InclusiViz

Tool

Visual analytics system for multi-level analysis of urban segregation using human mobility data. Integrates deep learning model to predict mobility patterns across social groups, augmented with explainable AI. Features innovative visualizations for exploring segregation patterns and evaluating urban planning interventions with real-time feedback for creating inclusive cities.

2501.03594

HMI (Human-centered Multimodal Interpreter)

Tool

Multimodal interpreter system for autonomous vehicles that provides visual, textual, and auditory feedback. Combines Bird's Eye View interface, map display, text, and voice interaction using fine-tuned large language model. Provides real-time, context-sensitive explanations of vehicle actions to increase passenger trust by over 8%.

2501.05322

ExoFabric

Tool

Re-moldable fabric system created by embedding thermoplastic threads into fabric through computerized embroidery. Enables tunability between rigid plastic and conformable fabric states. Includes design primitive library for geometric formability, stiffness, and stretchability. Supports applications in form-fitting wearables, structural support, and shape-changeable furniture.

2501.05664

ScooterLab

Framework

Community research testbed comprising fleet of customizable battery-powered micromobility vehicles retrofitted with advanced sensing, communication, and control capabilities. Enables researchers to design and deploy customized sensing experiments and access curated datasets for interdisciplinary research in computing, mobility, and urban planning.

2501.06177

MinMo

Model

Multimodal Large Language Model with approximately 8B parameters for seamless voice interaction. Trained through multi-stage alignment (speech-to-text, text-to-speech, speech-to-speech, duplex) on 1.4 million hours of speech data. Achieves 100ms speech-to-text latency and 600ms full-duplex latency. Supports emotion, dialect, speaking rate control, and voice mimicking.

2501.06282

BEHAVIOR-1K Automation Preference Dataset

Dataset

Dataset combining BEHAVIOR-1K data with American Time-Use Survey and Well-Being Module to analyze correlations between desire for robotic automation, time spent on activities, and associated feelings (happiness, meaningfulness, sadness, pain, stress, tiredness). Includes analysis tool for exploring automation preferences across gender and income levels.

2501.06348

Improving Human-Robot Teaching by Quantifying and Reducing Mental Model Mismatch

Quantifies the gap between what humans *think* they're teaching robots and what robots actually learn. Turns out measuring the mismatch itself improves knowledge transfer.

2501.04156

AdaptiveCoPilot: Design and Testing of a NeuroAdaptive LLM Cockpit Guidance System in both Novice and Expert Pilots

Reads pilot brain activity to adjust LLM guidance cues in real-time cockpit scenarios. Tests whether adaptive interfaces reduce cognitive overload without annoying experts.

2501.03190

Multimodal Machine Learning Can Predict Videoconference Fluidity and Enjoyment

Trains models on audio, video, and interaction patterns to predict when videoconferences feel awkward. The goal: catch friction moments before participants consciously notice them.

2501.05790

Understanding Impact of Human Feedback via Influence Functions

Uses influence functions to trace which specific human feedback samples corrupt RLHF reward models. Reveals how noisy preferences propagate through alignment pipelines.

2502.00016

Large Language Models for Education: ChemTAsk -- An Open-Source Paradigm for Automated Q&A in the Graduate Classroom

Bolts retrieval-augmented generation onto LLMs for graduate chemistry Q&A. Open-source pipeline designed to fail gracefully when models confabulate domain-specific knowledge.

2503.15492

World of ScoreCraft: Novel Multi Scorer Experiment on the Impact of a Decision Support System in Sleep Staging

Deploys AI assistance across multiple sleep scorers to measure impact on accuracy and inter-rater agreement. Finds decision support doesn't eliminate variability—it redistributes it.

2501.04860

Exploring the Use of Robots for Diary Studies

Tests whether robots can replace paper diaries for longitudinal HRI data collection in private spaces. Methodological experiment on robots as research instruments, not just subjects.

2501.05434

GraspR: A Computational Model of Spatial User Preferences for Adaptive Grasp UI Design

Builds preference models for XR grasp interfaces without ground-truth labels—because users can't articulate ideal UI placement while holding objects. Adaptive design from implicit signals.

REFLECTION(4)

Alignment costs more than integration

This week's research reveals a stubborn pattern: adding AI to human workflows doesn't reduce friction—it redistributes it. Systems that promise seamless integration instead demand new forms of calibration, retraining, and trust negotiation, often degrading performance before improving it. The question isn't whether to integrate AI, but whether we're willing to pay the alignment tax.

Clinicians using AI diagnostics show *degraded* performance on subsequent unassisted tasks, yet we frame this as a training problem rather than a system design failure. Are we optimizing for human-AI teaming or just normalizing human dependence on flawed automation?

1 / 4

Week 01January 2025

Week 03January 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—89 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

Aligning Humans and AI Systems

This cluster investigates how humans and AI systems coordinate expectations and behavior. Core questions: How do users calibrate trust in AI outputs? How do mental models between humans and machines diverge, and what interventions reduce misalignment? Research spans preference elicitation in RLHF, real-time cognitive load adaptation in safety-critical domains, and interaction design for robots and conversational agents. Methods combine behavioral observation, neuroimaging, LLM-based feedback mechanisms, and user studies. Audience: HCI researchers, AI safety engineers, and roboticists designing human-centered AI systems.

1/11

Top Papers in this Theme

2501.03190

Multimodal Machine Learning Can Predict Videoconference Fluidity and Enjoyment

2501.04156

AdaptiveCoPilot: Design and Testing of a NeuroAdaptive LLM Cockpit Guidance System in both Novice and Expert Pilots

2501.04755

Improving Human-Robot Teaching by Quantifying and Reducing Mental Model Mismatch

2501.04929

What Drives You to Interact?: The Role of User Motivation for a Robot in the Wild

2501.05790