Week 7 / February 2026

Human Involvement Shapes AI Outcomes Independent of Technical Capability

Controlled studies isolate how attribution, guidance, and interpretation alter collaborative performance

Synthesized using AI

Analyzed 121 papers. AI models can occasionally hallucinate, please verify critical details.

AI systems are entering collaborative workflows where human involvement shapes outcomes independent of technical capability. A controlled experiment with 60 participants building web apps over 90 minutes found that human-guided AI pairs completed 23% more functional requirements than AI working alone, despite the AI-only condition generating more total code. The mechanism: humans provided strategic direction that prevented the AI from pursuing technically valid but functionally irrelevant solutions—a volume-versus-value trap. This pattern extends beyond code generation. Learners who believed identical LLM feedback came from a human TA engaged more deeply than those told it came from AI, spending more time revising and making more substantive changes despite content being held constant. Guardian Angel demonstrates a different facet: LLMs can operate directly on raw, noisy sensor waveforms to deliver health guidance, achieving 94.12% validity on signals that break traditional algorithms 29.71% of the time. The system still requires human interpretation of plain-language output, but it shifts the interface burden from user articulation to system interpretation.

The accessibility infrastructure work complicates this picture by revealing that human involvement isn't just about collaboration—it's about who gets to participate at all. Blind and low-vision researchers report that 89% encounter barriers in at least one research workflow stage, with data analysis and visualization most problematic. The issue isn't that accessibility features are missing; it's that visual representations dominate research infrastructure by design. Decades of retrofitting haven't solved this because the problem is architectural, not implementational. Research tools being rebuilt around AI represent a rare opportunity to design non-visual modalities—data sonification, tactile representations, semantic descriptions—as first-class citizens rather than compliance add-ons. The tension: we're instrumenting AI systems to augment human capabilities while simultaneously encoding assumptions about which humans and which capabilities matter.

Featured(1/5)

2602.10473

Why Human Guidance Matters in Collaborative Vibe Coding

Haoyu Hu, Raja Marjieh, Katherine M Collins, Chenyi Li, Thomas L. Griffiths, Ilia Sucholutsky, Nori Jacoby

Preprint·2026-02-11

Human-guided AI pairs completed 23% more functional requirements and produced code rated higher on maintainability. The AI-only condition generated more lines of code but lower feature completeness—a volume-versus-value trap.

AI code generation through natural language—"vibe coding"—is popular, but its impact on productivity and the role of humans in the process remains unclear.

Method: Researchers ran a controlled experiment where 60 participants built a web app in pairs over 90 minutes. One condition used AI alone (GPT-4), another paired human guidance with AI. The human-guided condition produced significantly more functional features and higher code quality. The mechanism: humans provided strategic direction and context that prevented the AI from pursuing technically valid but functionally irrelevant solutions.

Caveats: 90-minute sessions with predefined tasks. Longer projects with evolving requirements may show different patterns.

Reflections: At what project complexity does human guidance become essential versus optional? · Can AI be trained to recognize when it needs strategic human input? · How does the guidance pattern change for expert versus novice developers?

ai-interactioncollaborationprogramming-tools

2602.09522

Earinter: A Closed-Loop System for Eating Pace Regulation with Just-in-Time Intervention Using Commodity Earbuds

Jun Fang, Ka I Chan, Xiyuxing Zhang, Yuntao Wang, Mingze Gao, Leyi Peng, Jiajin Li, Zihang Zhan, Zhixin Zhao, Yuanchun Shi

Preprint·2026-02-10

Deploy passive pace monitoring in existing consumer hardware—no custom wearables needed. Best for interventions where unobtrusiveness matters: users reported the system felt natural and didn't interrupt conversations or enjoyment.

Rapid eating is common and unhealthy, but people rarely notice pace changes and sustained self-monitoring is effortful during meals.

Method: Earinter repurposes commodity earbuds' bone-conduction voice sensors to capture chewing vibrations and estimate eating pace as chews per minute. When pace exceeds a personalized threshold, it delivers just-in-time audio nudges. A 28-day field study with 20 participants showed the system reduced eating pace by 16.8% on average and extended meal duration by 2.1 minutes without disrupting the eating experience.

Caveats: Tested on adults without eating disorders. Clinical populations and children may respond differently to pace feedback.

Reflections: Does sustained use over months maintain effectiveness or trigger habituation? · Can the system detect and adapt to different food textures that affect chewing patterns? · Would social eating contexts (family meals, restaurants) require different intervention timing?

wearableshealthcareevaluation-methods

2602.11311

Same Feedback, Different Source: How AI vs. Human Feedback Shapes Learner Engagement

Caitlin Morris, Pattie Maes

Preprint·2026-02-11

Don't assume AI feedback will be processed the same as human feedback, even when quality is identical. If depth of engagement matters, consider hybrid attribution or explicit framing that signals the feedback's pedagogical intent, not just its source.

As AI is used alongside human instructors, learners' beliefs about feedback source may shape how they engage with it, but source attribution effects remain unclear.

Method: Researchers designed a creative coding interface where all participants received identical LLM-generated feedback, but half saw it attributed to AI and half to a human TA. The study isolated source attribution while controlling content. Learners who believed feedback came from a human spent significantly more time revising their code and made more substantive changes. The AI-attributed group treated feedback more superficially despite identical content quality.

Caveats: Single-session creative coding task. Longitudinal effects and other domains (writing, math) may differ.

Reflections: Does transparency about AI authorship reduce engagement, or does learner expectation mismatch cause the effect? · Can AI feedback be framed to trigger the same engagement depth as human feedback? · Do expert learners show the same attribution bias as novices?

ai-interactioneducationtrust-safety

1 / 5

Featured

Findings(1/5)

Conversational interfaces move from text-first to sensor-grounded reasoning·Authentication architectures shift from device-specific to cross-device physiological bridges·Age verification fails at the behavioral layer despite policy compliance·Threat modeling expands from technical vulnerabilities to interconnected systemic risks·Institutional data-making prioritizes organizational imperatives over representational accuracy

LLM-based systems are coupling directly with live sensor streams rather than requiring users to describe context. Kissan-Dost processes soil and climate measurements into plain-language guidance over WhatsApp, while Guardian Angel operates on raw PPG waveforms from low-cost hardware, handling signal quality issues that render 29.71% of data invalid for traditional algorithms. This shifts the interface burden from user articulation to system interpretation, particularly valuable where digital literacy is low.

2602.08593

Kissan-Dost: Bridging the Last Mile in Smallholder Precision Agriculture with Conversational IoT

2602.08701

LLM-Enhanced Wearables for Comprehensible Health Guidance in LMICs

Surprises(1/3)

Attribution effects matter more than feedback quality in learning systems·Human guidance remains essential even when AI can generate code from natural language·Professional research tools systematically exclude the researchers they claim to serve

Learners who believed identical LLM-generated feedback came from a human TA engaged differently than those told it came from AI, despite content being held constant. The source attribution shaped engagement independent of feedback quality. This inverts the assumption that optimizing feedback content is the primary lever—belief about origin may matter as much as the feedback itself when designing hybrid AI-human educational systems.

2602.11311

Same Feedback, Different Source: How AI vs. Human Feedback Shapes Learner Engagement

TOOLBOX(8)

T2VTree

Tool

Visual analytics system for agent-assisted thought-to-video authoring. Represents authoring as a tree where each node binds editable specifications (intent, inputs, workflow, prompts, parameters) with multimodal outputs. Integrates collaborating agents that translate step-level intent into executable plans, supporting branching authoring with in-place preview and stitching for multi-scene video creation.

2602.08368

Kissan-Dost

Tool

Multilingual, sensor-grounded conversational system for smallholder agriculture. Couples commodity soil and climate sensors with retrieval-augmented generation to deliver plain-language guidance over WhatsApp text or voice. Achieved over 90 percent correctness on 99 sensor-grounded crop queries with subsecond end-to-end latency in controlled tests. Modular pipeline enforces grounding, traceability, and proactive alerts.

2602.08593

Guardian Angel

Tool

Low-cost, screenless wearable paired with WhatsApp-based LLM agent for personal health monitoring in LMICs. Operates directly on raw, noisy sensor waveforms with 100 percent availability coverage under field noise conditions (compared to 70.29 percent valid outputs from standard algorithms). Delivers plain-language, personalized physiological insights to users with low digital literacy.

2602.08701

PPGTransID

Tool

Cross-device authentication system leveraging photoplethysmography (PPG) signals for smart wearables. Compares PPG signals from wearable devices with remote PPG (rPPG) extracted from smartphone camera to transfer authentication status. Achieves 95.5 percent balanced accuracy across multiple wearable form factors (rings, earphones, glasses, wristbands) with resilience to lighting, camera placement, and user behavior variations.

2602.08972

Earinter

Tool

Commodity-earbud-based closed-loop system for eating pace regulation. Repurposes earbud's bone-conduction voice sensor to capture chewing vibrations and estimate eating pace as chews per swallow (CPS). Achieves F1 = 0.97 for chewing detection and MAE of 0.18 ± 0.13 chews/min, 3.65 ± 3.86 chews/swallow for pace estimation. Delivers theory-grounded just-in-time interventions during daily meals.

2602.09522

BabyMamba-HAR

Framework

Lightweight selective state space model framework for human activity recognition on resource-constrained devices. Comprises two architectures: CI-BabyMamba-HAR (channel independent) and Crossover-BiDir-BabyMamba-HAR (early fusion). Achieves 86.52 percent average macro F1-score with approximately 27K parameters and 2.21M MACs, requiring 11x fewer MACs than TinyHAR on high-channel datasets while matching accuracy.

2602.09872

ECHO

Framework

Open research platform (Evaluation of Chat, Human behavior, and Outcomes) for reproducible, mixed-method studies of human interaction with conversational AI and Web search. Orchestrates end-to-end experimental workflows integrating consent surveys, chat/search sessions, writing tasks, and evaluations within unified framework. Logs fine-grained interaction traces and exports structured datasets for downstream analysis.

2602.10295

LCIP

Tool

Loss-Controlled Inverse Projection method for high-dimensional image data. Maps 2D projection space back to data space to support data augmentation, classifier analysis, and imputation. Works generically with any projection technique and dataset, controlled by two user-set parameters. Enables 'sweeping' the data space under user control rather than generating fixed surface-like structures.

2602.11141

Large Language Models in Peer-Run Community Behavioral Health Services: Understanding Peer Specialists and Service Users' Perspectives on Opportunities, Risks, and Mitigation Strategies

Examines what happens when conversational AI enters peer-run mental health support—where trust is built on lived experience, not algorithmic scale. The opacity problem gets existential here.

2602.07802

TouchScribe: Augmenting Non-Visual Hand-Object Interactions with Automated Live Visual Descriptions

Generates real-time visual descriptions as blind users touch objects. Finally addresses the gap between tactile information and visual features like color, text, or wear patterns.

2602.11567

Behavioral Indicators of Overreliance During Interaction with Conversational Language Models

Identifies behavioral signals that reveal when users are trusting LLM outputs too much. Turns the fuzzy concept of overreliance into something you can actually measure during interaction.

2602.10403

Disability-First AI Dataset Annotation: Co-designing Stuttered Speech Annotation Guidelines with People Who Stutter

Co-designs annotation guidelines with people who stutter instead of outsourcing to crowdworkers. Shows why expertise matters more than volume when labeling disability data.

2602.10701

Don't blame me: How Intelligent Support Affects Moral Responsibility in Human Oversight

Tests whether intelligent support systems make human overseers feel less responsible when things go wrong. Spoiler: they do, which complicates the entire human-in-the-loop narrative.

2602.10001

Human-AI Synergy Supports Collective Creative Search

Uses a word-guessing game to measure whether AI actually improves collective creativity or just makes everyone converge on similar ideas. The controlled task design is clever.

2602.09629

Stop Testing Attacks, Start Diagnosing Defenses: The Four-Checkpoint Framework Reveals Where LLM Safety Breaks

Proposes diagnosing where safety mechanisms fail rather than just proving jailbreaks work. Shifts focus from "can we break it" to "which component broke and why."

2602.08830

Enhancing Generative AI Image Refinement with Scribbles and Annotations: A Comparative Study of Multimodal Prompts

Compares text prompts, inpainting, and scribbles for refining AI-generated images. Addresses the gap between rapid ideation and precise control in generative design tools.

REFLECTION(4)

Users want comfort, systems need friction

The research reveals a stubborn pattern: people consistently prefer AI interactions that feel frictionless and advisory—yet these same preferences produce worse outcomes and erode their own agency. This isn't a failure of communication or design education; it's a collision between what feels safe and what actually works.

Users resist monitoring systems while demanding personalized support, yet personalization requires the very data collection they reject. If you remove the friction of transparency, do you gain trust or just hide the trade-off?

1 / 4

Week 06February 2026

ABOUT THIS ISSUE

How was this newsletter 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—121 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, Control, and Collaboration

This cluster investigates how humans calibrate trust, distribute control, and coordinate with AI systems across high-stakes domains. Core tensions emerge: users prefer AI advisors but achieve better outcomes with autonomous delegates; feedback source shapes engagement more than content; hybrid systems perform best when humans retain directional authority. Research spans clinical decision-making, peer support, education, and creative work, consistently revealing misalignments between user preferences and system performance, and between normative trust frameworks and behavioral trust formation mechanisms.

1/10

Top Papers in this Theme

2602.08187

Large Language Models in Peer-Run Community Behavioral Health Services: Understanding Peer Specialists and Service Users' Perspectives on Opportunities, Risks, and Mitigation Strategies

2602.11567

Behavioral Indicators of Overreliance During Interaction with Conversational Language Models

2602.09286

Human Control Is the Anchor, Not the Answer: Early Divergence of Oversight in Agentic AI Communities

2602.09629

Stop Testing Attacks, Start Diagnosing Defenses: The Four-Checkpoint Framework Reveals Where LLM Safety Breaks

2602.09856

Synthesized using AI

Why Human Guidance Matters in Collaborative Vibe Coding

Earinter: A Closed-Loop System for Eating Pace Regulation with Just-in-Time Intervention Using Commodity Earbuds

Same Feedback, Different Source: How AI vs. Human Feedback Shapes Learner Engagement

Kissan-Dost: Bridging the Last Mile in Smallholder Precision Agriculture with Conversational IoT

LLM-Enhanced Wearables for Comprehensible Health Guidance in LMICs

Same Feedback, Different Source: How AI vs. Human Feedback Shapes Learner Engagement

T2VTree

Kissan-Dost

Guardian Angel

PPGTransID

Earinter

BabyMamba-HAR

ECHO

LCIP

Large Language Models in Peer-Run Community Behavioral Health Services: Understanding Peer Specialists and Service Users' Perspectives on Opportunities, Risks, and Mitigation Strategies

TouchScribe: Augmenting Non-Visual Hand-Object Interactions with Automated Live Visual Descriptions

Behavioral Indicators of Overreliance During Interaction with Conversational Language Models

Disability-First AI Dataset Annotation: Co-designing Stuttered Speech Annotation Guidelines with People Who Stutter

Don't blame me: How Intelligent Support Affects Moral Responsibility in Human Oversight

Human-AI Synergy Supports Collective Creative Search

Stop Testing Attacks, Start Diagnosing Defenses: The Four-Checkpoint Framework Reveals Where LLM Safety Breaks

Enhancing Generative AI Image Refinement with Scribbles and Annotations: A Comparative Study of Multimodal Prompts

Users want comfort, systems need friction

How was this newsletter synthesized?

Methodology

Selection Criteria

Key Themes Discovered

Field Report: ai-interaction

Trust, Control, and Collaboration

Top Papers in this Theme

Large Language Models in Peer-Run Community Behavioral Health Services: Understanding Peer Specialists and Service Users' Perspectives on Opportunities, Risks, and Mitigation Strategies

Behavioral Indicators of Overreliance During Interaction with Conversational Language Models

Human Control Is the Anchor, Not the Answer: Early Divergence of Oversight in Agentic AI Communities

Stop Testing Attacks, Start Diagnosing Defenses: The Four-Checkpoint Framework Reveals Where LLM Safety Breaks

Code2World: A GUI World Model via Renderable Code Generation