Looks Fine to Me: Understanding Oversight in Agentic Interactions

Dinithi Dissanayake; Yasith Samaradivakara; Hyung Woon Lee; Yadeesha Weerasinghe; Sanjay Saha; Pattie Maes; Suranga Nanayakkara

arXiv · 2610.27801

Looks Fine to Me: Understanding Oversight in Agentic Interactions

Dinithi Dissanayake·Yasith Samaradivakara·Hyung Woon Lee·Yadeesha Weerasinghe·Sanjay Saha·Pattie Maes·Suranga Nanayakkara

View original source ↗Read source document ↗Export metadata ↓

Abstract

Agentic AI systems are deployed to execute complex tasks across domains such as writing, coding, and data analysis. As execution is offloaded to AI agents, the human role shifts toward oversight. Effective oversight requires recognizing when closer scrutiny is needed, even though flawed outputs may appear convincing. Our formative study (N=12) revealed that users accepted plausible outputs without thorough evaluation, and only recognized these lapses in retrospect. To examine this systematically, we conducted a controlled within-subject study (N=36) in which participants completed programming tasks with an agent that produced correct outputs, silent defects, or visible failures. Combining task outcomes, interaction logs, gaze, physiological signals, and self-reports, we examined perceived, behavioral, and effective oversight. Participants were often highly confident while missing silent defects: they performed generic checks, failed to generate diagnostic evidence, or even misinterpreted defect evidence. We discuss how studying oversight across multiple dimensions can support future work on timely, task-calibrated re-engagement.

Explore related subjects

cs.HC

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Dinithi Dissanayake, Yasith Samaradivakara, Hyung Woon Lee, Yadeesha Weerasinghe, Sanjay Saha, Pattie Maes, Suranga Nanayakkara. 2026-10-20. Looks Fine to Me: Understanding Oversight in Agentic Interactions. https://arxiv.org/abs/2610.27801

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Unpacking "Personal" Health Informatics for Proactive Collective Care

Care is primarily a collective phenomenon, with a practice that involves sharing health and wellbeing information within a trusted "care circle" of family and companions for sensemaking, interpretation, decision-making, and follow-through. However, current digital health tools and systems are designed for individuals and primarily intended for Personal Health Informatics (PHI). This mismatch between collective practice and individualistic design creates new challenges for the proactive use of such systems in care settings and limits adoption, sustained engagement, and meaningful use. To examine how people practice collective care and how (if) they perceive, adopt, and integrate PHI systems for proactive care, we conducted a sequential mixed-methods study. Through an initial survey (n=87) and semi-structured interviews (n=22), we found that their practices involve collectively analyzing and sensemaking health information. However, we also found that their use of existing systems to support such practices is constrained by factors at personal, relational, technological, and structural levels that evolve over time. To address these constraints and explore redesigning PHI toward "Collective Health Informatics", we conducted eight stakeholder interviews (n=12), a follow-up survey (n=114), and six co-design workshops (n=12) to understand the dynamics required in collective settings while retaining agency. Using a design probe evaluation (n=38), we refine a design vision for coordinated, trustworthy action across such care relationships. Our findings motivate CC-Proact, an operational map that translates ecological influences into three design levers: Agency, Elicitation, and Engagement. Using this map, our work empirically examines collective care practices and offers six design recommendations to build responsible systems that proactively support collective care.

cs.HC↗

arXiv2026

When Language Learners Code-Switch: Korean-English Code-Switching in LLM-based Speaking Practice

Code-switching (CSW), the use of multiple languages within the same interaction, has been widely studied in NLP for multilingual communication. While CSW also occurs in language-learning contexts, the functions and pedagogical implications of learner CSW in learner-AI interactions remain underexplored. To address this gap, we introduce LAKE, a dataset of Learner-AI dialogues with Korean-English CSW. This dataset captures South Korean learners' CSW behavior during English speaking practice with LLM-based conversation partners, comprising 675 dialogues, 19.6K utterances, and 43.9 hours of learner speech. We develop a taxonomy of learner CSW functions, finding that learners switch to Korean for diverse purposes, including signaling lexical gaps, negotiating meaning, and conveying culturally specific content. We also observe learners reusing and adapting English expressions introduced by the LLM partner following their code-switches, suggesting opportunities for contextualized lexical support. To support research on speech recognition for learner CSW, we release 4.9 hours of code-switched learner speech with human-verified transcripts and benchmark speech-to-text models on Automatic Speech Recognition (ASR), revealing difficulties in recognizing learner CSW. These findings highlight learner CSW as a communicative and pedagogically meaningful resource that can inform the design of AI-based conversational language learning systems.

cs.HC↗

arXiv2026

Evaluation format, not only model capability, shapes measured triage failure in consumer health AI

A recent Nature Medicine study reported that ChatGPT Health under-triages 51.6% of emergencies and concluded that consumer-facing AI triage poses safety risks. Its protocol, however, was an exam-style scaffold (forced A/B/C/D output, knowledge suppression, no clarifying questions) unlike how consumers use health chatbots. We ask how much of the headline error rate is a property of the measurement rather than of the models. In a first, mechanistic study, five frontier LLMs on a 17-scenario bank scored 6.4 points higher under naturalistic patient-style messages than under the constrained scaffold (p=0.015). In a second, faithful replication we ran the authors' own 60 released vignettes through six frontier models under four matched formats, with clinician validation of the rewrites and a blinded clinician audit of the LLM adjudicators. Here the direction reversed: free-text rewrites scored slightly below the exact prompt (78.7% vs 81.8%, p=0.020), removing only the answer scaffold changed little (80.6% vs 81.4%, p=0.81), and naturalistic input with a forced categorical answer beat both the exact prompt (84.4%, p=0.023) and free text (p=0.0001). On the four vignettes defining the original emergency rate, under-triage was 17% with the exact scaffold, 50% in free text, and 21% with a forced letter on the same message; every free-text "under-triage" was a same-day recommendation scored C rather than D, and 60% made escalation conditional on information the patient was asked to check, behavior a single-turn benchmark cannot score. The headline rate did not reproduce as a stable cross-model finding; much of it depends on output format and on mapping prose onto a four-point scale. Each format elicits a separate response, so the design measures format sensitivity, not format versus capability. Safety claims should report sensitivity to input wording, output format and adjudication.

cs.HC↗