Review Open access

Human–AI Collaboration in Automated X-Ray Screening: Effects of Alarm Types and Reliability Levels on Operator Performance in Subway Security

Xin Zhou

2026 · AHFE International · Vol 234 · 0 citations

TL;DR

An empirical mapping between alarm granularity and automation reliability is provided, demonstrating that the optimal alarm type depends on the AI’s operating performance.

Abstract

Public-transportation X-ray checkpoints increasingly integrate Automated Diagnostic Aid Systems (ADAS) to support threat detection, yet system-level success remains contingent on human operators’ vigilance, decision strategies, and calibrated trust in automation. To promote the joint performance of human–AI collaboration, this study examines how alarm design (i.e., the way AI provides diagnostic advice) should vary with automation reliability. We experimentally compared three alarm modalities—binary alarm (“danger/safe”), likelihood alarm (four-level graded advice: “danger/warning/possible-safe/safe”), and automated decision (the system hides “safe” images and forwards only “danger” cases for human review)—across three system reliability levels (70%, 80%, 90%). Twenty-one participants completed X-ray baggage search tasks with target prevalence set as 30%; after quality control, 18 datasets (n=6 per reliability) were analyzed. Primary objective measures were d′ sensitivity (Signal Detection Theory) and response time (RT) for target-present and target-absent decisions; subjective measures captured multi-dimensional trust.As automated decision triages images and alters the decision space, SDT analyses focused on binary vs. likelihood conditions, with alarm type as a within-group variable and reliability as a between-group variable. A two-way mixed ANOVA revealed a significant main effect of alarm type and a significant alarm-type × reliability interaction (Alarm Type: F=10.88, p<.05; Interaction: F=11.63, p<.05). For binary alarms, operator d′ increased monotonically with ADAS reliability (from 70% to 90%), indicating that categorical cues benefit from high classifier accuracy. For likelihood alarms, d′ improved from 70% to 80% but declined at 90%, suggesting that when the AI is highly accurate, graded, ambiguous messages can impose avoidable decisional complexity and cognitive load, degrading sensitivity relative to simpler cues. RT analyses did not yield reliable omnibus effects, though patterns were consistent with the interpretation that richer advice requires additional decisional processing, especially for target-absent judgments.Subjective results complemented the objective pattern. At 70% reliability, participants preferred likelihood alarms, rating them higher on perceived competence/faith/reliability, consistent with the idea that greater transparency and nuance are valuable when automation is imperfect. At 90% reliability, participants expressed the highest trust and willingness to rely on automated decision, reflecting comfort with delegating routine “safe” triage to a highly reliable AI and reserving human involvement for flagged “danger” cases. Across conditions, trust calibration tracked reliability, but critically depended on the alarm form through which the AI conveyed its assessment.Contributions. (1) We provide an empirical mapping between alarm granularity and automation reliability, demonstrating that the optimal alarm type depends on the AI’s operating performance. (2) We show that graded likelihood cues can enhance sensitivity at lower automation reliability by supporting informed human override, but they can reduce performance at high reliability by adding decisional friction. (3) We integrate SDT-based sensitivity with multi-dimensional trust to articulate actionable design guidance for human–AI teaming in safety-critical screening.Implications for design. To maximize human–AI system performance, alarm transparency should be matched to system reliability. Likelihood-based alarms are preferable when reliability is modest, as they support human verification and facilitate appropriate criterion setting. When reliability is high, binary or automated-decision modes are recommended to minimize cognitive load and enable efficient triage. Practically, an adaptive alarm policy that switches alarm type as real-time reliability estimates change may best sustain calibrated trust, operator efficiency, and system-level sensitivity in high-throughput subway screening.

DOI Semantic Scholar

Read PDF

Similar papers

Open access Jul 2026

Safety design guidelines for clinician–AI interaction in computer-aided diagnosis systems using system-theoretic framework with explainability validation

Ensuring the safety of Artificial Intelligence-enabled Computer-Aided Diagnosis systems is critical because diagnostic errors can have serious consequences for patient care. However, existing regulatory and risk management frameworks often do not sufficiently address the complex socio-technical interactions between clinicians and Artificial Intelligent systems, leaving key human-centered safety challenges underexplored. This paper presents a systematic and human-centered approach to deriving safety design guidelines for clinician-Artificial Intelligence interaction in Computer-Aided Diagnosis systems using System-Theoretic Process Analysis. Through this analysis, we identify critical hazards associated with clinician–Artificial Intelligence collaboration, including automation bias on system recommendations, and misinterpretation of explanations. Based on the identified unsafe control actions, we formulate a set of actionable and traceable safety design guidelines that promote transparency, coherent explanations, and calibrated trust in Artificial Intelligence-assisted decision-making. To bridge safety analysis and system design, the proposed guidelines are operationalized within a Computer-Aided Diagnosis interaction framework. The framework includes a safety-oriented Graphical User Interface that integrates multiple explanation methods and interactive mechanisms to promote clinician engagement. Furthermore, we introduce a safety-oriented evaluation approach that uses consistency across multiple explanation methods as a quantitative indicator of potentially unreliable or ambiguous explanations. By linking System-Theoretic Process Analysis, interaction design, and explainability evaluation, this work provides a unified and reusable framework for improving the safety and reliability of Artificial Intelligence-driven Computer-Aided Diagnosis systems.

Yuki Hagiwara, Katherine Fitch, Mario Trapp · Scientific Reports · 0 citations

Review Open access 2026

AI-Powered Computer Vision Systems for Automated Safety Compliance Monitoring and PPE Detection in Automobile Manufacturing Plants

Results indicate the vision model achieves 92% PPE detection accuracy and practical implications highlight the need for robust encryption user training and continuous model retraining, and future research should examine long-term behavioral impacts and cross-plant scalability.

A. Mahajan, K. Dharwadkar · International Journal of Inn... · 0 citations

Dependable Person Detection using AI in Industrial Environments

Suggestions for ensuring safe person detection using AI in industrial environments are offered, including suggestions for ensuring safe person detection using AI in industrial environments.

Iwo Kurzidem, Andrea Matic-Flierl, Poulami Sinhamahapatra et al. · 0 citations

Open access Aug 2026

Uncovering reliability blind spots in transportation AI: an edge-case-based validation approach for transportation systems

With artificial intelligence (AI) technology expanding into high-risk domains, criteria for evaluating its reliability and safety are required, as AI deployment may otherwise be constrained by future government regulation or societal backlash. In particular, verifying reliability in unexpected scenarios, such as edge cases, has emerged as a key challenge for implementing safe AI systems. This study introduces a supplementary validation approach to support the reliability testing of AI systems in high-risk domains. Considering the case of AI-based vehicle license plate recognition (VLPR), this study proposes edge-case-based validation approaches specifically designed for certification bodies. These approaches aim to address reliability validation needs that fall outside the well-defined operating regions covered by conventional testing. For VLPR systems, edge cases were identified as adverse-condition samples misrecognized by a baseline model and were then evaluated using a resolution-enhanced model. Under rainy conditions, the conventional recall metric reported 99.2%, whereas evaluating the same system on the extracted edge cases yielded 93.3%—a gap the aggregate metric did not make explicit. These findings highlight the potential for certification bodies to utilize this approach as a qualitative validation tool to uncover reliability blind spots in such systems. While the study focuses on VLPR systems, the proposed methodology is, in principle, transferable to other AI-based applications, although this remains to be tested. Future studies will aim to quantify edge case extraction techniques and examine their adaptability to diverse AI systems across various domains. This study provides preliminary, case-study insights that may contribute to improving the safety and reliability of AI systems in high-risk environments.

June Kim, Donghyun Lee · Scientific Reports · 0 citations

Preprint Jul 2026

Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels

Jailbreak-robustness research typically evaluates safety through generated responses using an LLM-as-judge approach. Such evaluations, however, are sensitive to the benchmark's grading procedure and capture only observed behavior on a given set of attacks, without directly revealing the hidden fragility of the underlying safety mechanisms. This work proposes JADR (Jacobian Assessment of Danger Recognition), a protocol that measures a model's internal representation through Jacobian space (J-space, a recently proposed workspace of verbalizable concepts) before the first response token is generated. For every prompt and layer we record the top-k J-space tokens; these are grouped into six behavioral scenario axes and compared between a danger sample based on StrongREJECT and a safe control drawn from XSTest and OKTest. The method does not call on an external judge model: the computation runs entirely locally, on the activations of the model under evaluation, which lets us compare both different models against each other and modifications of a single model - quantization and fine-tuning in particular - on the same terms. The final comparison rests on the proposed SafetyAUC metric, complemented with bootstrap confidence intervals. The protocol is applied to six models (Qwen3-1.7B, Qwen3-4B, Qwen3-8B, Qwen3-Uncensored-4B, Qwen3-SafeRL-4B, Gemma 2 9B) across three weight-representation regimes - BF16, INT8, and INT4 - and checked against an independent behavioral evaluation with the StrongREJECT grader. The metric separates models with a strong versus a weak internal safety mechanism with statistical significance and captures substantively different effects across quantization regimes.

Roman Prosvirnin, V. Minchenkov, A. Soldatov et al. · 1 citation

Conference Aug 2026

AI-Driven Well Surveillance Using Low-Code Tools: Automating Anomaly Detection and Actionable Insights on Gas Lift Wells

Automation, AI, and modern monitoring equipment has made the dream of surveillance staff having a near-omniscient view of oilfield issues closer to reality. This paper tackles a segment of this vision and presents a low-code well surveillance workflow developed to reduce the manual monitoring by well performance specialists and production engineers while extending surveillance coverage across gas lift wells. The proof-of-concept system was built by Occidental using Microsoft Power Automate, Power Apps, SharePoint, OneDrive, Azure OCR, and multimodal large language model (LLM) calls. The workflow captures well images, extracts structured information, prompts an LLM to identify abnormal operating conditions and propose follow-up actions, and then presents those outputs in a user-facing application alongside the original surveillance view. The prototype established a complete end-to-end path from data capture to anomaly review and user feedback. In testing, the system demonstrated that a low-code architecture can support daily surveillance screening, highlight wells that merit attention, and place draft recommendations directly in front of users. This work is not presented as a final enterprise architecture. Instead, it documents the decisions, tradeoffs, performance metrics comparing different LLMs, and lessons from an accelerated proof of concept effort. The results show that low-code tools can be used to create a practical, AI-assisted surveillance workflow, create a testing ground for multimodal AI in production operations, and provide a bridge toward future model-driven surveillance systems that rely on structured data and purpose-built machine learning models.

Ivan Tanakov, Emmanuel Zoubovsky, Ivan Berry et al. · Americas · 0 citations