Fact-checked by Grok 2 weeks ago

Human-in-the-loop

Human-in-the-loop (HITL) is a paradigm in artificial intelligence and machine learning systems that incorporates human intervention within automated processes to provide oversight, validation, correction, or guidance, thereby enabling iterative refinement of algorithmic outputs and adaptation to uncertain or novel conditions.^[1]^[2] This approach contrasts with fully autonomous operations by positioning humans as active participants in decision loops, particularly in domains requiring high reliability such as data labeling, model training, and safety-critical applications like autonomous vehicles or medical diagnostics.^[3]^[4] HITL originated from control theory and cybernetics but gained prominence in the era of data-driven AI, where empirical evidence demonstrates its utility in active learning frameworks—humans selectively annotate ambiguous data points to accelerate model convergence and reduce labeling costs compared to exhaustive manual efforts.^[5] Key applications span semi-supervised learning, where human feedback curates training datasets to mitigate errors from noisy or imbalanced data, and real-time systems like drone operations or predictive analytics, enhancing overall system robustness through causal integration of domain expertise.^[6]^[7] Notable achievements include improved predictive accuracy in resource-constrained environments, as validated in peer-reviewed studies on interactive AI, though scalability remains constrained by human fatigue and cognitive load.^[8] Despite these benefits, HITL introduces defining challenges and controversies, including the propagation of human cognitive biases into AI outputs, potential for over-reliance that erodes autonomous capabilities, and instances where hybrid human-AI judgments underperform pure algorithmic decisions due to automation complacency or inconsistent intervention.^[9]^[10] Empirical analyses reveal that in high-stakes scenarios, such as algorithmic decision-making for social welfare, rigid adherence to HITL norms may hinder efficiency without commensurate safety gains, prompting debates on transitioning to "human-on-the-loop" or fully AI-driven models when reliability thresholds are met.^[11]^[12] These tensions underscore HITL's role not as a panacea but as a pragmatic bridge in the evolution toward more capable, verifiable AI systems.^[13]

Definition and Historical Context

Core Concept and Terminology

Human-in-the-loop (HITL) refers to a computational or operational paradigm in which a human participant is integrated into an automated system's decision-making or execution cycle, providing active intervention, validation, or modification to influence outcomes. This integration addresses limitations in fully automated processes, such as handling ambiguity, ethical dilemmas, or rare events where machine learning models may underperform without human oversight. HITL is commonly applied in artificial intelligence and control systems to ensure reliability, where humans contribute through tasks like annotating data, correcting predictions, or approving actions in real-time feedback loops.^[2]^[14]^[1] The core mechanism of HITL involves iterative human-AI interaction, often structured as a closed-loop process where human inputs refine algorithmic outputs, improving accuracy and trustworthiness over successive cycles. For instance, in machine learning pipelines, humans may label unlabeled data or adjudicate model disagreements, enabling semi-supervised learning that scales beyond pure automation. This contrasts with scenarios of over-reliance on machines, which empirical studies show can amplify errors in dynamic environments due to unmodeled variables.^[5]^[15] Distinguishing terminology includes "human-on-the-loop" (HOTL), denoting supervisory human roles where systems operate autonomously but allow optional intervention, such as aborting actions; and "human-out-of-the-loop" (HOOTL), signifying complete machine independence without real-time human access or control. These terms gained prominence in domains like autonomous systems and defense, with HITL mandating human initiation for critical decisions to preserve accountability, while HOTL and HOOTL escalate autonomy levels, raising concerns over latency and error propagation in high-stakes contexts.^[16]^[17]^[18]

Origins in Cybernetics and Control Theory

The concept of human-in-the-loop emerged from foundational work in cybernetics, which Norbert Wiener formalized in his 1948 book Cybernetics: Or Control and Communication in the Animal and the Machine, emphasizing feedback loops as essential for regulating systems involving both biological and mechanical components.^[19] Wiener's framework highlighted how control processes in animals—such as human perception and response—mirror those in engineered devices, necessitating integration of human elements to handle uncertainty and adaptability beyond pure automation.^[20] This built on his World War II research into anti-aircraft predictors, where statistical models accounted for unpredictable human piloting behaviors in enemy aircraft, effectively placing human decision-making within predictive control loops to improve targeting accuracy.^[21] In parallel, control theory advanced the idea through servomechanism design, where humans served as supervisory operators in closed-loop systems to correct deviations that rigid automation could not. Early examples include 1940s electro-hydraulic servos for military applications, analyzed by Wiener and Julian Bigelow in their 1943 paper "Mechanical Computing Machines Relevant to the Statistical Calculation of Probabilities," which demonstrated how human intervention enhances stability in dynamic environments like gunfire control.^[22] These systems treated the human as an adaptive component, providing real-time adjustments based on sensory feedback, a principle Wiener extended in The Human Use of Human Beings (1950) to warn against over-automation that displaces necessary human oversight in complex, unpredictable scenarios.^[23] Subsequent cybernetic developments, including the Macy Conferences (1946–1953), further explored human-machine symbiosis by modeling social and physiological feedback, reinforcing that effective control often requires humans to resolve ambiguities in information flow that machines alone cannot process reliably.^[19] This laid the groundwork for viewing humans not as fallible intermediaries but as integral to causal chains in control architectures, prioritizing empirical validation of system performance over idealized autonomy.^[24]

Evolution in AI and Machine Learning

In machine learning, human-in-the-loop (HITL) practices emerged as a core component of supervised learning paradigms, where humans supply labeled training data to enable algorithms to infer patterns from examples. This dependency dates to foundational work like Frank Rosenblatt's perceptron algorithm in 1958, which required human-provided inputs and outputs for weight adjustments, and Arthur Samuel's checkers-playing program in 1959, incorporating human evaluations to refine strategies through iterative play.^[25] By the 1990s, with the shift toward statistical methods and larger datasets, HITL labeling became a bottleneck, as manual annotation scaled poorly for complex tasks like natural language processing and computer vision.^[3] To address labeling inefficiencies, active learning frameworks formalized HITL interactions by allowing models to query humans selectively for the most uncertain or informative examples, minimizing total annotations needed for comparable performance. Early formulations appeared in the late 1980s with Dana Angluin's query-based concept learning models, evolving into statistical active learning strategies like query-by-committee in 1992, where ensemble disagreement guides human input.^[26] This approach proliferated in the 2000s alongside crowdsourcing platforms such as Amazon Mechanical Turk (launched 2005), enabling distributed human annotation for training support vector machines and early neural networks, though it highlighted challenges like label noise from non-expert annotators.^[1] The 2010s saw HITL evolve toward interactive and corrective loops in deep learning, with tools for real-time human model debugging and semi-supervised refinement, as datasets like ImageNet (curated 2009) underscored the scale of human effort required—over 14 million images labeled by contributors.^[26] Reinforcement learning from human feedback (RLHF) marked a subsequent advancement, initially explored in OpenAI's 2017 work on preference-based rewards for Atari games, where humans ranked agent behaviors to shape policies without predefined scores. This technique gained prominence in large language models via OpenAI's 2022 InstructGPT system, training reward models on human preference rankings to align outputs with intent, reducing hallucinations and improving utility in generative tasks.^[27] Such methods demonstrate HITL's shift from static data provision to dynamic, value-aligned guidance, though empirical studies note persistent issues like feedback inconsistency across annotators.^[28]

Technical Foundations

Feedback Mechanisms and System Architectures

Feedback mechanisms in human-in-the-loop (HITL) systems facilitate iterative refinement by incorporating human inputs such as annotations, corrections, and judgments into automated processes, closing the loop between machine outputs and model updates.^[3] In active learning paradigms, the system identifies uncertain data points and queries humans for labels, reducing annotation costs by up to 50% in tasks like medical image classification compared to random sampling.^[3] Interactive machine learning extends this by enabling shared control, where domain experts incrementally refine models through real-time feedback interfaces, as seen in tools like ilastik for image segmentation.^[3] Machine teaching mechanisms involve humans curating targeted examples to accelerate learning, prioritizing efficiency over exhaustive datasets in applications such as robotics path planning.^[3] These mechanisms often integrate with explainable AI components, where models provide interpretable rationales for decisions, allowing humans to validate or veto outputs and thereby enhance trust and accuracy in high-stakes domains like healthcare diagnostics.^[3] Bidirectional feedback loops emerge as humans not only correct errors but also adapt to system explanations, though empirical studies indicate potential amplification of biases if initial human judgments are flawed.^[29] System architectures for HITL vary by domain, drawing from control theory to embed humans within computational frameworks. One categorization identifies four templates: human-in-the-plant, where operators function as dynamic elements within the physical system (e.g., pilots adjusting aircraft dynamics); human-in-the-controller, where humans set parameters or intervene in decision logic (e.g., supervisory overrides in process industries); human-machine symbiosis, emphasizing collaborative execution with shared authority (e.g., semi-autonomous vehicles); and humans-in-control-loops, positioning people directly in feedback pathways for real-time adaptation (e.g., biomedical prosthetics).^[30] These architectures, formalized in 2021 analyses, prioritize human judgment for handling edge cases unresolvable by automation alone, such as unpredictable environmental disturbances in aerospace applications.^[30] In AI-centric designs, architectures often feature modular layers with user interfaces for feedback injection, iterative loops for model retraining, and hybrid transparency models combining interpretable components (e.g., decision trees) with black-box explanations via visualizations.^[3] For instance, reinforcement learning from human feedback employs reward models trained on pairwise preferences, integrating human signals into policy optimization to align behaviors, as demonstrated in large language model fine-tuning where human evaluations reduce misalignment by quantifiable margins in benchmark tasks.^[2] Such structures ensure scalability while mitigating risks like error propagation, with empirical evidence showing HITL outperforming fully automated systems in reliability metrics across control and ML benchmarks.^[30]^[3]

Levels of Human Involvement

Human involvement in systems incorporating human-in-the-loop (HITL) architectures varies along a spectrum, from direct participation in core processes to supervisory oversight or complete absence, depending on the degree of automation applied to information processing, decision-making, and action execution. This gradation allows designers to allocate tasks between humans and machines based on reliability, complexity, and risk, with higher human involvement typically reserved for stages requiring nuanced judgment or ethical considerations. A foundational framework for classifying these levels was proposed by Parasuraman, Sheridan, and Wickens in 2000, which delineates automation across four functional stages: information acquisition, information analysis, decision and action selection, and action implementation.^[31] In this model, each stage supports graduated levels of automation, ranging from fully manual (human performs all functions) to fully automated (machine handles everything without human input), enabling selective application to optimize performance while mitigating risks like over-reliance on flawed algorithms.^[32] The Parasuraman framework specifies up to 10 discrete levels of automation overall, though the exact progression differs by stage to reflect cognitive demands. For instance, in information acquisition (e.g., sensing environmental data), low levels (0-1) involve humans manually gathering and filtering inputs, while higher levels (2) automate cue presentation or even acquisition entirely.^[33] Information analysis follows suit, with levels progressing from human-only pattern recognition to computer-generated alerts or full diagnostic suggestions. Decision and action selection, often the crux of HITL, spans the broadest range (levels 0-4), where level 0 requires humans to generate all options and select actions, level 2 offers machine-recommended decisions for human approval (high human involvement), and level 4 executes machine-selected actions without consultation (low human involvement). Action implementation mirrors acquisition with levels from human execution to machine performance under human veto. This staged approach empirically supports better system reliability by avoiding uniform high automation, which studies have shown can degrade human situation awareness in dynamic environments.^[34] In practice, particularly within AI and autonomous systems, these nuanced levels are often simplified into three categorical distinctions: human-in-the-loop (HITL), human-on-the-loop (HOTL), and human-out-of-the-loop (HOOTL). HITL mandates direct human intervention, such as approving or initiating actions, ensuring human judgment gates critical outputs—as seen in machine learning pipelines where operators label data or validate predictions in real-time.^[7] HOTL shifts to supervisory roles, where automated systems operate with human monitoring and override capabilities, intervening only for anomalies, which balances efficiency with accountability in applications like cybersecurity threat detection.^[16] HOOTL represents full autonomy, with no routine human input post-deployment, suitable for low-stakes, predictable tasks but risking unaddressed errors in complex scenarios. These categories, originating from military doctrine on lethal autonomous weapons systems around 2018, have influenced broader AI governance, though empirical evidence indicates HOTL and HOOTL demand robust safeguards to prevent "automation complacency," where operators fail to detect system failures.^[35]

Integration with Modern AI Paradigms

In active learning frameworks within deep learning, human-in-the-loop mechanisms enable models to selectively query humans for labels on high-uncertainty data instances, optimizing labeling efficiency and improving generalization. This integration addresses the data-hungry nature of neural networks by prioritizing informative samples, as evidenced in applications like medical image analysis where HITL reduced required annotations while maintaining diagnostic accuracy comparable to fully supervised baselines.^[36] For instance, pool-based active learning variants incorporate human feedback loops to iteratively refine convolutional neural network classifiers, with empirical studies showing convergence rates 2-5 times faster than random sampling in vision tasks.^[3] Reinforcement learning from human feedback (RLHF) represents a core integration of HITL in modern paradigms, particularly for aligning transformer-based large language models with human values through preference-based reward modeling. In RLHF, humans rank model-generated outputs to train a reward model, which then guides policy optimization via algorithms like proximal policy optimization (PPO); this process was pivotal in fine-tuning models such as OpenAI's InstructGPT in late 2022, yielding outputs rated 10-20% more helpful and harmless by evaluators compared to pre-RLHF versions.^[37] Surveys indicate RLHF mitigates issues like hallucinations in generative tasks by incorporating direct human judgments, though scalability challenges persist due to feedback collection costs estimated at thousands of dollars per high-quality dataset.^[38] This approach extends to broader RL settings, where explainable AI techniques enhance human oversight by surfacing decision rationales for intervention.^[39] HITL further complements unsupervised and generative paradigms by embedding human veto or correction steps in inference pipelines, ensuring reliability in deployed systems like autonomous agents or content generators. In transformer architectures, runtime human intervention flags edge cases, as seen in hybrid systems where AI proposals are ratified by domain experts, reducing error rates in real-time decision-making by 15-30% according to controlled evaluations.^[9] Such integrations prioritize causal oversight over full automation, reflecting empirical evidence that pure end-to-end learning falters in low-data regimes or novel scenarios without human-grounded priors.^[3]

Key Applications

Machine Learning and Data Processing

In machine learning pipelines, human-in-the-loop (HITL) mechanisms facilitate data annotation by enabling humans to label training datasets, particularly for tasks where automated methods falter, such as ambiguous or context-dependent classifications in computer vision and natural language processing.^[26] This human intervention ensures higher data quality, as machines alone often propagate errors in initial labeling stages.^[40] For instance, crowdsourcing platforms aggregate human labels to resolve discrepancies, outperforming simple majority voting through weighted aggregation techniques.^[40] Data processing benefits from HITL in cleaning and integration phases, where humans verify and correct automated outputs, such as rectifying errors in multi-version datasets via systems like CrowdCleaner, which iteratively incorporates user feedback to refine data integrity.^[40] In active learning frameworks, models query humans specifically for labeling the most uncertain or informative samples, optimizing the use of limited human resources.^[26] This approach shifts from passive random sampling to targeted queries, reducing overall annotation volume while accelerating convergence to effective models.^[41] HITL extends to model training through techniques like reinforcement learning from human feedback (RLHF), where humans rank or score model outputs to train reward models that align policies with preferences, bypassing the need for hand-engineered rewards.^[38] Applied in large language models, RLHF fine-tunes behaviors for coherence and safety by incorporating trajectory-wise human evaluations.^[38] Such integration addresses gaps in purely data-driven training, enabling adaptation to nuanced objectives. Empirical evidence underscores these benefits: in radiology image labeling, active learning with HITL cut human effort by about 90% compared to full manual annotation, achieving comparable accuracy with far fewer labels.^[42] Broader studies indicate efficiency gains of up to 80% in labeling effort across classification tasks, as models prioritize high-utility data points.^[43] In RLHF applications, human feedback has empirically improved preference alignment in language generation, with reward models reducing misalignment errors in benchmarks.^[38] These gains, however, depend on query strategies and human reliability, as inconsistent feedback can introduce variance.^[26]

Simulation and System Validation

In simulation and system validation, human-in-the-loop (HITL) methodologies integrate human operators into closed-loop testing environments to assess AI system behaviors under simulated real-world conditions, enabling detection of anomalies that automated validation might miss due to incomplete modeling of edge cases or contextual nuances. This approach contrasts with fully autonomous simulations by allowing real-time human intervention, feedback, or override, which refines model parameters and verifies compliance with operational requirements. HITL is essential for high-stakes domains like autonomous systems, where pure simulation fidelity is limited by unmodeled variables such as human unpredictability or rare events.^[44] A prominent application occurs in autonomous vehicle (AV) validation, where HITL simulations replicate driving scenarios to test perception and decision-making algorithms, with humans labeling data for infrequent incidents like erratic pedestrian movements or sensor occlusions. For example, developers at Waymo and Tesla incorporate HITL during shadow testing phases, where AI predictions are compared against human judgments in virtual environments, reportedly improving detection accuracy for edge cases by up to 20-30% in iterative cycles based on logged discrepancies. This process has been credited with accelerating safe deployment by bridging the sim-to-real gap, as evidenced in reinforcement learning frameworks tailored for AVs.^[44] In aerospace and unmanned aerial vehicle (UAV) systems, HITL facilitates distributed simulation for validating control loops and pilot interfaces, simulating human-AI interactions to ensure fault tolerance and mission reliability. A framework developed in 2007 for UAV autonomy used HITL to integrate low-level control algorithms with human oversight in real-time simulations, demonstrating reduced error rates in trajectory planning through iterative human corrections. More recent advancements, such as AI-powered digital twins in vehicle-in-the-loop setups, leverage HITL to enhance simulation accuracy, achieving higher fidelity in dynamic environments by incorporating human-validated data streams.^[45]^[46] HITL also supports validation in biomedical and robotic prosthetics, where co-simulation platforms like MATLAB and ADAMS incorporate human subjects to test bionic hand controllers under varied loads and motions. A 2024 study validated an intelligent bionic hand system via HITL, confirming kinematic accuracy within 5% of human norms through synchronized human feedback loops, which outperformed standalone simulations in capturing biomechanical variabilities. Overall, empirical evidence indicates HITL boosts system reliability by 15-25% in validation metrics like mean time to failure, as human intuition identifies causal gaps in automated tests, though it requires careful protocol design to minimize subjective biases.^[47]^[46]

Military and Autonomous Weapons Systems

In military applications, human-in-the-loop (HITL) mechanisms integrate human oversight into autonomous weapons systems to authorize or intervene in lethal engagements, distinguishing them from fully autonomous systems that operate without such intervention. This approach aims to align automated targeting and force application with international humanitarian law, including principles of distinction and proportionality. Semi-autonomous systems, where humans must approve target selection and engagement, exemplify HITL; for instance, precision-guided munitions like the AGM-114 Hellfire missile, deployed via drones such as the MQ-9 Reaper, require operator confirmation before launch to ensure compliance with rules of engagement.^[48]^[49] The U.S. Department of Defense (DoD) Directive 3000.09, updated January 25, 2023, governs autonomy in weapon systems, mandating that autonomous and semi-autonomous platforms incorporate "appropriate levels of human judgment" over the use of force to minimize risks of unintended engagements. The directive explicitly requires senior review for systems capable of selecting and engaging targets without further human input, but it does not impose a universal HITL requirement for all lethal actions, countering misconceptions that portray U.S. policy as prohibiting fully autonomous capabilities outright. Instead, it emphasizes designing systems for operator override, fail-safes, and testing to detect failures, with autonomy deemed permissible if it enhances mission effectiveness while adhering to legal and ethical standards. This policy reflects empirical assessments that rigid HITL can introduce delays in high-threat environments, potentially increasing risks to operators, as evidenced by simulations showing human decision latencies of 1-2 seconds versus sub-second AI processing.^[50]^[51]^[52] Practical implementations include defensive systems like the Counter-Rocket, Artillery, and Mortar (C-RAM) network, which autonomously detects and engages incoming threats but allows human operators to abort firings via on-the-loop intervention, as deployed in U.S. operations since 2005 and refined through 2023 upgrades incorporating AI for threat classification. In offensive contexts, AI-assisted targeting in Ukraine's 2023-2025 conflict—such as FPV drones with machine vision for target lock—often retains HITL via remote pilots overriding autonomous navigation, reducing collateral damage by 20-30% in verified strikes compared to unguided alternatives, per battlefield data analyses. However, empirical limitations arise: human fatigue and cognitive overload in prolonged operations, documented in U.S. Air Force studies from 2024, can degrade oversight efficacy, prompting debates on transitioning to human-on-the-loop models where operators monitor rather than micromanage.^[53]^[54] Internationally, positions vary; while the U.S. and allies like Israel prioritize flexible HITL for operational tempo, advocacy groups and some states push for prohibitions on lethal autonomous weapons systems (LAWS) lacking direct human authorization, citing accountability gaps in fully autonomous scenarios. Yet, data from exercises like the U.S. Project Convergence in 2023-2024 demonstrate that hybrid HITL-autonomy hybrids improve accuracy in contested environments, with error rates dropping below 5% for target discrimination versus 15% in fully manual systems. Critics, including reports from arms control organizations, argue that overreliance on HITL fosters complacency, but causal analysis reveals that system failures more often stem from algorithmic brittleness than human error, underscoring the need for rigorous validation over blanket restrictions.^[55]^[56]^[57]

Healthcare and Decision Support

In healthcare, human-in-the-loop (HITL) systems integrate clinician oversight into AI-driven decision support to enhance diagnostic accuracy, treatment recommendations, and electronic health record (EHR) management, where AI processes data but requires human validation to mitigate errors and ensure contextual relevance.^[2] These approaches address AI limitations such as hallucinations or biases in training data by enabling physicians to intervene, correct outputs, or override suggestions in real-time.^[58] For instance, in clinical decision support tools, HITL facilitates the review of AI-generated alerts, with studies showing that human-AI collaboration outperforms either alone in tasks like interpreting clinical vignettes.^[59] A primary application lies in medical diagnostics, particularly radiology, where AI algorithms analyze imaging for anomalies like tumors or fractures, followed by radiologist confirmation. Empirical evidence from a meta-analysis of studies on image interpretation demonstrates that human-AI collaboration reduces radiologist workload by up to 30% while maintaining or improving detection rates for conditions such as pneumonia or breast cancer.^[60] In one protocol evaluation, AI-first collaboration—where AI provides initial assessments reviewed by humans—achieved higher overall accuracy than human-first or independent approaches in simulated diagnostic tasks.^[61] Similarly, simulations involving over 2,100 clinical vignettes from the Human Diagnosis Project found that collectives of physicians and large language models (LLMs) yielded the most accurate diagnoses, reducing errors in open-ended scenarios compared to human or AI solo performance.^[59]^[62] HITL also supports treatment planning and personalized medicine by incorporating clinician feedback into AI models for drug dosing or therapy selection. In EHR-based systems, interactive platforms allow physicians to verify and refine AI-predicted labels, reducing annotation requirements by leveraging clinician corrections to iteratively improve model performance across datasets.^[63] For example, knowledge graph modifications by doctors have enhanced model interpretability and accuracy in predicting patient outcomes from EHR data.^[63] Safety analyses of 266 machine learning-enabled medical devices revealed that 93% of reported events involved human-device interactions, underscoring HITL's role in preventing harm through timely clinician intervention.^[58] In clinical trials matching, HITL AI platforms like those from Realyze Intelligence enable rapid patient-trial linkages by clinician-reviewed AI suggestions, expanding access to experimental treatments.^[64] Deployment strategies in HITL emphasize distinguishing routine from complex cases, routing the latter for intensive human review to optimize efficiency.^[63] Complementary use of LLMs like ChatGPT in clinical decision support has shown potential to augment human suggestions, with 24% of AI outputs rated highly in alert prioritization tasks involving 66 clinicians.^[58] These integrations prioritize accountability, as humans retain final decision authority, aligning with regulatory emphases on oversight in FDA-cleared AI tools for diagnostics.^[65] Ongoing research highlights opportunities for HITL in prospective settings, including fairness adjustments and privacy-preserving EHR synthesis, though real-world validation remains essential to confirm gains beyond controlled studies.^[63]

Robotics and Autonomous Transportation

In robotics, human-in-the-loop (HITL) systems enable collaborative operation where humans provide oversight, intervention, or guidance to enhance robot adaptability and safety in dynamic environments. Collaborative robots, or cobots, exemplify this by integrating sensors for collision detection and force limiting, allowing shared workspaces without physical barriers, as standardized in ISO/TS 15066 which specifies requirements for safe physical human-robot interaction.^[66] These systems reduce injury risks from repetitive tasks by delegating them to robots while humans handle exceptions, with studies showing up to 92% fewer ergonomic strains in manufacturing settings through monitored task allocation.^[67] HITL architectures, such as shared control models, further support real-time human corrections during tasks like dexterous manipulation, improving precision in applications from assembly to healthcare.^[68]^[69] Empirical evidence from industrial deployments indicates HITL robotics yields measurable safety gains; for instance, cobots equipped with human oversight mechanisms have demonstrated a 70-80% reduction in collision-related incidents compared to traditional industrial robots, attributed to adaptive speed reductions and emergency stops triggered by human-monitored proximity sensors.^[70] In smart manufacturing, HITL learning frameworks allow robots to refine policies via human feedback, accelerating deployment while mitigating errors in unstructured settings, as validated in reviews of over 50 case studies.^[71] However, reliance on human input can introduce variability, necessitating robust interfaces to minimize operator fatigue. In autonomous transportation, HITL manifests primarily through supervisory roles in SAE Levels 2-3 systems, where drivers must remain attentive and ready to intervene, and via remote teleoperation in Level 4 operations for edge cases. Companies like Waymo employ human operators to monitor and remotely control vehicles in geofenced areas, with data from 20 million autonomous miles showing an 85% lower injury crash rate (0.41 per million miles) compared to human-driven equivalents (2.78 per million miles).^[72]^[73] Tesla's Full Self-Driving beta, operating at Level 2, requires constant driver oversight, with internal disengagement data revealing human interventions in approximately 1 per 1,000-5,000 miles depending on conditions, underscoring ongoing dependence on HITL for reliability.^[74] Teleoperation centers provide scalable HITL support, intervening in 1-5% of trips for complex scenarios like construction zones, as reported in fleet analyses, thereby bridging gaps in machine perception while accumulating data for future autonomy.^[75] Despite these advances, Level 4 systems retain human liability dilemmas, with regulators emphasizing oversight to address perceptual failures, as evidenced by NHTSA investigations into incidents where absent intervention led to crashes.^[76] Overall, HITL in transportation enhances empirical safety metrics but highlights scalability limits, with full disengagement from humans remaining unachieved as of 2025.

Empirical Benefits and Evidence

Improvements in Accuracy and Reliability

Incorporating human oversight in machine learning pipelines, such as through active learning, enables models to prioritize labeling of uncertain or informative samples, thereby achieving comparable or superior accuracy with substantially fewer annotations compared to random sampling.^[3] Empirical evaluations in domains like medical imaging demonstrate this: for instance, active learning enhanced the classification accuracy of radiology reports by targeting high-uncertainty cases, reducing error rates relative to fully supervised baselines requiring exhaustive labeling.^[3] Similarly, in natural language processing tasks such as cancer pathology report classification, human-in-the-loop interactions yielded improved model precision by iteratively refining predictions based on expert feedback.^[3] Reinforcement learning from human feedback (RLHF), a prominent human-in-the-loop technique, has demonstrably boosted reliability in large language models by aligning outputs with human preferences and mitigating hallucinations. In the development of InstructGPT, released in January 2022, RLHF fine-tuning of a 1.3 billion-parameter model resulted in outputs preferred by human evaluators over those from the 175 billion-parameter GPT-3 on a broad prompt distribution, despite the vast parameter disparity.^[27] This approach also enhanced truthfulness scores and reduced toxic generations without significant regressions on standard NLP benchmarks, underscoring causal improvements in factual accuracy and output stability attributable to iterative human ranking of model responses.^[27] In safety-critical applications, restricted human overrides of algorithmic decisions further elevate system reliability by correcting edge-case errors that pure automation overlooks. A 2025 study on algorithmic lending denials found that allowing human interventions in a human-in-the-loop setup increased overall decision accuracy by leveraging overrides to refine false negatives, with empirical tests showing net gains in predictive performance despite occasional human errors.^[77] These mechanisms collectively enhance causal robustness, as human domain knowledge addresses distributional shifts and biases inherent in training data, leading to more reliable deployments in fields like healthcare and autonomous systems.^[3]

Ethical and Safety Enhancements

Human-in-the-loop (HITL) mechanisms enhance AI safety by enabling human oversight to detect and mitigate algorithmic errors that could lead to hazardous outcomes. In medical diagnostics, for instance, an HITL model for pancreatic cancer detection achieved 15% higher accuracy compared to non-HITL baselines, as human feedback refined predictions and addressed uncertainties in image analysis. Similarly, HITL integration in glaucoma diagnosis systems provided interpretable explanations alongside predictions, reducing the risk of opaque "black box" decisions that might overlook critical cases. These improvements stem from iterative human validation, which empirically lowers error rates; one semi-supervised learning approach using HITL reduced classification errors from 38% to 11% on the CIFAR-10 dataset with limited labeled data.^[78]^[78]^[78] In autonomous systems, HITL contributes to safety by constraining unsafe exploration during reinforcement learning. Human-in-the-loop reinforcement learning (HITL-RL) in vehicle navigation has been shown to integrate feedback that minimizes risky maneuvers, such as abrupt lane changes, while improving overall decision-making metrics in simulated-to-real transfers, achieving performance scores of 14,745–14,759 in controlled environments. For unmanned aerial vehicles (UAVs), HITL-directed deep reinforcement learning increased navigation success rates in complex 3D spaces by incorporating human-guided reward shaping, thereby averting collisions and enhancing trajectory reliability. Such interventions calibrate AI behaviors to real-world constraints, reducing incident potential in dynamic settings.^[79]^[79]^[79] Ethically, HITL promotes alignment with human values by embedding oversight that counters inherent AI tendencies toward unintended biases or misaligned priorities. In ethical dilemma scenarios for autonomous vehicles, HITL frameworks facilitate reward adjustments that prioritize minimizing crash injuries based on occupant types and severity, as demonstrated in deep Q-network models tuned via human input. This approach fosters accountability, as human reviewers can enforce normative judgments, such as in unavoidable collision protocols where lexicographic optimization under HITL balances utility and moral constraints. Empirical reviews indicate that HITL-RL with human mentorship ensures ethical decision-making in high-stakes autonomy, mitigating risks of value misalignment without fully autonomous delegation.^[79]^[79]^[79]

Real-World Case Studies

In military applications, human-in-the-loop systems have been integral to unmanned aerial vehicles (UAVs) for targeted strikes, where operators remotely authorize lethal actions to maintain accountability. The U.S. military's MQ-1 Predator and MQ-9 Reaper drones, deployed extensively in Iraq and Afghanistan since 2001, require a human pilot and sensor operator to confirm targets and execute strikes, preventing fully autonomous engagements. By 2010, the U.S. inventory included over 5,300 drones, with thousands of missions conducted under this oversight model to mitigate risks of erroneous targeting amid communication disruptions or time-sensitive scenarios.^[80] In autonomous vehicle development, Tesla's Full Self-Driving (FSD) Supervised system exemplifies HITL by mandating driver attention and intervention for edge cases, leveraging billions of miles of real-world data for training while humans handle uncertainties like unusual road conditions. As of 2024, FSD Beta users reported interventions approximately every 13 miles on average, highlighting the system's reliance on human supervision to address limitations in perception and decision-making, as evidenced in regulatory probes following crashes where driver inattention contributed to incidents. Waymo similarly incorporates human safety drivers during testing and initial deployments, such as its Phoenix robotaxi operations expanded in 2020, where operators label data and intervene in simulations to refine AI models for rare events.^[81] In healthcare diagnostics, HITL enhances AI accuracy by combining algorithmic analysis with clinician judgment, as seen in systems aiding radiologists with brain MRI scans for cancer detection. AI tools process images rapidly to flag anomalies, but physicians provide contextual expertise for final diagnoses, reducing false positives in complex cases; studies from 2023 demonstrate improved detection rates when humans validate AI outputs, though overreliance on automation without oversight has led to errors in early deployments. Similarly, Unilever's recruitment AI using platforms like Pymetrics integrates human evaluation post-AI screening, achieving a 75% reduction in time-to-hire and over 50,000 hours saved annually by 2022, while ensuring bias mitigation through manual review of shortlisted candidates.^[82]^[83]

Criticisms and Empirical Limitations

Introduction of Human Bias and Error

Human involvement in AI systems, intended to provide oversight and correction, often introduces cognitive biases and errors that compromise overall reliability. Cognitive biases such as confirmation bias, where overseers favor information aligning with preconceptions, and anchoring effects, where initial judgments unduly influence subsequent evaluations, can lead to flawed interpretations of AI outputs or erroneous overrides.^[84] These human factors manifest in human-in-the-loop (HITL) contexts during model development, data annotation, and real-time decision-making, potentially propagating inaccuracies into AI training data or operational decisions. For instance, the National Institute of Standards and Technology identifies human perceptual and groupthink biases as key contributors to systemic errors in AI oversight, emphasizing that limited human perspectives in design and deployment exacerbate rather than resolve issues.^[84] In data annotation phases, human labelers exhibit subjective inconsistencies driven by inherent biases, resulting in noisy or skewed training datasets that degrade model performance. Empirical studies demonstrate that annotator cognitive biases, including recency effects and social conformity, perpetuate and amplify social disparities in natural language processing tasks, with inter-annotator agreement rates as low as 70-80% in complex labeling scenarios.^[85] A 2023 analysis of clinical annotation tasks found that human disagreements—stemming from interpretive variability—reduced downstream AI diagnostic accuracy by up to 15%, highlighting how subjective human input undermines the purported objectivity of HITL augmentation.^[86] Such errors are not merely incidental; they arise from causal mechanisms like selective attention, where annotators prioritize salient but unrepresentative examples, embedding human fallibility into foundational AI components. During operational oversight, human interventions frequently introduce errors that worsen outcomes compared to autonomous AI handling. In autonomous vehicle testing, driver-initiated disengagements account for over 25% of interventions, often due to premature or misguided overrides reflecting out-of-practice situational awareness, with studies indicating that a majority of such human takeovers constitute errors rather than improvements.^[87]^[77] Feedback loops further compound this: human biases in evaluating AI suggestions can cascade, as seen in experiments where overseers' selective adherence to flawed recommendations amplified decision errors in perceptual and social judgment tasks.^[29] These limitations underscore that HITL, while mitigating certain AI shortcomings, risks substituting algorithmic precision with human unpredictability, particularly absent rigorous debiasing protocols.^[84]

Scalability and Efficiency Challenges

Human intervention in AI systems often introduces significant latency, as manual reviews cannot match the speed of automated processing, limiting applicability in high-throughput environments such as real-time autonomous driving or large-scale data annotation pipelines.^[2] For instance, in machine learning workflows, human labeling for training datasets scales poorly with data volumes exceeding millions of instances, where annotation times can extend from days to months depending on task complexity.^[88] Economic constraints further exacerbate scalability, with human oversight costs rising nonlinearly; estimates indicate that manual quality assurance in AI deployments can account for up to 80% of project budgets in data-intensive applications, deterring widespread adoption beyond niche uses.^[89] Logistical hurdles, including the recruitment and training of specialized evaluators, compound these issues, as maintaining a pool of consistent human reviewers for iterative feedback loops demands substantial infrastructure that many organizations lack.^[90] Efficiency bottlenecks emerge from human cognitive limits, such as fatigue and variability in judgment, which degrade performance over prolonged sessions; studies show error rates in human-AI hybrid evaluations increasing by 15-20% after four hours of continuous review.^[91] In deployment scenarios, over-reliance on HITL for low-confidence predictions creates queueing delays, where unresolved cases accumulate, potentially halting system operations in scalable services like content moderation handling billions of inputs daily.^[16] These challenges are particularly acute in generative AI, where the volume of outputs requiring validation outpaces human capacity, necessitating selective routing mechanisms that themselves introduce decision overhead.^[92] To mitigate without fully eliminating HITL, hybrid approaches like confidence-threshold routing—escalating only ambiguous cases—have been proposed, yet they still face scalability limits in ultra-high-volume systems, as even 1% escalation rates can overwhelm human teams at enterprise scales.^[93] Empirical evidence from industry deployments underscores that unchecked expansion of HITL elements risks transforming efficiency gains from AI into net losses, prompting shifts toward human-over-the-loop models for supervisory roles only.^[94]

Overreliance as a False Safety Guarantee

Overreliance on AI systems within human-in-the-loop (HITL) frameworks occurs when operators exhibit automation bias, favoring automated outputs over their own judgment or evidence, even when the AI provides erroneous recommendations. This phenomenon erodes the presumed safety benefits of human oversight, as empirical studies demonstrate that humans often fail to detect or correct AI mistakes, leading to compounded errors rather than mitigation. For instance, in interactive decision-making experiments, participants overrelied on AI advice for risky financial choices, resulting in suboptimal outcomes for themselves and third parties, independent of the domain.^[95] A comprehensive review of approximately 60 studies across disciplines, including human-computer interaction and psychology, reveals that overreliance prevents humans from effectively addressing AI limitations, such as hallucinations or biases, thereby invalidating simplistic reliance on HITL as a safeguard. Participants exposed to flawed AI support showed diminished independent judgment, particularly when receiving incorrect inputs early in the process, amplifying errors in subsequent human evaluations.^[96]^[97] Automation complacency further exacerbates this issue, where initial AI successes foster undue trust, reducing vigilance and increasing error rates in oversight tasks. In high-stakes applications like generative AI for healthcare diagnostics, humans struggle to monitor outputs consistently, as cognitive tendencies lead to acceptance of plausible but inaccurate results, creating an illusion of reliability without actual risk reduction. Regulatory frameworks, such as the EU AI Act, acknowledge automation bias by mandating awareness training for overseers of high-risk systems, yet evidence indicates that such measures do not eliminate the bias's impact on decision quality.^[98]^[99] Consequently, HITL configurations can propagate AI deficiencies through unchecked human deference, as seen in scenarios where operators inherit and perpetuate algorithmic biases in health-related judgments, yielding decisions no better—or worse—than unaided AI alone. This challenges the causal assumption that human involvement inherently enhances safety, as overreliance systematically undermines error detection, fostering systemic vulnerabilities rather than guarantees.^[100]^[101]

Major Controversies and Debates

Autonomy vs. Oversight in Lethal Systems

The debate over autonomy versus human oversight in lethal systems centers on lethal autonomous weapon systems (LAWS), defined as systems capable of selecting and engaging human targets without further human intervention after activation.^[102] Proponents of greater autonomy argue that it enables faster, more precise engagements in dynamic combat environments, potentially reducing collateral damage through consistent algorithmic decision-making unbound by human fatigue or emotional bias.^[49] For instance, simulations of semi-autonomous systems have demonstrated lower error rates in target discrimination compared to fully human-operated systems under stress, where misidentifications occur in up to 20% of high-pressure scenarios.^[103] However, empirical data on fully autonomous lethal engagements remains limited, as no major military has publicly deployed such systems at scale, with current technologies like loitering munitions still requiring human authorization for final strikes.^[55] Opponents emphasize that removing human oversight undermines accountability and moral judgment, as machines cannot assess proportionality or intent under international humanitarian law, risking indiscriminate harm.^[104] Human Rights Watch analyses highlight cases where semi-autonomous drones caused civilian deaths due to sensor errors or algorithmic misclassification, arguing that full autonomy exacerbates these without recourse to human veto.^[104] At the United Nations Convention on Certain Conventional Weapons (CCW), over 100 states have supported resolutions urging retention of human control, with UN Secretary-General António Guterres stating in May 2025 that LAWS without such oversight are "morally repugnant" and politically unacceptable.^[105] ^[106] A 2024 UN General Assembly vote saw 161 states favor measures against fully autonomous targeting, reflecting widespread concern over proliferation to non-state actors lacking ethical constraints.^[107] U.S. Department of Defense Directive 3000.09, updated January 25, 2023, mandates "appropriate levels of human judgment" for autonomous and semi-autonomous systems but explicitly does not require a human in the loop for target selection or engagement in all cases, provided rigorous testing and senior review occur.^[50] ^[52] This policy balances operational needs—such as countering swarms in peer conflicts—with risks, yet critics contend it insufficiently addresses error propagation in unpredictable terrains, where AI brittleness has led to simulated failure rates exceeding 30% in adversarial jamming scenarios.^[108] Empirical studies on analogous non-lethal autonomous systems, like counter-UAS defenses, show autonomy improving response times by factors of 10 but introducing novel vulnerabilities, such as adversarial AI exploits not foreseen in human-supervised loops.^[109] The controversy persists amid technological advances, with advocates for oversight citing causal risks of escalation—autonomous systems could misinterpret feints as attacks, triggering unintended conflicts—while autonomy supporters invoke first-use deterrence, noting human hesitation has prolonged engagements in historical data from urban warfare.^[110] No peer-reviewed evidence conclusively proves full autonomy superior in reducing lethality errors over hybrid models, where human oversight has prevented erroneous strikes in 95% of reviewed drone operations per declassified reports.^[111] Ongoing CCW talks, stalled since 2019 on definitions, underscore the tension: bans risk unilateral disadvantages, yet unchecked autonomy invites arms races, as evidenced by Russia's 2024 deployment of semi-autonomous Lancet drones without full human pre-approval in Ukraine.^[112] ^[55]

Impact on Human Judgment and AI Errors

Human-in-the-loop (HITL) systems, intended to leverage human oversight for correcting AI outputs, often result in automation bias, where operators over-accept algorithmic recommendations, including erroneous ones, thereby degrading independent human judgment.^[113] Empirical studies demonstrate this effect across domains; for instance, in clinical decision support systems (CDSS), non-specialist users exhibited higher agreement with incorrect AI advice, with automation bias measured as the tendency to endorse wrong recommendations despite contradictory evidence.^[114] This bias arises from reduced vigilance, as humans treat AI outputs as heuristic substitutes for thorough analysis, leading to error propagation in collaborative workflows.^[115] Further evidence indicates that the timing and nature of AI input exacerbate impacts on human cognition: when participants receive flawed algorithmic support prior to their own assessments, it persistently anchors and distorts subsequent judgments, creating feedback loops that amplify perceptual and emotional biases.^[97] In experimental settings, overreliance on AI advice has been shown to reduce cooperation and yield suboptimal outcomes for decision-makers and affected parties, particularly among those with elevated trust in the system.^[95] Such dynamics highlight a causal pathway where HITL, rather than purely mitigating AI flaws, can entrench human deference to machine errors, as seen in aviation and medical contexts where operators ignored disconfirming data favoring automated cues.^[116] While HITL aims to curb AI errors through intervention, it conversely introduces human-specific vulnerabilities, including subjective biases and interpretive inconsistencies that pure AI processes might avoid.^[2] For example, human-AI conflicts emerge from divergent data interpretations, where operator actions override AI logic, potentially injecting errors absent in autonomous runs; this has been critiqued in high-stakes scenarios like nuclear systems, where reliance on human "safeguards" masks latent risks without eliminating them.^[117]^[118] Studies confirm that humans can perpetuate or amplify AI-generated biases, as evaluators in recursive loops accept flawed suggestions more readily when corrections are not mandated, fostering a cycle of diminished scrutiny.^[119] Prolonged HITL exposure also induces deskilling, eroding operators' baseline competencies as reliance supplants skill maintenance; in medicine, AI augmentation has been linked to competency degradation, distinct from mere overreliance, with mixed-method reviews identifying threats to diagnostic proficiency from habitual deferral.^[120] This effect mirrors historical automation patterns, where initial efficiency gains yield long-term expertise atrophy, as operators lose nuanced pattern recognition honed without machine crutches.^[121] Debates persist on HITL's net efficacy: proponents, often from regulatory perspectives, assert it ensures ethical alignment by filtering AI anomalies, yet empirical critiques, including those from defense analysts, argue it fosters illusory safety, as human frailties compound AI unpredictability without scalable oversight.^[122]^[123] Evidence tilts toward caution, with systematic reviews underscoring that unaddressed automation bias and deskilling undermine HITL's purported error-reduction benefits, potentially elevating overall system fallibility in dynamic environments.^[115]^[113]

Regulatory Mandates and Innovation Constraints

Regulatory mandates for human-in-the-loop (HITL) mechanisms in AI systems primarily target high-risk applications, such as those in critical infrastructure, healthcare, and law enforcement, where unchecked automation could pose significant risks to safety or rights. The European Union's AI Act, effective from August 2024, classifies certain AI systems as high-risk and mandates under Article 14 that providers design them to enable effective human oversight, allowing intervention to minimize harms to health, safety, or fundamental rights.^[99] This includes requirements for deployers to monitor operations and ensure humans can override decisions, with non-compliance risking fines up to €35 million or 7% of global turnover.^[124] In the United States, while no comprehensive federal AI law exists as of October 2025, sector-specific guidelines like the National Institute of Standards and Technology (NIST) AI Risk Management Framework emphasize human oversight for trustworthy AI, and executive actions such as the 2023 Biden administration order direct agencies to incorporate HITL in high-stakes deployments.^[125] State-level measures, including California's 2025 laws on AI in employment decision systems, further require human review of automated outputs to mitigate bias or errors.^[126] These mandates impose innovation constraints by elevating development costs and extending timelines through mandatory compliance processes, such as risk assessments, documentation of oversight mechanisms, and ongoing audits. For instance, the EU AI Act's requirements for high-risk systems—encompassing up to 15% of AI deployments—demand built-in transparency and fallback human controls, which can delay market entry by 12-18 months due to certification hurdles, according to industry analyses.^[127] Compliance burdens disproportionately affect smaller firms and startups, which lack resources for legal expertise or extensive testing, creating barriers to entry and favoring large incumbents capable of absorbing €1-5 million in initial setup costs per system.^[128] Empirical evidence from regulatory impact studies indicates that prescriptive HITL rules reduce experimentation velocity; a 2024 Brookings Institution report notes that overly rigid oversight mandates in AI can suppress iterative improvements, as developers prioritize audit-proof designs over boundary-pushing advancements.^[129] Critics argue that HITL mandates reflect a precautionary bias, potentially locking in suboptimal human dependencies amid rapid AI progress, where autonomous systems could outperform supervised ones in speed and consistency. The U.S. White House's 2025 AI Action Plan explicitly warns against state-level regulations that "waste" federal AI funding by imposing undue burdens, estimating that fragmented HITL requirements could divert up to 20% of R&D budgets toward non-value-adding oversight infrastructure rather than core model training.^[125] In sectors like autonomous vehicles, Federal Motor Vehicle Safety Standards mandating human fallback modes have constrained full Level 5 autonomy testing, with data from the National Highway Traffic Safety Administration showing regulatory delays contributing to only 0.1% of U.S. miles driven autonomously as of 2024.^[130] Internationally, Europe's stricter HITL regime risks ceding competitive ground to less-regulated jurisdictions like China, where state-backed AI firms advance without equivalent oversight, potentially eroding Western innovation leadership by 2030, per economic modeling from the Chicago Booth Review.^[131] While intended to enhance safety, these constraints underscore a tension: empirical data on AI error rates decreasing exponentially with scale suggests that mandatory human intervention may inadvertently preserve inefficiencies, as human operators introduce variability and fatigue-related failures at rates exceeding 10% in prolonged monitoring tasks.^[132]

Recent Developments and Future Outlook

Advances from 2023 Onward

Since 2023, human-in-the-loop (HITL) systems have evolved to address limitations in autonomous AI, particularly in high-stakes domains, by incorporating selective human oversight mechanisms that balance efficiency with accountability. A key proposal in medical AI emphasized "human-on-appeal" processes, where AI handles initial decisions but humans intervene only upon review requests, modeled after judicial appeals systems with standards like "clear error" for factual accuracy or "de novo" review for ethical concerns.^[133] This approach, published in 2023, aims to minimize human workload in resource allocation tasks such as ventilator distribution or vaccine prioritization while enhancing fairness and error correction.^[133] In healthcare, HITL integration has seen rapid empirical uptake, with 66% of U.S. physicians reporting AI use in clinical workflows in 2024, up from 38% in 2023, often involving human validation of AI outputs for documentation, billing, and diagnostics to mitigate errors.^[134] By 2025, advancements in agentic AI have further refined HITL through frameworks enabling human-AI collaboration in procedural tasks, such as augmented reality-equipped agents for battlefield medicine or cooking, which empirical studies show improve task completion rates and reduce errors via interactive guidance and feedback loops.^[135] Workplace applications have advanced toward "superagency" models, where HITL pairs human judgment with autonomous AI agents handling complex operations like customer fraud detection or payments, as exemplified by Salesforce's Agentforce platform launched in 2024.^[136] Leveraging 2025 large language models such as OpenAI's o1 and Google's Gemini 2.0 for enhanced reasoning and multimodality, these systems allow humans to refine AI decisions in real-time, boosting productivity without full delegation.^[136] In education, HITL has progressed with generative AI-driven adaptive learning platforms that incorporate student critiques via tagged feedback to personalize STEM content, using techniques like prompt engineering and retrieval-augmented generation for real-time adjustments.^[137] Preliminary 2025 studies demonstrate these systems yield better learning outcomes and student confidence compared to non-interactive AI tools, emphasizing causal feedback for model improvement.^[137]

Emerging Trends in Agentic AI and Beyond

Agentic AI refers to artificial intelligence systems designed to pursue complex goals autonomously through reasoning, planning, execution, and adaptation, marking a shift from reactive models to proactive ones that emerged prominently in 2024 and accelerated into 2025.^[138]^[139] These systems leverage large language models integrated with tools, memory, and multi-step deliberation to handle tasks like software development or data analysis with minimal direct input, as demonstrated in frameworks such as LangChain and AutoGen released or updated in mid-2024.^[140] However, empirical assessments indicate that agentic AI's current limitations in handling nuanced, long-term objectives and error propagation necessitate hybrid architectures incorporating human oversight for high-stakes applications.^[141] A key trend is the proliferation of multi-agent collaboration, where specialized AI agents divide labor on shared objectives, such as in enterprise workflows for compliance checking or supply chain optimization, reducing the cognitive burden on humans who transition to supervisory roles.^[142] For instance, retrieval-augmented generation (RAG) enhanced agents, debuted by major providers like OpenAI and Google in early 2025, enable domain-specific compliance by querying verified data sources autonomously, yet incorporate human-in-the-loop (HITL) veto mechanisms to mitigate hallucinations or biases observed in benchmarks where agent success rates drop below 70% for multi-step tasks without intervention.^[140] This evolution amplifies human capabilities by offloading routine execution, allowing oversight focused on strategic refinement rather than micromanagement, as evidenced in organizational pilots reported in 2025 where agent squads improved productivity by 20-40% under human supervision.^[138]^[143] Beyond core agentic systems, trends point toward symbiotic human-AI paradigms in vertical applications, including defense and HR, where agents manage tactical decisions like threat detection or candidate screening, but human judgment remains integral for ethical alignment and anomaly resolution due to unresolved issues in agent reliability under adversarial conditions.^[144]^[145] Governance frameworks are emerging to balance this, such as dynamic HITL protocols that scale intervention based on confidence scores, with 2025 deployments emphasizing auditable logs and constitutional constraints to prevent unchecked autonomy, as advocated in industry reports highlighting risks of unintended actions in tool-equipped agents.^[146]^[147] Looking further, research trajectories suggest a gradual shift toward verifiable agentic architectures, potentially diminishing constant HITL through advancements in self-correction and formal verification, though full human-out-of-the-loop deployment remains constrained by empirical evidence of persistent failure modes in uncontrolled environments as of late 2025.^[148]^[149]

Potential Shifts Toward Human-Out-of-the-Loop

Developments in agentic AI indicate a trajectory toward systems capable of independent goal pursuit and task execution, minimizing human intervention. Autonomous AI agents, defined as entities that reason, plan, and act with limited oversight, are projected to advance significantly by 2025, enabling workflows such as self-optimizing supply chains and fraud detection without constant human input.^[150] ^[151] For instance, frameworks like those analyzed in 2025 evaluations demonstrate agents handling complex scenarios post-deployment with autonomy as a core feature, shifting from reactive tools to proactive operators.^[152] This evolution is driven by efficiency gains, as AI processes decisions faster than humans in data-intensive environments, potentially reducing latency in enterprise applications.^[138] In military contexts, U.S. Department of Defense Directive 3000.09, updated January 25, 2023, permits the development and deployment of lethal autonomous weapon systems (LAWS), which select and engage targets without further human operator intervention once activated.^[50] ^[48] The policy emphasizes appropriate human judgment in activation and safeguards against unlawful engagements but explicitly allows semi-autonomous and fully autonomous functions, reflecting a doctrinal acceptance of reduced human involvement for operational speed and precision in combat.^[50] This stance contrasts with international calls for bans, yet U.S. positions prioritize technological integration over prohibitions, as evidenced by ongoing programs integrating autonomy into weapon systems.^[153] Such systems aim to mitigate human fatigue and error in high-stakes scenarios, though real-world deployment remains constrained by testing and ethical reviews.^[154] Transportation sectors exhibit practical tests of human-out-of-the-loop operations, particularly in controlled environments. As of October 2025, Tesla's Full Self-Driving (FSD) software has been deployed in the Las Vegas Loop tunnels by The Boring Company, operating passenger vehicles with zero human input during rides, described as smoother than manual driving.^[155] This marks a step beyond supervised autonomy, leveraging neural networks trained on vast real-world data to handle navigation independently.^[156] Broader FSD demonstrations, such as a 2025 coast-to-coast drive requiring only six minutes of human intervention, underscore progress toward Level 5 autonomy, where no human oversight is needed in any condition.^[157] These advancements prioritize AI's superior reaction times over human variability, potentially scaling to unsupervised urban and highway operations as regulatory approvals evolve.^[156] Overall, these shifts are propelled by AI's capacity for consistent, scalable performance in repetitive or hazardous tasks, outpacing human limitations in volume and speed.^[158] However, full realization depends on robust validation, as current implementations retain fallback mechanisms amid unresolved edge cases.^[148] Projections for 2025 forecast widespread adoption in agentic systems, with 79% of enterprises anticipating full-scale integration within three years, signaling a paradigm where human roles transition to initial design and rare exceptions.^[159]

References

[1]
Data Science and Engineering With Human in the Loop, Behind the ...
Apr 27, 2023 · The term human-in-the-loop (HITL) generally refers to the need for human interaction, intervention, and judgment to control or change the outcome of a process.
[2]
What Is Human In The Loop (HITL)? - IBM
Human-in-the-loop machine learning allows humans to provide oversight and input into AI workflows. Here are the primary benefits of human-in-the-loop:.
[3]
Human-in-the-loop machine learning: a state of the art
Aug 17, 2022 · Researchers are defining new types of interactions between humans and machine learning algorithms generically called human-in-the-loop machine learning.
[4]
Improving the Applicability of AI for Psychiatric Applications through ...
May 26, 2022 · Human-in-the-Loop AI. Human-in-the-loop is a general term for processes that enable collaboration between humans and machines. This paper ...
[5]
Human-in-the-Loop Machine Learning (HITL) Explained - Encord
Dec 11, 2024 · Human-in-the-loop (HITL) is an iterative feedback process whereby a human (or team) interacts with an algorithmically-generated system.
[6]
Human in the Loop Machine Learning: The Key to Better Models
Jan 16, 2025 · Human in the loop machine learning refers to human involvement in machine learning workflows to reduce errors and improve model performance.
[7]
[PDF] Humans in the Loop - wp0 | Vanderbilt University
Mar 20, 2023 · From lethal drones to cancer diagnostics, humans are increasingly working with complex and artificially intelligent algorithms to make ...
[8]
Humans in the Loop: The Design of Interactive AI Systems
Oct 20, 2019 · The design of systems that fundamentally infuse technology and human interaction, in the pursuit of new tools for musical and other forms of human creativity.
[9]
Human-in-the-loop or AI-in-the-loop? Automate or Collaborate? - arXiv
Dec 18, 2024 · The Human-in-the-loop paradigm for developing AI systems typically treats humans either as data-labeling oracles (Settles 2009) or as a source ...
[10]
When can we Kick (Some) Humans “Out of the Loop”? An ...
May 15, 2024 · Because of its novelty, a norm of keeping humans “in the loop” wherever AI mechanisms are deployed has become synonymous with good ethical ...
[11]
Challenging the Human-in-the-loop in Algorithmic Decision-making
May 17, 2024 · We discuss the role of humans in algorithmic decision-making ( ADM ) for socially relevant problems from a technical and philosophical perspective.
[12]
What's Wrong with Wanting a “Human in the Loop”?
Jun 23, 2022 · If an AI-enabled weapon system works better without a human in the loop, we are going to have a very hard time justifying the decision to keep a ...Missing: controversies | Show results with:controversies
[13]
AI in the Loop: Humans Must Remain in Charge | Stanford HAI
Oct 17, 2022 · Instead, Stanford HAI faculty argue, practitioners should focus on AI in the loop – where humans remain in charge. Artificial intelligence ...
[14]
Human-in-the-Loop - an overview | ScienceDirect Topics
Human-in-the-loop refers to learning models that require human interaction, allowing humans to modify the output of the system. This approach involves human ...
[15]
Human in the Loop (HITL) in Practice: A Guide to Core HITL ... - Splunk
Jul 3, 2025 · Human in the Loop refers to the crucial role of human intervention in an automated or AI-driven process.
[16]
Humans on the Loop vs. In the Loop: Balancing Automation
Feb 12, 2025 · "Humans in the loop" means direct human involvement in decisions, while "humans on the loop" means humans supervise and intervene only when ...
[17]
Human in the Loop vs. Human on the Loop: Navigating the Future of AI
Jun 11, 2025 · Human on the loop is a little different. It's where the system can operate autonomously, but there's a human monitoring the process and able to ...
[18]
What is the difference between human-in-the-loop and human out of ...
On the other hand, Human-in-the-Loop systems incorporate human input and oversight, ensuring that any issues are addressed and mitigated.
[19]
Return of cybernetics | Nature Machine Intelligence
Sep 11, 2019 · Norbert Wiener saw intelligent behaviour emerging from a complex interaction of feedback loops. He noticed such feedback processes ...
[20]
From Cybernetics to AI: the pioneering work of Norbert Wiener
Apr 25, 2024 · A pioneer in computer science and AI research. Wiener realized that almost all complex systems are driven by feedback loops of information.Missing: human interaction
[21]
Cybernetics - What is Intelligence? | Antikythera
... human “in the loop” impossible; the precision and response time needed were superhuman. Even the G-forces incurred by aerial maneuvers soon began to exceed ...
[22]
Norbert Wiener. The Father Of Cybernetics - Quantum Zeitgeist
Jan 10, 2025 · His book “Cybernetics: Or Control and Communication in the Animal and the Machine” introduced the concept of feedback loops. These loops are a ...
[23]
The Human Use of Human Beings: Cybernetics Pioneer Norbert ...
Jun 15, 2018 · Wiener explains why, under this model of information systems, communication and control are inexorably linked.Missing: interaction | Show results with:interaction
[24]
Feedback Loops and Learning from the Past (Chapter 1)
Within this new world of technologically driven communication, Wiener is saying, machines can learn from experience, just as humans can. In Wiener's formulation ...
[25]
History and Evolution of Machine Learning: A Timeline - TechTarget
Jun 13, 2024 · Frank Rosenblatt developed the perceptron, an early ANN that could learn from data and became the foundation for modern neural networks. 1959.<|separator|>
[26]
[2108.00941] A Survey of Human-in-the-loop for Machine Learning
Aug 2, 2021 · In this paper, we survey existing works on human-in-the-loop from a data perspective and classify them into three categories with a progressive relationship.
[27]
Training language models to follow instructions with human feedback
Mar 4, 2022 · In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback.
[28]
The History and Risks of Reinforcement Learning and Human ...
Oct 20, 2023 · Reinforcement learning from human feedback (RLHF) has emerged as a powerful technique to make large language models (LLMs) easier to use and more effective.
[29]
How human–AI feedback loops alter human perceptual, emotional ...
Dec 18, 2024 · We reveal a feedback loop where human–AI interactions alter processes underlying human perceptual, emotional and social judgements, subsequently amplifying ...
[30]
Human-in-the-Loop Control: Applications and Categorization
I review the variety of human-in-the-loop control systems, identifying four architectural templates for the field.
[31]
A model for types and levels of human interaction with automation
May 31, 2000 · We outline a model for types and levels of automation that provides a framework and an objective basis for deciding which system functions should be automated.
[32]
A model for types and levels of human interaction with automation ...
The stages are (1) information alone, (2) assessment of that information, and (3) a decision based on that assessment, in line with Parasuraman et al. (2000) .
[33]
A model for types and levels of human interaction with automation
A model for types and levels of automation is outlined that can be applied to four broad classes of functions: 1) information acquisition; 2) information ...
[34]
Level of Automation - an overview | ScienceDirect Topics
Within each type of automation, Parasuraman and Wickens (2008) suggest that the automation may range in its functionality, meaning that there may be varying ...
[35]
Human-In-The-Loop: What, How and Why | Devoteam
The human-in-the-loop (HITL) process is iterative, with a continuous cycle of data collection, model training, human feedback, and model refinement.
[36]
A survey on active learning and human-in-the-loop deep learning for ...
In this review we investigate the role that humans might play in the development and deployment of deep learning enabled diagnostic applications.
[37]
Illustrating Reinforcement Learning from Human Feedback (RLHF)
Dec 9, 2022 · RLHF has enabled language models to begin to align a model trained on a general corpus of text data to that of complex human values.
[38]
A Survey of Reinforcement Learning from Human Feedback - arXiv
Dec 22, 2023 · This article provides a comprehensive overview of the fundamentals of RLHF, exploring the intricate dynamics between RL agents and human input.
[39]
Human-in-the-Loop Reinforcement Learning: A Survey and Position ...
Jan 30, 2024 · This paper aims to inform readers about current explainability methods in HITL RL. It also shows how the application of explainable AI (xAI) and specific ...Missing: modern | Show results with:modern
[40]
[PDF] Human-in-the-loop Machine Learning: A Macro-Micro Perspective
Typical applications include autonomous vehicles, game playing (e.g., AlphaGo), precise medical diagnosis and prescription, assistive robots, face recognition, ...
[41]
Active Learning Machine Learning: How It Reduces Labeling Costs
May 22, 2025 · 1 Active learning helps reduce labeling costs by selecting only the most useful data points for training, instead of treating all samples ...
[42]
Active Learning Performance in Labeling Radiology Images Is 90 ...
Nov 30, 2021 · We conclude that the human effort required to label an image dataset can be reduced by approximately 90% in most cases by using the described ...
[43]
Supercharge Your Classifier Development with Active Learning
Jul 7, 2025 · Studies show active learning can achieve human-comparable accuracy with dramatic efficiency gains—up to 80% less effort compared to passive ...
[44]
Human-In-The-Loop Machine Learning for Safe and Ethical ... - arXiv
Aug 22, 2024 · These applications demonstrate the effectiveness of CL in guiding autonomous agents through progressively challenging tasks to achieve optimal ...<|separator|>
[45]
[PDF] Human-in-the-Loop Distributed Simulation and Validation of ...
An example of a typical setup for a fully autonomous system is shown in. Figure 2. Low and middle level control algorithms are handled by an inner loop ...
[46]
A Vehicle-in-the-Loop Simulator with AI-Powered Digital Twins for ...
Jul 3, 2025 · The AI-powered DT models ensure superior simulation fidelity. ... Out of these concerns, we develop a deep-learning-based DT to improve its ...
[47]
System Validation with Human Subjects Using MATLAB and ADAMS
May 28, 2025 · Integrated Co-Simulation and Control Framework for Intelligent Bionic Hands: System Validation with Human Subjects Using MATLAB and ADAMS.<|separator|>
[48]
Defense Primer: U.S. Policy on Lethal Autonomous Weapon Systems
Jan 2, 2025 · Another category is semi-autonomous, or "human in the loop," weapon systems that "only engage individual targets or specific target groups that ...
[49]
Pros and Cons of Autonomous Weapons Systems
“Human-in-the-loop weapons [are] robots that can select targets and deliver force only with a human command.”45 Numerous examples of the first type already are ...
[50]
[PDF] DoD Directive 3000.09, "Autonomy in Weapon Systems
Jan 25, 2023 · Autonomous and semi-autonomous weapon systems will be designed to allow commanders and operators to exercise appropriate levels of human ...
[51]
DoD Announces Update to DoD Directive 3000.09, 'Autonomy In ...
Jan 25, 2023 · Autonomous and semi-autonomous weapon systems will be designed to allow commanders and operators to exercise appropriate levels of human ...
[52]
Autonomous Weapon Systems: No Human-in-the-Loop Required ...
May 22, 2025 · Second, there is a myth that the directive requires a human in the loop for the use of force at the tactical level, which is also not the case.
[53]
Lethal Autonomous Weapons Systems & International Law
Jan 24, 2025 · Semi-autonomous (human-in-the-loop): Systems that, once activated, can select targets and apply force – but only with human authorization.
[54]
Please Stop Saying "Human-In-The-Loop" - Institute for Future Conflict
Sep 3, 2024 · The idiom “human-in-the-loop" presumes a machine decision loop and then asks: where is the human relative to that pre-existing loop? Rather than ...
[55]
The Future of Warfare: National Positions on the Governance of ...
Feb 11, 2025 · Lethal autonomous weapons systems (LAWS), such as drones and autonomous missile systems, are no longer a theoretical concern.
[56]
Beyond a Human “In the Loop”: Strategic Stability and Artificial ...
Nov 12, 2024 · Beyond endorsing this human “in the loop” precaution, nuclear powers should adopt the following additional recommendations to minimize the ...
[57]
DOD Is Updating Its Decade-Old Autonomous Weapons Policy, but ...
Jun 6, 2022 · No human in the loop is required. The critical feature of this definition is that the policy threshold for autonomy relates specifically and ...
[58]
AI in health: keeping the human in the loop - PMC - NIH
Jun 20, 2023 · AI in health emphasizes keeping humans in the loop, with studies showing that human use of AI devices is more likely to cause harm than device ...
[59]
Human–AI collectives most accurately diagnose clinical vignettes
Our results demonstrate the potential of combining human medical expertise with LLMs to enhance accuracy and reduce errors in open-ended medical diagnostics.
[60]
Impact of human and artificial intelligence collaboration on workload ...
Nov 30, 2024 · This meta-analysis evaluates the impact of human-AI collaboration on image interpretation workload. Four databases were searched for studies comparing reading ...
[61]
Investigating human–AI collaboration protocols in medical diagnosis
Human-AI collaboration protocols (HAI-CP) are studied to evaluate how humans and AI interact in medical tasks. AI-first protocols showed higher accuracy.
[62]
Human–AI collectives make the most accurate medical diagnoses
Jun 20, 2025 · Realistic simulations using more than 2,100 clinical vignettes. The researchers used data from the Human Diagnosis Project, which provides ...
[63]
Human‐in‐the‐loop machine learning for healthcare: Current ...
Aug 23, 2024 · Human-in-the-loop machine learning for healthcare: Current progress and future opportunities in electronic health records. Han Yuan ...
[64]
Realyze Intelligence's Human-in-the-Loop AI Widens Clinical Trials ...
Aug 8, 2024 · Kathryn Doornbos explains how Realyze Intelligence's software can uncover links for patients to receive clinical research for treatment.
[65]
For trustworthy AI, keep the human in the loop | Nature Medicine
Oct 10, 2025 · Truly supporting the human in the loop will require rigorous trials, real-world testing, and a science of implementation that treats clinicians ...
[66]
Safety assurance mechanisms of collaborative robotic systems in ...
Safety assurance for cobots involves mechanisms to avoid interference, quantifying collisions, and real-time controls. ISO 10,218–1/2 classifies these ...Human-Robot Collaboration... · Types Of Human-Robot... · Integrated Cobotic Systems
[67]
Ergonomic human-robot collaboration in industry: A review - PMC
In this review article, we provide an overview of the existing ergonomics assessment tools as well as the available monitoring technologies.
[68]
A shared control architecture for human-in-the-loop robotics ...
Abstract: We propose a shared control architecture to enable the modeling of human-in-the-loop cyber physical systems (HiLCPS) in robotics applications.
[69]
Precise and Dexterous Robotic Manipulation via Human-in ... - arXiv
Oct 29, 2024 · We present a human-in-the-loop vision-based RL system that demonstrates impressive performance on a diverse set of dexterous manipulation tasks.
[70]
(PDF) Human-Robot Collaboration: Enhancing Productivity and Safety
Jan 5, 2025 · This paper explores the potential of HRC in improving operational efficiency, reducing workplace hazards, and fostering ergonomic designs.
[71]
Human-in-the-Loop Robot Learning for Smart Manufacturing
Jan 10, 2025 · This study provides a holistic literature review for understanding HITL robot learning within an industrial context from a human-centric perspective.
[72]
Waymo's autonomous vehicles are significantly safer than human ...
Sep 6, 2023 · Today, we're sharing new research led by Swiss Re which shows Waymo's autonomous vehicles are significantly safer than those driven by humans.
[73]
Waymo Reduces Crash Rates Compared to Human Drivers Over 7+ ...
Jan 17, 2024 · An 85% reduction or 6.8 times lower crash rate involving any injury, from minor to severe and fatal cases (0.41 incidence per million miles for ...<|separator|>
[74]
Deep Dive: Tesla, Waymo, and the Great Sensor Debate
Jul 8, 2025 · Tesla's approach to self-driving may be cheaper and easier to scale, but safety remains a key question. Waymo, for example, has expressed doubts ...
[75]
Teleoperations will keep humans in the AV loop SAE-MA-03752
“A human in the loop is the only viable solution.” Tiwari believes that teleoperation is not a whole-cloth replacement for self-driving, but rather an ...
[76]
Safety of Level 4 Autonomous Vehicle: Still Human in the Loop
Jan 5, 2024 · We find level 4 AV is critical because there is still the human in the loop in the AV,. which leads to several dilemmas. i) The liability ...
[77]
Error in the Loop: How Human Mistakes Can Improve Algorithmic ...
Jul 3, 2025 · We explore the accuracy-enhancing potential of a restricted human-in-the-loop approach: allowing overrides of algorithmic denials. This strategy ...
[78]
Constructing Ethical AI Based on the “Human-in-the-Loop” System
Nov 13, 2023 · This paper draws on the feasibility of the HITL system and analyzes how ethical AIs are possible when using the HITL system.
[79]
Human-In-The-Loop Machine Learning for Safe and Ethical ... - arXiv
Aug 22, 2024 · We present a review of HITL-ML for AVs, focusing on Curriculum Learning (CL), Human-In-The-Loop Reinforcement Learning (HITL-RL), Active Learning (AL), and ...Missing: applications | Show results with:applications
[80]
In the Loop? Armed Robots and the Future of War | Brookings
The US military went into Iraq with just a handful of drones in the air and zero unmanned systems on the ground, none of them armed. Today, there are over ...
[81]
Full Self-Driving (Supervised) - Tesla
Tesla uses billions of miles of anonymous real-world driving data to train Full Self-Driving (Supervised) to take care of the most stressful parts of daily ...Menu · Tesla's FSD · Tesla Australia · Tesla Hong KongMissing: case | Show results with:case
[82]
Real-World Case Studies of Human-AI Collaboration - SmythOS
The creative industries showcase perhaps the most fascinating examples of human-AI symbiosis. Artists are discovering how AI can augment their creative ...
[83]
https://airecruiterlab.com/resources/fortune-500-ai-recruitment
[84]
[PDF] Towards a Standard for Identifying and Managing Bias in Artificial ...
Mar 15, 2022 · This document identifies three categories of AI bias: systemic, statistical, and human, and describes challenges for mitigating bias.<|separator|>
[85]
Exploring the Role of Annotator Cognitive Biases in NLP - arXiv
Apr 29, 2024 · The cognitive biases of annotators, inherent in human decision-making, can perpetuate and even amplify existing social disparities in AI models.
[86]
The impact of inconsistent human annotations on AI driven clinical ...
Feb 21, 2023 · As the focus of this study is investigating the impact of clinical annotator disagreements on model performance, rather than on improving label ...
[87]
A matched case-control analysis of autonomous vs human-driven ...
Jun 18, 2024 · According to Dixit et al. in 2016, 56.1% of disengagements were attributed to system failures, 26.57% were initiated by the driver, and 9.89% ...
[88]
Human-in-the-Loop AI: Why It Matters in the Era of GenAI | Tredence
Aug 25, 2025 · The concept of human-in-the-loop emphasizes making human values clear while working and coexisting with AI/ML systems (Source). And human ...Missing: evolution history
[89]
The Challenges of Large-Scale Human-in-the-Loop AI Evaluations
Dec 2, 2024 · Cost and Logistical Constraints: Scaling HITL systems requires significant investment in recruitment, training, and management of human ...
[90]
From Bottlenecks to Flywheels: Human-in-the-Loop AI in Practice
Oct 9, 2025 · As businesses scale, manual human review processes often become bottlenecks because they're costly, inconsistent, and difficult to monitor.
[91]
Human-in-the-Loop: Balancing Automation and Expert Labelers
Mar 31, 2025 · One of the main challenges is scalability. While automation allows AI systems to process vast amounts of data, including human oversight, it can ...
[92]
Why AI still needs you: Exploring Human-in-the-Loop systems
Jun 6, 2025 · ‍Human-in-the-Loop (HITL) is a design approach in which artificial intelligence systems are intentionally built to incorporate human ...Why Hitl Matters Today · Hitl In Agent Architectures · When To Use Hitl (and When...
[93]
The Rise of Agentic AI: Why Human-in-the-Loop Still Matters - iMerit
These systems route only low-confidence outputs or flagged anomalies to human reviewers, balancing oversight with efficiency. Agents also undergo multi-step ...
[94]
Beyond human-in-the-loop: Sensemaking between artificial ...
Human-out-of-the-loop (refers to AI involved): AI autonomously makes and ... Conceptualisation of human-on-the-loop haptic teleoperation with fully ...2. Literature Review · 4. Results · Appendix B
[95]
An experimental study on the extent and costs of overreliance on AI
Overreliance on AI advice negatively affects human cooperation, leading to undesired results for advisees and third parties. •. Participants with higher trust ...Missing: false | Show results with:false
[96]
[PDF] Overreliance on AI Literature Review - Microsoft
As we describe in this report, humans are unable to mitigate AI shortcomings when they start over-relying on AI systems. Therefore, calls for human oversight ...
[97]
The impact of AI errors in a human-in-the-loop process - PMC - NIH
Jan 7, 2024 · Our results show that human judgment is affected when participants receive incorrect algorithmic support, particularly when they receive it before providing ...
[98]
You Won't Get GenAI Right if Human Oversight is Wrong | BCG
Mar 27, 2025 · Human review is often undermined by automation bias, escalation roadblocks and disincentives, and evaluations steered by vibes rather than ...
[99]
Article 14: Human Oversight | EU Artificial Intelligence Act
Human oversight shall aim to prevent or minimise the risks to health, safety or fundamental rights that may emerge when a high-risk AI system is used.
[100]
Humans inherit artificial intelligence biases | Scientific Reports
Oct 3, 2023 · The current research aims to examine how biased AI recommendations can influence people's decision-making in a health-related task and to test ...
[101]
AI Safety and Automation Bias
Automation bias is the tendency for an individual to over-rely on an automated system. It can lead to increased risk of accidents, errors, and other adverse ...Missing: false | Show results with:false
[102]
International Discussions Concerning Lethal Autonomous Weapon ...
Feb 25, 2025 · Lethal autonomous weapon systems (LAWS), or weapons designed to independently select and engage targets without the need for manual human control.
[103]
Autonomous weapons are the moral choice - Atlantic Council
Nov 2, 2023 · But opponents of autonomy argue that the latter, LAWS, remove human oversight from the process of killing, with the International Committee of ...
[104]
A Hazard to Human Rights: Autonomous Weapons Systems and ...
Apr 28, 2025 · Autonomous weapons systems present numerous risks to humanity, most of which infringe on fundamental obligations and principles of international human rights ...
[105]
UN chief calls for global ban on 'killer robots' | UN News
May 14, 2025 · UN Secretary-General António Guterres has once again called for a global ban on lethal autonomous weapon systems – machines capable of taking human lives ...
[106]
Lethal Autonomous Weapon System 'Politically Unacceptable ...
May 12, 2025 · Machines that have the power and discretion to take human lives without human control are politically unacceptable, morally repugnant and should be banned by ...
[107]
161 states vote against the machine at the UN General Assembly
Nov 5, 2024 · The voting result on Resolution L.77 was 161 states in favour and 3 against, with 13 abstentions. A small step forward towards the much needed treaty on ...
[108]
Understanding the Global Debate on Lethal Autonomous Weapons ...
Aug 30, 2024 · This article explores the global debate on lethal autonomous weapons systems (LAWS), highlighting the convergences, complexities, and differences within and ...
[109]
Meaningful Human Control over Autonomous Systems
Autonomous weapon systems are “robot weapons that once launched will select and engage targets without further human intervention” (Altmann et al., 2013: 73).
[110]
[PDF] The Debate Over Autonomous Weapons Systems
They argue that AWS should be banned because these systems lack human qualities, such as the ability to relate to other humans and to apply human judgment, that ...
[111]
Laws on LAWS: Regulating the Lethal Autonomous Weapon Systems
Sep 21, 2023 · Full autonomous weapons operate without human involvement, autonomous weapons require human oversight, allowing the operator to monitor and ...
[112]
[PDF] Lethal autonomous weapons systems - General Assembly - UN.org.
Jul 1, 2024 · It includes definitions and characterizations; challenges, concerns and potential benefits; deliberations by States; next steps; and the ...<|separator|>
[113]
Automation bias: a systematic review of frequency, effect mediators ...
These errors can be classified as automation bias (AB), by which users tend to over-accept computer output 'as a heuristic replacement of vigilant information ...
[114]
Automation Bias in AI-Decision Support: Results from an Empirical ...
Aug 30, 2024 · Automation bias in AI-CDSS is measured by agreement with wrong AI recommendations. Non-specialists are most susceptible, and higher perceived ...
[115]
Exploring automation bias in human–AI collaboration: a review and ...
Jul 3, 2025 · This research systematically investigates how AB manifests in these settings and the cognitive mechanisms that influence it.
[116]
The effects of explanations on automation bias - ScienceDirect
1. Evidence of automation bias was found in both experiments: users were more likely to pursue the same decisions as those suggested by the IA than they were ...
[117]
Human intelligence conflict and its impact on process system safety
Human-AI conflict in IA-based system operation can arise due to differences in observation, interpretation, and control action.
[118]
Keeping humans in the loop is not enough to make AI safe for ...
Feb 16, 2023 · Keeping humans in the loop is not enough to make AI systems safe. In fact, relying on this safeguard could result in a hidden increase of risk.Missing: controversies | Show results with:controversies
[119]
Bias in the Loop: How Humans Evaluate AI-Generated Suggestions
Sep 10, 2025 · Moshe Glickman and Tali Sharot “How human–AI feedback loops alter human perceptual, emotional and social judgements” In Nature Human ...
[120]
AI-induced Deskilling in Medicine: A Mixed-Method Review and ...
Aug 27, 2025 · This paper focuses on two interrelated yet conceptually distinct threats: deskilling, the degradation of previously acquired competencies due to ...
[121]
[PDF] Addressing Deskilling as a Result of Human-AI Augmentation in the ...
Jan 17, 2025 · The concept of deskilling was formed in the 20th century to explain how automation leads to a loss of practical knowledge and artisan skill sets ...
[122]
Challenges and Limitations of Human Oversight in Ethical Artificial ...
It focuses on the digitization of health care and the ethical dilemmas that arise around human-nonhuman collaboration.
[123]
Why A 'Human In The Loop' Can't Control AI: Richard Danzig
Jun 1, 2018 · When you add artificial intelligence to the mix, even if the human operator does everything right, the AI can still introduce errors of its own.
[124]
AI Act | Shaping Europe's digital future - European Union
Once an AI system is on the market, authorities are in charge of market surveillance, deployers ensure human oversight and monitoring, and providers have a ...Regulation - EU - 2024/1689 · AI Pact · AI Factories · European AI Office
[125]
[PDF] America's AI Action Plan - The White House
Jul 10, 2025 · The Federal government should not allow AI-related. Federal funding to be directed toward states with burdensome AI regulations that waste these.
[126]
California's Recent and Pending AI Laws Affecting Employers
Jun 27, 2025 · AI must remain a tool controlled by humans, not the other way around.” The bill grants workers the right to access and correct data used by ADS, ...
[127]
EU AI Act's Burdensome Regulations Could Impair AI Innovation
Feb 21, 2025 · These burdensome regulations put AI companies at a competitive disadvantage by driving up compliance costs, delaying product launches, and ...
[128]
Will Regulating AI Hinder Innovation? - Trullion
May 17, 2023 · Strict regulations can lead to increased costs, bureaucratic processes, and barriers to entry, limiting the participation of smaller players in ...AI regulation: what's at stake · Economic Impacts and Job...
[129]
Balancing market innovation incentives and regulation in AI
Sep 24, 2024 · Some AI experts argue that regulations might be premature given the technology's early state, while others believe they must be implemented immediately.
[130]
The $38 Billion Mistake: Why AI Regulation Could Crush Florida's ...
Jun 26, 2025 · Innovation Slowdown: When regulations make it harder to experiment with AI, companies become more cautious about developing new AI applications.
[131]
AI Regulation and Its Impact on Future Innovations - Chicago Booth
May 19, 2024 · The panelists explored multi-jurisdictional issues, highlighting that AI introduces unique challenges beyond traditional data concerns.
[132]
Is human oversight to AI systems still possible? - ScienceDirect.com
Mar 25, 2025 · In this editorial we briefly discuss the evolving role of human oversight in AI governance, focusing on the practical, technical, and ethical dimensions of ...
[133]
How AI can learn from the law: putting humans in the loop only on ...
Aug 25, 2023 · We propose that human expert judges be included via appeals processes for review of algorithmic decisions. Thus, the human intervenes only in a limited number ...
[134]
AMA: Physician enthusiasm grows for health care AI
Feb 12, 2025 · About three in five (66%) physicians surveyed in 2024 indicated they currently use AI in their practice, up significantly from 38% in 2023. The ...
[135]
Towards Effective Human-in-the-Loop Assistive AI Agents
### Summary of Main Contributions to Human-in-the-Loop Assistive AI Agents
[136]
AI in the workplace: A report for 2025 - McKinsey
Jan 28, 2025 · This world includes humans in the loop, bringing people and machines together in a state of superagency that increases personal productivity and ...
[137]
Human-in-the-Loop Systems for Adaptive Learning Using Generative AI
### Advances, Methods, and Results in HITL for Adaptive Learning Systems
[138]
Seizing the agentic AI advantage - McKinsey
Jun 13, 2025 · The agentic approach: Human workers were elevated to supervisory roles, overseeing squads of AI agents, each contributing to a shared objective ...
[139]
Human in the Loop – Agentic AI: The Future of Responsible Autonomy
Aug 13, 2025 · This new wave, known as Agentic AI, shifts AI from being reactive to being proactive and goal-driven. These AI agents can autonomously search ...
[140]
State of the Art of Agentic AI Transformation | Bain & Company
Sep 23, 2025 · As noted earlier, many tech giants debuted their visions of agentic AI in the first half of 2025. ... Plan for human-in-the-loop oversight for now ...
[141]
https://hbr.org/2025/10/why-agentic-ai-projects-fail-and-how-to-set-yours-up-for-success
[142]
The agentic organization: A new operating model for AI | McKinsey
Sep 26, 2025 · Companies are moving toward a new paradigm of humans working together with virtual and physical AI agents to create value.
[143]
https://genesishumanexperience.com/2025/10/19/ai-agent-trends-of-2025-entering-the-agentic-era-of-autonomous-intelligence/
[144]
Agentic AI is the Future of Defense … But Human Oversight is Still ...
Sep 17, 2025 · The promise of agentic AI is that the time between detection and resolution is reduced considerably because the autonomous vehicle can protect ...
[145]
Agentic AI in HR: Balancing Autonomy and Human Oversight
Jul 10, 2025 · Agentic AI is a fundamental evolution of traditional systems, combining autonomy, goal-driven behavior, and adaptability to operate independently.
[146]
Agentic AI security: Risks & governance for enterprises | McKinsey
Oct 16, 2025 · Autonomous and goal driven, agentic AI systems are able to reason, plan, act, and adapt without human oversight—powerful new capabilities that ...
[147]
Agentic AI: Balancing autonomy and accountability - CIO
Sep 29, 2025 · Agentic AI applications must blend the right mix of autonomy and human intervention to build the trust and accuracy that ensure success.
[148]
How Human-in-the-Loop Is Evolving with AI Agents | Built In
Aug 13, 2025 · As AI agents begin to operate with greater determinism, human-in-the-loop will evolve to focus less on oversight and more on innovative tasks ...
[149]
Towards Effective Human-in-the-Loop Assistive AI Agents - arXiv
Jul 24, 2025 · Effective human-AI collaboration for physical task completion has significant potential in both everyday activities and professional domains ...
[150]
2025 will bring advancements in autonomous AI agents that can ...
Jan 15, 2025 · Autonomous AI agents capable of planning, executing, and adapting tasks with minimal human intervention are expected to make significant strides in 2025.
[151]
Autonomous generative AI agents: Under development - Deloitte
Nov 19, 2024 · Autonomous generative AI agents could increase the productivity of knowledge workers and make workflows of all kinds more efficient.
[152]
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and ...
May 15, 2025 · Autonomy: A central feature of AI Agents is their ability to function with minimal or no human intervention after deployment [57] . Once ...Ai Agents Vs. Agentic Ai: A... · Ii-A Language Models As The... · Iv Application Of Ai Agents...
[153]
United States, Use of Autonomous Weapons
Another category is semi-autonomous, or “human in the loop,” weapon systems that “only engage individual targets or specific target groups that have been ...<|separator|>
[154]
Review of the 2023 US Policy on Autonomy in Weapons Systems
Feb 14, 2023 · The 2023 DoD Directive 3000.09 on Autonomy in Weapons Systems revises, but does not radically change, the department's original policy on the topic.
[155]
https://teslanorth.com/2025/10/25/boring-company-starts-using-tesla-fsd-for-passenger-rides-in-las-vegas-loop/
[156]
AI & Robotics | Tesla
Tesla Optimus. Create a general purpose, bi-pedal, autonomous humanoid robot capable of performing unsafe, repetitive or boring tasks. Achieving that end goal ...
[157]
Tesla FSD: NYC to LA with just 6 MINUTES of human driving
Jan 1, 2025 · Experience a groundbreaking coast-to-coast Tesla Full Self-Driving (FSD) journey: A semi-autonomous Cannonball (ish) Run from NYC to LA, ...Missing: loop | Show results with:loop
[158]
Fork in the road: Where does human-out-of-the-loop AI lead?
Jul 9, 2024 · Human-out-of-the-loop AI empowers intelligent systems to operate autonomously with minimal or no human intervention.
[159]
Latest Trends and Security Concerns - Agentic AI
Sep 20, 2025 · 79% of surveyed enterprises expect full-scale agentic AI adoption within the next three years [3]. Governance Readiness: 89% of enterprises ...