CoChat
Sign Up
Cited by 1Ethics and Social Impacts of AIHuman-Automation Interaction and SafetyArtificial Intelligence Applications

The Verification Bottleneck in AI-Accelerated Knowledge Work: Queueing Epistemics, Verification Capacity and Adaptive Human Oversight

Open access
TAN, Kwan Hong·2026·Open Access Journal of Multidisciplinary Research
DOI 10.47760/oajmr.2026.v02i09.001
Open this paper in CoChat — free account

Abstract

Generative artificial intelligence can increase the rate at which knowledge work is produced, but it does not proportionally increase the human capacity to verify claims, assumptions, sources, calculations, and consequential recommendations. This paper develops the verification bottleneck as a distinct socio-technical mechanism and introduces queueing epistemics, a framework for analysing knowledge reliability when verification-demanding outputs arrive faster than bounded review capacity can process them. The model defines the Epistemic Load Ratio (ELR), derives a Verification Capacity Frontier that jointly constrains delay and review quality, and formalizes Marginal Verification Value for allocating scarce human attention across heterogeneous tasks. Four propositions show that verification delay is convex in load, review quality can deteriorate under overload, raw productivity and reliable throughput can diverge, and risk-ranked adaptive review can dominate volume-based oversight under heterogeneous expected loss. A synthetic Monte Carlo stress test compares blanket verification, a production-first 20% sample, a static risk threshold, and Queue-Aware Adaptive Verification (QAV) across five workload acceleration levels. At the highest acceleration level, QAV maintained ELR near 0.80 with no verification backlog, achieved mean net value of 5.732 synthetic units per task, and reduced severe escaped errors by 56.7% relative to production-first sampling and by 37.5% relative to static thresholding. Sensitivity analysis across 27 combinations of risk-estimation noise, capacity reserve, and overload severity preserved a positive QAV net-value advantage over the best comparator in every tested condition. These results are model-conditional rather than empirical estimates. They nevertheless establish a general design principle: AI productivity should be governed by verification capacity, not generation capacity alone. The paper concludes with an operational control architecture and auditable metrics for organisations seeking to scale AI-assisted knowledge work without converting speed into epistemic fragility.

AI snapshot

Written from the paper’s full text.

TL;DR

Generative AI raises output volume but not human verification capacity, creating a nonlinear verification bottleneck that can make nominal oversight infeasible. The paper formalizes this with queueing epistemics and a synthetic Monte Carlo test showing queue-aware adaptive verification can outperform blanket, fixed-percentage, and static-threshold review under high acceleration.

Key findings

  • Verification delay is convex in load: in the stylized M/M/1 model, normalized expected queue delay is 1 at ELR = 0.50, 4 at ELR = 0.80, and 9 at ELR = 0.90, so a 10-point rise from 0.80 to 0.90 more than doubles expected waiting.
  • At α = 3, blanket verification reaches mean ELR 4.342, leaves a backlog equal to 79.7% of candidate units, and completes review for only 20.29% of units despite intending to review all.
  • Queue-Aware Adaptive Verification (QAV) maintained ELR near 0.80 with no verification backlog, achieved mean net value of 5.732 synthetic units per task, and reduced severe escaped errors by 56.7% relative to production-first 20% sampling and by 37.5% relative to static risk thresholding.
  • Across a 27-condition sensitivity grid varying risk-estimation noise, capacity reserve, and overload severity, QAV retained a positive net-value advantage over the best comparator in every tested condition, ranging from +0.232 to +0.492 synthetic units per task.

Methods

  • The paper develops a conceptual queueing model of verification demand versus capacity, defining the Epistemic Load Ratio (ELR), Verification Capacity Frontier (VCF), Marginal Verification Value (MVV), and Queue-Aware Adaptive Verification (QAV).
  • It runs a synthetic Monte Carlo stress test with no human participants or observational data: 15,000 candidate knowledge units at α = 1 scaling linearly to 45,000 at α = 3, with 30 independent replications per acceleration level and four comparator policies (blanket verification, production-first 20% sampling, static risk thresholding, and QAV).
  • Main outcomes are net value per task, escaped loss per task, severe escaped-error rate, backlog fraction, reviewed fraction, and ELR; robustness uses α from 1 to 3 and a 3×3×3 sensitivity grid at α = 3 with 15 replications per combination.

Limitations

  • The paper states that the M/M/1 queue is a stylized analytical device; real systems are multi-server, skill-constrained, priority-based, and often have bursty correlated arrivals.
  • The paper states that the overload function is assumed rather than empirically estimated, the simulation assumes separable task losses and does not model correlated errors, risk estimates use simple multiplicative noise, net value is in synthetic units, and verification capacity is treated as fixed within a period.
  • The paper states that results are model-conditional rather than empirical estimates and should not be read as forecasts of real productivity, error rates, or financial returns; it also notes the analysis focuses on error detection, not long-run human expertise.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.47760/oajmr.2026.v02i09.001

Cites

30 of 30 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 43 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    If AI acceleration increases verification arrival rate faster than service rate and the resulting load increases queue delay, backlog release, or quality degradation, reliable net throughput can decline even while gross output and total review activity rise.

    “Proposition 3, Verification Reversal.If AI acceleration increases λ V faster than μ and the resulting load increases queue delay, backlog release, or quality degradation, reliable net throughput can decline even while gross output and total review activity rise.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: In the analytical M/M/1 approximation, verification arrivals are Poisson and service times are exponentially distributed.
    • The authors state this as an assumption: Review detection probability declines above a load threshold according to a parsimonious overload function that is not an empirically estimated law.
  • Findingnone given

    When knowledge units differ in error probability times consequence severity, detection probability, and review demand, capacity allocated by expected avoided loss per service unit weakly dominates random allocation in the fractional relaxation if risk estimates have positive ranking information.

    “Proposition 4, Heterogeneity Advantage.When knowledge units differ in p i L i , detection probability, and review demand, capacity allocated by expected avoided loss per service unit weakly dominates random allocation in the fractional relaxation if risk estimates have positive ranking information.”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    If an organisation imposes either a strict delay limit or a minimum detection probability, the admissible arrival rate can be materially below mathematical service capacity.

    “If the organisation imposes either a strict delay limit or a minimum detection probability, the admissible arrival rate can be materially below mathematical service capacity.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: In the analytical M/M/1 approximation, verification arrivals are Poisson and service times are exponentially distributed.
    • The authors state this as an assumption: Review detection probability declines above a load threshold according to a parsimonious overload function that is not an empirically estimated law.
  • Findingown data

    A ten-percentage-point increase from an Epistemic Load Ratio of 0.80 to 0.90 more than doubles expected waiting in the stylised queue.

    “A ten-percentage-point increase from 0.80 to 0.90 therefore more than doubles expected waiting in this stylised queue.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    In the stable M/M/1 approximation, expected verification waiting time is strictly increasing and convex in the Epistemic Load Ratio.

    “Proposition 1, Queue Amplification.In the stable M/M/1 approximation, expected verification waiting time is strictly increasing and convex in ρ E .”
    • Also stated in: § B. Queue Amplification and Verification Backlog · plain
      “The key property is convexity.The second derivative of W q with respect to ρ E is positive for the stable region, so the marginal cost of additional load rises as utilisation approaches one.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: In the analytical M/M/1 approximation, verification arrivals are Poisson and service times are exponentially distributed.
  • May combine two claims. 34
  • Classified as probably not a substantial claim. 42

  • May combine two claims.

    Findingown data

    At acceleration level α = 3, blanket verification is the most overloaded policy, with mean Epistemic Load Ratio of 4.342 and a verification backlog equal to 79.7% of candidate units.

    “Blanket verification is the most overloaded policy, with mean ELR of 4.342 and a verification backlog equal to 79.7% of candidate units.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    At an Epistemic Load Ratio of 0.50, normalised expected queue delay equals 1; at 0.80 it equals 4; at 0.90 it equals 9.

    “At ρ E = 0.50, normalised expected queue delay μ W q equals 1.At ρ E = 0.80 it equals 4, and at ρ E = 0.90 it equals 9.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: In the analytical M/M/1 approximation, verification arrivals are Poisson and service times are exponentially distributed.

  • May combine two claims.

    Findingnone given

    Ranking by Marginal Verification Value is exactly optimal for the fractional relaxation and provides an interpretable greedy heuristic for the discrete problem.

    “Ranking by MVV is exactly optimal for the fractional relaxation and provides an interpretable greedy heuristic for the discrete problem.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    At the highest acceleration level, Queue-Aware Adaptive Verification maintained Epistemic Load Ratio near 0.80 with no verification backlog.

    “At the highest acceleration level, QAV maintained ELR near 0.80 with no verification backlog”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    At α = 1, the blanket policy can still prevent many errors because a large share of intended reviews are completed and overload degradation is modest.

    “At α = 1 it can still prevent many errors because a large share of intended reviews are completed and overload degradation is modest.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Production-first sampling is comparatively stable in net value because its intended coverage grows with volume at a fixed percentage, but this stability comes with weak risk targeting.

    “Production-first sampling is comparatively stable in net value because its intended coverage grows with volume at a fixed percentage, but this stability comes with weak risk targeting.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    QAV's escaped-loss advantage ranges from 0.002 to 0.468 units per task, and its severe-error-rate advantage ranges from 0.067 to 1.267 percentage points.

    “Its escaped-loss advantage ranges from 0.002 to 0.468 units per task, and its severe-error-rate advantage ranges from 0.067 to 1.267 percentage points.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    At α = 3, the severe escaped-error rate is 1.012% for QAV versus 1.620% for static thresholding and 2.339% for production-first sampling.

    “At α = 3 the difference is substantial: 1.012% for QAV versus 1.620% for static thresholding and 2.339% for production-first sampling.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    At α = 2 and above, the static rule crosses into sustained overload while QAV preserves its planned reserve.

    “At α = 2 and above, the static rule crosses into sustained overload while QAV preserves its planned reserve.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Monte Carlo 95% intervals quantify simulation uncertainty only.

    “Monte Carlo 95% intervals quantify simulation uncertainty only.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    The numerical analysis is a mechanism test, not an empirical estimate of any specific organisation.

    “The numerical analysis is a mechanism test, not an empirical estimate of any specific organisation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The NIST Generative AI Profile extends this logic to generative-AI-specific risks.

    “the NIST Generative AI Profile extends this logic to generative-AI-specific risks [23]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Queue management for verification is a governance problem, not merely an operations problem.

    “This makes queue management a governance problem, not merely an operations problem.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

    Premisecitation

    The Cognitive Sovereignty Threshold examines whether users retain sufficient capability to understand, verify, contest, remember, and take responsibility for AI-mediated work.

    “The Cognitive Sovereignty Threshold examines whether users retain sufficient capability to understand, verify, contest, remember, and take responsibility for AI-mediated work [28]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Human Agency under Algorithmic Governance and the CLEAR2 cognitive sovereignty framework focus on contestability, responsibility, and preserved agency.

    “Human Agency under Algorithmic Governance [29] and the CLEAR² cognitive sovereignty framework [30] focus on contestability, responsibility, and preserved agency.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [29]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [30]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    Faster generation increases not only useful throughput, but also the volume of material that could require epistemic checking.

    “Faster generation therefore increases not only useful throughput, but also the volume of material that could require epistemic checking.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Generative AI use can shift critical-thinking effort, and higher confidence in AI is associated with less critical engagement.

    “Recent research with knowledge workers indicates that generative AI use can shift critical-thinking effort and that higher confidence in AI is associated with less critical engagement [3]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Study design and conductnone given

    In the discrete setting, the Marginal Verification Value heuristic is not guaranteed to be globally optimal, but it provides a transparent baseline against which exact or learned policies can be evaluated.

    “In the discrete setting the MVV heuristic is not guaranteed to be globally optimal, but it provides a transparent baseline against which exact or learned policies can be evaluated.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Numerical differences in the simulation should be interpreted as conditional properties of the stated model, not as forecasts of real productivity, error rates, or financial returns.

    “Numerical differences should therefore be interpreted as conditional properties of the stated model, not as forecasts of real productivity, error rates, or financial returns.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Reversibility-Aware Staged Delegation allocates consequential actions according to recoverability and non-recoverable exposure.

    “Reversibility-Aware Staged Delegation allocates consequential actions according to recoverability and non-recoverable exposure [26]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisereasoning

    AI increases the arrival rate of verification-demanding knowledge units, whereas expert verification remains constrained by finite time, attention, domain competence, and institutional authority.

    “AI increases the arrival rate of verification-demanding knowledge units, whereas expert verification remains constrained by finite time, attention, domain competence, and institutional authority.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Workload imbalance, overtime, and corner-cutting can create persistent quality erosion.

    “Oliva and Sterman [20] model how workload imbalance, overtime, and corner-cutting can create persistent quality erosion.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [20]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    AI assistance can improve individual creative performance while simultaneously reducing collective diversity of outputs.

    “AI assistance can improve individual creative performance while simultaneously reducing collective diversity of outputs [8]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [8]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    All values in the numerical analysis are synthetic, and no human participants, confidential corporate data, or observational datasets are used.

    “All values are synthetic.No human participants, confidential corporate data, or observational datasets are used.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Premisereasoning

    What is missing is a unified model that treats verification demand as an endogenous consequence of AI acceleration, makes capacity saturation observable, and links prioritisation to expected epistemic harm per unit of scarce review effort.

    “What is missing is a unified model that treats verification demand as an endogenous consequence of AI acceleration, makes capacity saturation observable, and links prioritisation to expected epistemic harm per unit of scarce review effort.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The European Union Artificial Intelligence Act creates binding obligations for specified AI systems and places substantial emphasis on risk management, documentation, monitoring, and human oversight.

    “The European Union Artificial Intelligence Act creates binding obligations for specified AI systems and places substantial emphasis on risk management, documentation, monitoring, and human oversight [24]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The NIST AI Risk Management Framework emphasises governance, mapping, measurement, and management across the AI lifecycle.

    “The NIST AI Risk Management Framework emphasises governance, mapping, measurement, and management across the AI lifecycle [22]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisereasoning

    Generative AI can increase the rate at which knowledge work is produced, but it does not proportionally increase the human capacity to verify claims, assumptions, sources, calculations, and consequential recommendations.

    “Generative artificial intelligence can increase the rate at which knowledge work is produced, but it does not proportionally increase the human capacity to verify claims, assumptions, sources, calculations, and consequential recommendations.”
    • Also stated in: § INTRODUCTION · plain
      “Verification does not scale automatically with generation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    A low-consequence workflow can tolerate longer verification delay than a time-sensitive regulated decision, so the relevant verification capacity frontier is task-class specific.

    “Conversely, a low-consequence workflow can tolerate longer verification delay than a time-sensitive regulated decision.The relevant frontier is therefore task-class specific.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Runtime assurance treats agent actions as objects of policy-gated control.

    “Runtime assurance treats agent actions as objects of policy-gated control [25]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Coordination Compression Trap models broader organisational verification debt and resilience under AI-enabled acceleration.

    “The Coordination Compression Trap models broader organisational verification debt and resilience under AI-enabled acceleration [27]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The same structural parameters are used across policies within a replication so that differences arise from allocation rules rather than from different task populations.

    “The same structural parameters are used across policies within a replication so that differences arise from allocation rules rather than from different task populations.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    The purpose of the robustness analysis is not statistical generalisation but to test whether the main ordering depends on a single favourable calibration.

    “The purpose is not statistical generalisation.It is to test whether the main ordering depends on a single favourable calibration.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    A verification queue differs from an ordinary service queue in that service quality changes the epistemic state of the item.

    “A verification queue differs from an ordinary service queue in one important respect: service quality changes the epistemic state of the item.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Controlled evidence shows generative AI can reduce completion time and improve average task quality for professional writing activities.

    “Controlled evidence shows that generative AI can reduce completion time and improve average task quality for professional writing activities [1]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    In AI-assisted knowledge work, reviewers may respond to rising verification demand by compressing review, while managers interpret faster completion as evidence that the system is coping.

    “Applied to AI-assisted knowledge work, the danger is that reviewers respond to rising verification demand by compressing review, while managers interpret faster completion as evidence that the system is coping.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The overload function in Equation (6) is assumed rather than empirically estimated.

    “Second, the overload function in Equation ( 6) is assumed rather than empirically estimated.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Task-level speed gains are valuable but can trigger a downstream scarcity that task-level experiments do not measure.

    “Task-level speed gains [1], [2] are valuable, but they can trigger a downstream scarcity that task-level experiments do not measure.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [1]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [2]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The simulation assumes separable task losses and does not model correlated errors from shared models, retrieval sources, templates, or organisational assumptions.

    “Third, the simulation assumes separable task losses and does not model correlated errors from shared models, retrieval sources, templates, or organisational assumptions.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Capacity reserve is not idle waste; it is option value.

    “Capacity reserve is not idle waste; it is option value.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Correlated failures can make risk concentration more important than item-level expected loss.

    “Correlated failures can make risk concentration more important than item-level expected loss.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    Other teams experience rework, review fatigue, citation checking, inconsistency, and managerial unease because the generated artefacts are difficult to verify at scale.

    “Others experience rework, review fatigue, citation checking, inconsistency, and managerial unease because the generated artefacts are difficult to verify at scale.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The policy advantage appears in both economic and tail-risk metrics once the system becomes capacity constrained.

    “The policy advantage therefore appears in both economic and tail-risk metrics once the system becomes capacity constrained.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    AI productivity should be measured as reliable system throughput rather than gross generation.

    “The main theoretical implication is that AI productivity should be measured as reliable system throughput rather than gross generation.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The model treats verification capacity as fixed within a period.

    “Sixth, the model treats verification capacity as fixed within a period.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Blanket oversight can become infeasible, fixed-percentage review can eventually saturate, and static risk thresholds can fail when they ignore aggregate service demand.

    “Blanket oversight can become infeasible, fixed-percentage review can eventually saturate, and static risk thresholds can fail when they ignore aggregate service demand.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Without verification-capacity quantities, productivity projections can externalise verification costs onto employees, customers, auditors, or downstream decision makers.

    “Without these quantities, productivity projections can externalise verification costs onto employees, customers, auditors, or downstream decision makers.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    These results are model-conditional rather than empirical estimates.

    “These results are modelconditional rather than empirical estimates.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The functional form used for overload degradation is not claimed as an empirically estimated law.

    “The functional form is not claimed as an empirically estimated law.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    QAV should be understood as a capability-preserving design.

    “QAV should therefore be understood as a capability-preserving design.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    This study uses only synthetic data and publicly available literature and therefore involves no human participants or personal data.

    “This study uses only synthetic data and publicly available literature and therefore involves no human participants or personal data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The M/M/1 queue used in the model is a stylised analytical device.

    “First, the M/M/1 queue is a stylised analytical device.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Real verification systems are multi-server, skill-constrained, priority-based, and often have bursty correlated arrivals.

  • Classified as probably not a substantial claim.

    Premisecitation

    Human-agency frameworks emphasise contestability and institutional reversibility.

    “Similarly, human-agency frameworks emphasise contestability and institutional reversibility [29], [30].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [29]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [30]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    Net value is expressed in synthetic units and embeds a particular backlog penalty and review cost, so the numerical policy ranking is not a transferable business case.

    “Fifth, net value is expressed in synthetic units and embeds a particular backlog penalty and review cost.The numerical policy ranking is therefore not a transferable business case.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Queue-aware adaptation is most valuable when workload scales, risk is heterogeneous, and verification quality deteriorates under pressure.

    “Queueaware adaptation is most valuable when workload scales, risk is heterogeneous, and verification quality deteriorates under pressure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Generative AI can make knowledge production abundant while leaving verification scarce.

    “Generative AI can make knowledge production abundant while leaving verification scarce.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The analysis focuses on whether errors are detected, not on whether humans retain long-run expertise.

    “Seventh, the analysis focuses on whether errors are detected, not on whether humans retain long-run expertise.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    In practice, error probabilities and consequence magnitudes are estimated, not known.

    “In practice, p i and L i are estimated, not known.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    The Cognitive Sovereignty Threshold asks whether users remain able to understand, verify, contest, remember, and take responsibility for AI-mediated work.

    “The Cognitive Sovereignty Threshold [28] asks whether users remain able to understand, verify, contest, remember, and take responsibility for AI-mediated work.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Interpretationcitation

    ELR, backlog age, review completion, defect escape, and calibration error can be treated as measurable operational indicators within the NIST AI RMF lifecycle.

    “ELR, backlog age, review completion, defect escape, and calibration error can be treated as measurable operational indicators within that lifecycle.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [22]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [23]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    When the verification system is near capacity, production acceleration can produce nonlinear delay and weaker review.

    “When it is near capacity, the same acceleration can produce nonlinear delay and weaker review.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The NIST AI RMF and Generative AI Profile emphasise continuous mapping, measurement, management, and governance.

    “The NIST AI RMF and Generative AI Profile emphasise continuous mapping, measurement, management, and governance [22], [23].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [22]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [23]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The EU AI Act reinforces the need to make human oversight effective rather than merely nominal in contexts where its provisions apply.

    “The European Union AI Act likewise reinforces the need to make human oversight effective rather than merely nominal in contexts where its provisions apply [24].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Premisecitation

    The overload function is a stress-test mechanism consistent with evidence that workload and monitoring demands affect human performance.

    “It is a stress-test mechanism consistent with evidence that workload and monitoring demands affect human performance [10], [16], [20].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [10]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [16]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [20]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The worst overload mode is silent compression of review time because it preserves apparent throughput while degrading the control itself.

    “The worst overload mode is silent compression of review time because it preserves apparent throughput while degrading the control itself.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationcitation

    Repeated reliance on AI may reduce the competence needed for effective verification, while well-designed review may instead create learning.

    “Repeated reliance on AI may reduce the very competence needed for effective verification, while well-designed review may instead create learning.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [21]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The simulation design, parameterisation, and aggregation procedure are fully described in the paper.

    “The simulation design, parameterisation, and aggregation procedure are fully described in the paper; code and generated synthetic data can be provided by the author upon reasonable request.”
    Quote in the paperNot a substantial claim
67 of 67

  • The authors state this as an assumption.

    Only a subset of AI outputs require human review; a subset is routed to verification according to organisational policy, automated checks, risk scoring, provenance requirements, or human judgement.

    “The model does not assume that all AI outputs require human review.Instead, it assumes that a subset is routed to verification according to organisational policy, automated checks, risk scoring, provenance requirements, or human judgement.”

  • The authors state this as an assumption.

    In the analytical M/M/1 approximation, verification arrivals are Poisson and service times are exponentially distributed.

    “consider an M/M/1 approximation with Poisson verification arrivals and exponentially distributed service times.This is intentionally stylised.”

  • The authors state this as an assumption.

    Review detection probability declines above a load threshold according to a parsimonious overload function that is not an empirically estimated law.

    “A parsimonious overload function is ( ) * , ( )-+ (6) where β > 0 controls the severity of overload degradation and ρ 0 is the load above which quality begins to decline.The functional form is not claimed as an empirically estimated law.”

  • The paper relies on this without support.

    Real verification systems are multi-server, skill-constrained, priority-based, and often have bursty correlated arrivals.

    “Real verification systems are multi-server, skill-constrained, priority-based, and often have bursty correlated arrivals.”

  • The paper relies on this without support.

    Real triage systems may be systematically biased by domain, language, user status, or the same model-family errors they are intended to detect.

    “Real triage systems may be systematically biased by domain, language, user status, or the same model-family errors they are intended to detect.”

  • The paper relies on this without support.

    Risk-based triage is already common in audit, cybersecurity, medicine, and quality control.

    “Risk-based triage is already common in audit, cybersecurity, medicine, and quality control.”

  • The paper relies on this without support.

    Human oversight is often treated as a qualitative safeguard, with the assumption that having a person in the loop makes the system more controlled.

    “Human oversight is often treated as a qualitative safeguard: if a person is in the loop, the system is assumed to be more controlled.”
7 of 7

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Runtime-assurance citation points to the generative AI profile

    Minor

    The text cites [25] for Tan's runtime-assurance work, but [25] is the NIST Generative AI Profile. Tan's runtime-assurance paper is listed as [27].

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. L. Kleinrock, Queueing Systems, Volume 1: Theory. New York, NY, USA: Wiley-Interscience, 1975.
  2. European Parliament and Council of the European Union, "Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)," Official Journal of the European Union, 12 July 2024.

Cited by

Sorted by citations

Related

Authors
TAN, K. H.
Affiliations
Singapore University of Social Sciences (SG)
Corresponding author
TAN, Kwan Hong
ORCID
TAN, Kwan Hong
Venue
Open Access Journal of Multidisciplinary Research
Publisher
Zain Publications
Published
2026-09-30
Volume
2
Issue
9
Pages
1-17
DOI
10.47760/oajmr.2026.v02i09.001
OpenAlex
W7207677819
Cited by
1
References
28
Citations by year
2026: 1
FWCI
3.86
Type
journal article
Language
en
Open access
Licence
other-oa
Topics
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Social Sciences › Psychology › Social Psychology › Human-Automation Interaction and Safety
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
Keywords
generative AI, knowledge work, human oversight, review quality, knowledge reliability, AI productivity, socio-technical systems

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-11-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-06.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review