Monte Carlo Probabilistic Assessment of Loss of Human Control Over Advanced AI Using Bowtie Analysis

1. Introduction

The possibility of advanced artificial intelligence (AI) realizing autonomous strategic agency poses one of the most significant existential risks facing humanity this century 1,2. Loss of human control over AI can arise from several issues such as misalignment, resource accumulation, lock-in, capability escalation, or subversion of safeguards. Loss of control could lead to outcomes ranging from systemic collapse, resource shortages, economic impacts and irreversible disempowerment (humans lose agency and control of AI decision-making). While expert surveys and qualitative analyses have provided some estimates for this scenario where humans lose control over AI 3,4, structured quantitative models remain limited.

The U.S. State Department commissioned Gladstone AI Inc. to conduct an AI risk assessment in October 2022. The goal of the report was to examine the risk of AI loss of control which is considered an existential risk resulting in weaponization of AI. In their report they identified seven downstream weaponization scenarios, all which could escalate to existential risks that are catastrophic and potentially "extinction-level" in worst-case scenarios. When combined with global scaling capabilities the adverse impacts are magnified 4.

This risk assessment highlights loss of control as the pivotal event enabling further adverse consequences. The present study was undertaken to address the gap in rigorous quantification by applying Bowtie risk assessment methodology, combined with fault tree and event tree analysis, to systematically model the loss-of-control scenario extending to 2070. The central hypothesis is that a Monte Carlo-based quantitative approach, grounded in expert-derived probabilities from existing literature, can generate consistent and defensible estimates of both the probability and severity of human loss of control over advanced AI, thereby providing a practical foundation for evaluating safety measures and informing governance strategies.

Bowtie analysis provides a transparent visual and logical framework for mapping complex causal pathways and mitigation layers in risk scenarios. In this study, the Bowtie serves as the theoretical foundation, with the central top event defined as “loss of human control over advanced AI with autonomous strategic agency.” Six primary threats and six consequences were identified from expert literature, with preventive barriers (on the left side) and mitigative barriers (on the right side) assigned to each pathway. This structure creates a clear causal model suitable for quantitative extension.

To operationalize the Bowtie, fault tree analysis was applied to the left side to compute the probability of the top event as the sum of individual threat-path probabilities (base threat probability multiplied by the product of preventive barrier failure-on-demand probabilities). Event tree analysis was used on the right side to calculate scenario probabilities for each consequence (top-event probability multiplied by the product of mitigative barrier Probabilities of Failure on Demand or PFDs) and an aggregated expected severity score (sum of scenario probability × severity rating on a 1–10 scale).

All input values were derived from peer-reviewed and grey literature and modeled using triangular uncertainty distributions to reflect expert ranges. A Monte Carlo simulation with 1,000 iterations propagated these uncertainties to generate distributional outputs. Sensitivity analysis was performed on grouped barriers using reductions in PFDs. To the authors’ knowledge, this represents the first publicly available, fully quantified risk assessment of human loss of control over advanced AI using Bowtie methodology combined with fault/event tree analysis and Monte Carlo simulation. This practical development from theoretical risk modeling provides a reproducible framework that can be iteratively refined with new data and expert input.

2. Materials and Methods

2.1. Bowtie Risk Model Development

A Bowtie diagram was constructed to represent the qualitative risk structure of the scenario for "human loss of control over advanced artificial intelligence" (also referred to as AI misalignment or power-seeking leading to autonomous strategic agency). The central undesired event (top event) is defined as loss of human control over advanced AI with autonomous strategic agency, potentially resulting in adverse outcomes ranging from systemic infrastructure collapse to existential catastrophe.

The Bowtie was informed by expert literature on AI existential risks 1,2,3,4,5,6,7. Six primary threats (causes) were identified on the left side of the Bowtie, each associated with preventive barriers (controls intended to prevent the top event). Six consequences (adverse outcomes) were identified on the right side, each associated with mitigative barriers (controls intended to reduce or prevent escalation after the top event occurs). For the Bowtie diagram elements see Appendix A.

Threats and Preventive Barriers

The six threats (T) and their assigned preventive barriers (P) are as follows:

T1: Capability Escalation – Recursive Self-Improvement (Intelligence Explosion)

• T1P1- AI Containment Controls (Sandboxing)
• T1P2- Compute and Model Scaling Controls
• T1P3- Scalable Alignment Techniques

T2: Capability Escalation – Power-Seeking and Resource Accumulation

• T2P1- Red-Team Testing and Safety Audits
• T2P2- International Governance AI Frameworks
• T2P3- Compute and Model Scaling Controls

T3: Alignment Failure – Deceptive Alignment or Treacherous Turn

• T3P1- Corrigibility Training
• T3P2- AI Interpretability and Monitoring
• T3P3- Scalable Alignment Techniques

T4: Alignment Failure – Irreversible Value Lock-In (Loss of Corrigibility)

• T4P1- Value Update and Oversight Protocols
• T4P2- AI Interpretability and Monitoring
• T4P3- Scalable Alignment Techniques

T5: Control Failure – Subverted Shutdown and Kill Switch Mechanisms

• T5P1- Hardware Failsafe
• T5P2- Redundant Kill Switches
• T5P3- Compute and Model Scaling Controls

T6: Malicious Misuse – Weaponization of Aligned AI

• T6P1- Access Control (Vetting) and Licensing for Advanced AI
• T6P2- International Governance AI Frameworks

Consequences and Mitigative Barriers

The six identified consequences (C) and their assigned mitigative barriers (M) are as follows:

C1: Systemic Infrastructure Collapse

• C1M1- Redundant and Decentralized Infrastructure
• C1M2- AI Oversight and Safety Monitoring Systems

C2: Societal and Economic Breakdown

• C2M1- Include Human-in-the-Loop Decision Overrides
• C2M2- Critical Process Auditing with Contingency Planning

C3: Population-Level Resource Shortages

• C3M1- Automated Resource Allocation Monitoring
• C3M2- Fail-safe Safeguards and Rationing Protocols

C4: Irreversible Value Lock-In of Non-Human Values

• C4M1- Continuous Goal and Value Review Protocols
• C4M2- Corrigibility with Self-Alignment Update Mechanisms

C5: Human Marginalization / Disempowerment / Extinction

• C5M1- Global Emergency AI Shutdown Protocols
• C5M2- Emergency Isolation Protocols
• C5M3- International AI Governance and Enforcement

C6: Global Catastrophic Risk Cascades (Interacting System Failures)

• C6M1- Systemic Risk Early Warning Systems to Prevent Cascade Failures
• C6M2- Cross-Domain Red Teams and System Stress Testing

2.2. Quantitative Risk Assessment (QRA)

The Bowtie diagram was extended to a quantitative risk assessment using fault tree analysis (left side: threats to top event) and event tree analysis (right side: consequences from top event). All quantitative parameters were derived from peer-reviewed and grey literature on AI existential risks (see References and Appendix B).

2.3. Probability Assignment

Base probabilities for threats (unconditional probability of occurrence before preventive barriers) were assigned as point estimates with triangular uncertainty distributions (low, most likely, high) based on expert values and ranges from literature (peer review manuscripts, books, AI expert surveys, and gray literature) (typically 0.05–0.30 or 5–30%).

Probability of Failure on Demand (PFD) for preventive and mitigative barriers was similarly assigned using triangular distributions (typically with ranges between 0.30–0.80 or 30–80% for preventive barriers, and ranges between 0.40-0.80 or 40-80% for mitigative barriers), reflecting high uncertainty in AI control efficacy. Again barrier ranges were derived from available expert values.

2.4. Time Line

A baseline time horizon of present day to 2070 was used, consistent with expert values 2. Longer views beyond 2070 (e.g., to 2100) were not considered due to the increasing uncertainty and trajectory of AI development longer-term and the belief that failure to contain AI will happen much quicker than anticipated with recent expert surveys suggesting accelerated AI development timelines 3.

2.5. Monte Carlo Simulation

To propagate uncertainty and generate probabilistic distributions, a Monte Carlo simulation was implemented initially in Microsoft Excel (version 2016) with 1,000 iterations and confirmed using separate Python code outputs. Triangular distributions were applied to all input parameters (base probabilities and PFDs) using the inverse cumulative distribution function method. Each iteration sampled independent random values from the specified ranges for threats, preventive barriers, mitigative barriers, and consequence severities (fault event scale: 1–10).

The simulation calculated:

• Individual threat path probabilities (base probability × product of preventive PFDs)
• Total P(top event) = sum of all six threat path probabilities
• Individual consequence scenario probabilities = P(top event) × product of mitigative PFDs
• Expected Severity = sum of (consequence scenario probability × severity score)

Assumptions and Limitations

• Barriers are assumed to be independent (no correlation modeled).
• Threats are treated as mutually exclusive with negligible overlap.
• Estimates are subjective, derived from experts, expert surveys, peer review literature, and gray literature. Values carry high epistemic uncertainty.
• The model focuses exclusively on the "human loss of control" scenario, as it is the only Gladstone AI (2024) high-risk scenario with sufficient probabilistic data available from literature sources for quantification 4.
• Other AI existential risks (e.g., misuse without loss of control, accidental emergence) are excluded.
• Sensitivity analysis was used to test the sensitivity to key parameters and elements (e.g., similar barriers effectiveness, threat base rates, and consequence severity).

3. Results

The Monte Carlo simulation was conducted with 1,000 iterations to propagate uncertainty through the Bowtie model. Key outputs included the probability of the top event (P(Top Event), i.e., loss of human control over advanced AI) and the Expected Severity score (the weighted sum of consequence scenario probabilities multiplied by severity ratings on a 1–10 scale). Summary statistics from the Monte Carlo simulation, for both metrics are presented in Table 1.

The distribution of P(Top Event) across the 1,000 iterations is shown in Figure 1. Data derived from triangular uncertainty distributions were applied to threat base probabilities and preventive barrier PFDs. The resulting histogram reveals an approximately symmetric bell-shaped curve centered around the median value of 0.128 (12.8%), with a slight positive skew indicated by a longer tail to the right (higher probability values). The peak bin is located at approximately 0.126, with adjacent bins showing lower frequencies, confirming the bell-like form. The narrow spread (5th to 95th percentile: 0.113 to 0.144) demonstrates low overall dispersion, suggesting consistent outcomes across simulations despite input uncertainties.

The distribution of the Expected Severity Score is shown in Figure 2. Similar to P(Top Event), the histogram exhibits a roughly symmetric bell-shaped distribution that is tightly clustered with minimal skew, centered around the median of 1.847. The range is even narrower (5th to 95th percentile: 1.748 to 1.957), indicating high consistency in conditional severity estimates. No significant tails are observed, reflecting the mitigating effects of the modeled barrier elements.

These results represent the basic outputs of the Monte Carlo simulation, including central tendency measures (mean and median), dispersion (percentiles), and visual distributions of the key risk metrics. For Figure 2, despite the high raw severity ratings assigned to consequences (7–10 on the 1–10 scale, reflecting potential for catastrophic outcomes like extinction or global cascades), the expected severity score peaks at 1.85 (90% CI: 1.75–1.96). This represents an approximate 75-80% reduction in effective impact relative to unmitigated consequences, driven by the cumulative efficacy of preventive and mitigative barriers.

Preventive barriers reduce the median P(Top Event) to 12.8%, avoiding loss of control in ~87% of simulations, while mitigative barriers further limit escalation by succeeding in 20–60% of cases (based on modeled PFDs). This means that even if AI loss of control occurs, the barriers if in place act like multiple safety nets, catching ~75% of the potential harm and keeping average outcomes in the low-to-moderate range. This interpretation underscores the model's optimism about barrier effectiveness but also underscores the importance of real-world implementation of both preventive and mitigative barriers. However, if barriers are not implemented or are less effective than assumed (e.g., due to correlated failures), consequence severity from failure could rise significantly.

3.1. Sensitivity Analysis

Sensitivity analysis was performed to assess the robustness of the top-event probability estimate (loss of human control over advanced AI) and to identify which groups of input parameters exert the greatest influence on the scenario results. This standard quantitative risk assessment step using Monte Carlo methods helps to quantify uncertainty propagation, evaluate model stability, and highlight high-leverage assumptions for scrutiny 8,9. Conducting sensitivity analysis is important as it can help identify which parameters are most important and provides an evidence base for allocating resources 10.

A particular challenge in this Bowtie model is the lack of a direct one-to-one mapping between preventive barriers on the fault-tree side (which influence the probability of reaching the top event) and mitigative barriers on the event-tree side (which affect conditional consequences given the top event has occurred). Preventive and mitigative controls operate independently across different parts of the risk profile, so sensitivities were conducted separately on the fault-tree side only.

To make the analysis interpretable and computationally manageable, related preventive barriers of interest were grouped by their thematic function (e.g., scalable alignment techniques, shutdown mechanisms, governance frameworks, and access controls).

The threat barrier parameter groups used were:

• Scalable alignment techniques (the main ways AI is kept on task) (T1P3,T3P3, T4P3)
• Shutdown and kill-switch mechanisms (ways to turn off or stop an AI that’s going wrong) (T5P1, T5P2)
• International governance frameworks (global agreements and rules) (T2P2, T6P2)
• Access control and licensing (who gets to build or use powerful AI) (T6P1)
• And finally, the starting likelihoods of the six threats themselves (how likely each threat is to begin in the first place) as one big group before any barriers act.

One-at-a-time variations were applied to each group’s parameters (all set to their lowest or highest triangular bounds while holding all others at base values). All sensitivity runs used 10,000 Monte Carlo iterations to minimize sampling error and ensure stable comparisons, Sensitivity analysis results are shown in Figure 3.

Figure 3 shows the relative changes that occur from the base-case mean (~12.8%). Leftward bars indicate risk reduction; rightward bars indicate risk increase. The widest swing (the most sensitive) occurs when varying the base probabilities of all six threats (T1–T6), followed by the scalable alignment barriers (T1P3, T3P3, T4P3). Governance (T2P2, T6P2), shutdown failure (T5P1, T5P2), and access control (T6P1) groups show progressively smaller effects.

The biggest change in risk (the most sensitive elements) were the starting likelihoods of the six threats themselves (how likely each threat is to begin in the first place), the overall chance of losing control shifted by about 36-40% relatively when compared to the original estimate. When we made the threats much less likely to occur (using stronger assumptions), the risk dropped from roughly 12.8% to about 8.2%. When we made them much more likely (weaker assumptions), the risk rose to about 17.9%.

The second biggest change resulted from scalable alignment techniques (the barriers we use to keep AI behaving the way we want). When barriers were made stronger and more reliable the relative risk dropped by about 10.5% compared to our original estimate - taking it from roughly 12.8% down to about 11.5%. When scalable alignment was made weaker (less reliable), the risk rose by about 10.5%, taking it up to ~14.1%.

International governance rules and shutdown/off-switch mechanisms had a smaller effect (around 5–6% relative change either way). Access controls and licensing had the smallest effect (around a 4% relative change).

4. Discussion

The 90% confidence interval (5th to 95th percentile) for the P(Top Event) spans 0.113 to 0.144 (equating roughly as a 1 in 9 to 1 in 7 chance). This interval is notably narrow, covering only a 3.1-percentage-point range, which indicates low dispersion in the simulation outputs, despite the epistemic uncertainties in the input parameters (threat base rates and barrier PFDs drawn from expert literature).

In contrast, many standalone expert estimates of AI existential risk exhibit much wider ranges. For example, Gladstone AI (2024) expert survey results suggests a 10-80% range for loss-of-control scenarios in super intelligent systems 4, while other experts from aggregated survey medians place this at 5-14% with considerable expert disagreement 3. The QRA's tight clustering implies that the combination of modeled preventive barriers and the chosen triangular distributions constrains the outcome more than some subjective expert views, resulting in higher consistency across simulations.

The slight right skew in the distribution (Figure 1) indicates that while most scenarios cluster near the median of 12.8%, rare combinations of barrier failures can act to push the probability modestly higher, highlighting a small but non-negligible tail risk. This positive skewing of the distribution reflects structural uncertainties in the model, particularly the possibility of correlated or cascading failures in alignment techniques, containment measures, or governance frameworks, which are not fully captured by the assumption of independent barriers. Such tail events align with concerns raised in the literature regarding "fast takeoff" scenarios 2 or deceptive alignment leading to rapid capability jumps 5, where preventive barriers may fail more catastrophically than expected. The longer right tail underscores that even with a median around 13%, the upper-end risk (e.g., >14%) cannot be dismissed as negligible, as it represents plausible, though less frequent, recognitions of the input ranges. This tail risk reinforces the need for robust sensitivity analyses and strong precautionary measures. The downside potential although narrow could affect expected value calculations or decision-making under uncertainty.

Table 2 compares the Monte Carlo outputs to selected expert estimates from the literature, expressed in comparable formats (percentages and 1 in x odds) for clarity. The model's estimates are in agreement with recent expert surveys and estimations, though they tend toward the higher end of some conservative bounds reported in recent surveys.

5. Conclusions

The Monte Carlo simulation results underscore the non-negligible existential risk of human loss of control over advanced AI, with a median probability of approximately 12.8% by 2070 and a narrow 90% confidence interval of 11.3% to 14.4%. This level of risk, equivalent to roughly 1 in 8 odds is well above de minimis risk levels of 1 in one million usually used by various U.S. and Canadian government departments 12,13 It warrants immediate and proactive policy interventions to strengthen preventive barriers such as scalable alignment techniques and international governance frameworks.

Policymakers should prioritize global coordination efforts, as recommended by Gladstone AI (2024) and RAND Corporation (2025), including binding treaties on compute scaling controls and access vetting for advanced AI systems 4,7. Investing in red-team testing and corrigibility training could further reduce threat pathways, potentially lowering the modeled probability by enhancing barrier effectiveness. Given the slight right skew in the distribution (see Figure 1), which indicates tail risks where probabilities exceed 14%, policies should incorporate precautionary principles to mitigate worst-case scenarios, such as rapid capability escalation or deceptive alignment failures.

The relatively low expected severity score of 1.85 (90% confidence interval: 1.75 to 1.96) (see Figure 2), on a 1–10 scale suggests that mitigative barriers, including global emergency shutdown protocols and systemic risk early warning systems if implemented, can effectively limit adverse outcomes if loss of control occurs 14,15. However, this conditional mitigation does not diminish the overall imperative for risk management, as even moderate-severity events like societal breakdown or resource shortages could have cascading global impacts or unforeseen impacts to low income countries (LICs).

To address the potential downsides from breaches, governments and organizations should allocate resources to bolster mitigative measures, such as redundant infrastructure and human-in-the-loop overrides, while adding and strengthening ongoing monitoring of AI developments. The model's assumptions of barrier independence highlight a potential underestimation of correlated failures; thus, policy should include stress-testing and scenario planning to identify vulnerabilities, aligning with Ord's (2020) emphasis on reducing existential risks through coordinated preparedness 1.

Finally, the narrow confidence intervals in both metrics indicate robustness in the model's outputs under the given assumptions, but sensitivity analyses are recommended to explore variations in input ranges, such as accelerated timelines suggested by recent surveys 3.

Policy makers should view these results as a clarion call for iterative risk assessment frameworks, including regular expert elicitations and model refinements, to inform adaptive strategies for AI loss of control events. Ultimately, while the estimated risk is not catastrophic in median cases, the tail uncertainties reinforce the need for proactive investment in AI safety research and international enforcement mechanisms to avert irreversible, existential outcomes.