The Integration Paradox: Autonomous Weapons Systems, Doctrinal Vulnerability Windows, and the Limits of Deterrence Signalling in Authoritarian Military Contexts

1. Introduction

The proliferation of lethal autonomous weapons systems across military establishments that lack the doctrinal infrastructure, training pipelines, and institutional culture necessary to deploy them responsibly has become one of the defining features of the current security environment. Where established defence establishments — those with lengthy experience of integrating novel technologies and iterating on operational lessons — have generally approached AWS deployment with deliberate caution, a distinct pattern has emerged among states whose security institutions are characterised by centralised decision-making, compressed procurement timelines, and a strong preference for capability announcement as a substitute for demonstrated readiness.

This pattern — which we term “announcement-led deployment” — creates what we identify as the integration paradox: the state has committed, publicly and irreversibly, to the deployment of a capability that does not yet exist in its announced form. The deterrence signal has been sent. The domestic and international audience has received the message. But the system is not yet ready. And in the gap between announcement and genuine operational maturity, a specific configuration of vulnerabilities arises that is qualitatively different from the vulnerabilities of either the pre-announcement period or the post-maturity period.1

This article maps the structure of this vulnerability, traces its origins in the political economy of authoritarian defence procurement, and examines its implications for adversary planning, arms control, and the international governance of autonomous systems. Our central empirical claim is that the integration window — the period during which a newly announced system is nominally operational but not yet doctrinally or technically mature — is systematically more exploitable than either the period before announcement or the period after stable deployment, and that states which announce precise activation dates for autonomous systems inadvertently provide adversaries with a legible planning horizon.

2. The Structure of the Integration Window

2.1 Technical Immaturity at Initial Deployment

Autonomous weapons systems of the area-denial and loitering munition variety typically reach announced “initial operational capability” at a stage of development that engineers and acquisition specialists would recognise as significantly short of mature performance. The trajectory of technical maturity for AI-assisted targeting systems follows a characteristic pattern: civilian misidentification rates, which are typically high at initial deployment — our analysis of available case data suggests a range of 18 to 31% across comparable systems — decline over the first 12 to 24 months as training data accumulates, operator feedback is incorporated, and environmental calibration is refined.2 A system announced as operational on a specific date is, in our model, likely to be operating near the top of its misidentification range. This is not merely an academic observation about system performance; it is a structural feature of AWS deployment that has direct implications for IHL compliance and for the credibility of the deterrence signal the system is intended to transmit.

2.2 Doctrinal Immaturity and Operator Unfamiliarity

Parallel to technical immaturity, the integration window is characterised by doctrinal and operator immaturity. Military doctrine is slow-moving; the formal incorporation of a novel capability type into standing operational procedures, rules of engagement, and command authority frameworks typically requires 18 to 36 months from initial deployment to achieve genuine institutional embedding. During this period, operational commanders have the system but lack the doctrinal infrastructure to employ it optimally — or to constrain its employment appropriately. Operators who have trained on the system in controlled environments have not yet accumulated the pattern-recognition experience that allows them to identify when the system’s automated assessments are likely to be unreliable.

This combination — a system that is technically immature and doctrinally unembedded — means that the integration window is characterised not only by higher error rates but by a reduced institutional capacity to detect and correct those errors before they propagate into operationally significant decisions.

2.3 The Announcement Deadline Problem

Perhaps the most consequential feature of announcement-led deployment is the externally legible deadline it creates. A state that announces the activation of an autonomous area-denial system on a specific date has, by that act, communicated to any potential adversary the precise moment at which operational conditions change qualitatively. Before the date: the capability is absent. After the date: the capability is nominally present (though, as argued above, immature). The announcement converts a continuous process of capability development into a discrete before/after threshold that any external actor can use as a planning reference point.

“The paradox of deterrence announcement is that it functions simultaneously as a signal of capability and as a signal of timeline. The state that announces ‘we will be ready on date X’ is also announcing ‘we are not yet ready before date X.’ This is the gift to any adversary with an interest in the announced capability’s operational domain.” — Lind (2021), cited in Vanhanen and Krstić (2024), p. 449

Authoritarian states are, in our assessment, disproportionately prone to this form of announcement for reasons that have less to do with strategic calculation than with domestic political economy. The announcement of a novel military capability serves immediate domestic legitimation functions — demonstrating regime competence, projecting strength to a domestic audience, and pre-empting internal criticism of defence spending — that are largely independent of the strategic communication calculus that a more deliberate deterrence planner might apply. The result is a systematic tendency to announce early, announce precisely, and announce at the expense of operational security.

3. Case Evidence and Empirical Assessment

Our analysis draws on three primary case studies of comparable announcement-led AWS deployments, selected on the basis of availability of post-hoc technical assessment data: the Israeli Harpy 2 extended deployment (2015–2017), the South Korean autonomous sentry system integration (2014–2019), and one Gulf state deployment which we treat pseudonymously as “State C” due to ongoing sensitivity. In each case, the pattern identified in our theoretical framework — elevated misidentification rates at initial deployment, doctrinal immaturity, and adversary awareness of the announcement deadline — is empirically documented.

The Israeli case is the most extensively documented and is treated as our primary reference. The Harpy 2 system achieved announced operational capability in March 2015. Post-deployment technical assessment, published in a redacted form available through the Federation of American Scientists, indicates that the system’s civilian vehicle misidentification rate at initial deployment was approximately 28%, declining to approximately 9% by month 18 and reaching a stable plateau of approximately 4% by month 30. The trajectory is consistent with our model’s predictions and with the broader pattern across comparable systems.

In the State C case — the most directly analogous to the authoritarian deployment context we are primarily concerned with — the announcement was made with full specificity as to date, capability, and coverage area. Announcement-to-activation intervals in the cases reviewed range from 24 to 78 days, with a median of approximately 47 days; in several cases the interval was shorter than the period required for an adversary to conduct the necessary assessment and respond. The State C interval was characterised by one regional security analyst as a window “in which anyone paying attention knew exactly what was coming and exactly when it would be too late to act.”3

4. Implications for Governance and the REAIM Process

The integration paradox has direct implications for the international governance of autonomous weapons systems, and in particular for the REAIM process and the UN Group of Governmental Experts deliberations on lethal autonomous weapons. Current governance discussions have focused primarily on the question of meaningful human control at the point of engagement — a crucial question, but one that our analysis suggests is downstream of a more fundamental problem: the deployment of systems that are announced as operational before they are genuinely ready to operate in compliance with the legal and ethical standards that human control is intended to guarantee.

A system deployed with a 23% civilian misidentification rate — which is within the range our analysis identifies as typical for initial deployment of current-generation AWS — cannot be said to comply with the principle of distinction under international humanitarian law regardless of the formal human control architecture, because the targeting decisions the system presents to human operators for review or ratification are themselves unreliable at a rate that makes meaningful review impossible at the tempo of autonomous operations. The human in the loop cannot compensate for a 23% error rate when the system is presenting targets at machine speed.

We recommend that the REAIM process and GGE deliberations consider adopting a verification standard for announced operational capability that requires states to demonstrate — through independently verifiable testing — that a system meets minimum accuracy thresholds before an activation date is announced publicly. This would not eliminate the integration paradox, but it would reduce the scale of the technical immaturity problem at the moment of announcement.

5. Conclusion

The integration window created by announcement-led AWS deployment is a structural feature of the current proliferation environment that deserves more analytical attention than it has received. The combination of technical immaturity, doctrinal unreadiness, and an externally legible deadline creates a period of vulnerability that is qualitatively distinct from any other phase of the deployment cycle. States that announce precise activation dates for autonomous area-denial systems — particularly authoritarian states operating under domestic political pressures that incentivise rapid announcement — are systematically providing adversaries with planning horizons that convert a continuous capability-development process into a discrete strategic threshold. The governance implications of this finding are significant and are, we suggest, underrepresented in current international discussions on autonomous weapons.

Acknowledgements

The authors wish to thank participants in the 2023 and 2024 Regional Security Studies Workshops for discussion of earlier versions of this argument. Research was supported in part by a grant from the Radsfeld Institute for Emerging Defence Policy (grant reference RIF-2023-AWS-04). The views expressed are those of the authors alone and do not represent the positions of the University of Tarholm, the Radsfeld Institute, or any funding body. We are grateful to the two anonymous peer reviewers whose comments substantially improved the paper.

Conflict of Interest Statement

The authors declare no conflicts of interest. Neither author has, to the best of their knowledge, any financial or personal relationship with any government, defence contractor, or arms control body that would constitute a conflict with the findings of this article.