Since John McCarthy first coined the term “artificial intelligence” (AI) at Dartmouth in 1955, AI has undergone continuous evolution, steadily expanding its influence across sectors including governance. For instance, in 2025, Albania appointed Diella, an AI system, as the Minister of State for Artificial Intelligence with “the goal of enhancing online services for citizens of all ages through voice, visuals, and speed”. In the same year, the United Arab Emirates announced plans to integrate AI into its legal framework, aiming to become the first country to leverage artificial intelligence not only in drafting new legislation but also in reviewing and amending existing laws.
The acceleration and adoption of AI in state bureaucracy has expanded over the past decade, and AI is now being integrated into e-governance, security, justice, and surveillance systems as well. By automating administrative processes and assisting through data-driven decision-making, AI systems have the potential to improve resource allocation, enhance predictive analysis, and support real-time threat detection. Yet, what are the opportunities and threats of AI within the realm of security? For clarity, this paper uses “AI” broadly to mean computational systems, from machine-learning-based pattern recognition and predictive analytics to natural-language processing and generative tools, that analyze data to inform, accelerate, or automate decisions, rather than any single technology, product, or vendor. These advances raise a definitional question this paper treats as central rather than incidental. Security Sector Governance (SSG) describes the reality of how a state’s security sector actually works, its institutions, actors and practices that exercise or oversee the use of force. Security Sector Reform (SSR) is the political and technical process of improving that governance: Making security provision, management and oversight more effective and more accountable, within a framework of democratic civilian control, rule of law and respect for human rights. AI is already being absorbed into SSG. Whether it can ever serve SSR is a separate, and far more demanding, question, one this paper answers largely in the negative, absent specific safeguards.
In fragile, authoritarian and conflict-affected settings, where institutional accountability and oversight mechanisms are already severely compromised, the unregulated deployment of AI into SSG does not merely introduce new risks or accelerate existing governance failures. Rather, it creates the precise conditions under which unguided AI integration becomes a force multiplier for the dynamics that reform seeks to overcome. Used well, however, AI can strengthen transparency and institutional effectiveness. This paper proceeds in three steps: It first maps how AI is already being absorbed into SSG; it then argues that this absorption is making the underlying work of SSR harder rather than easier; and it finally identifies the narrow conditions under which AI could nonetheless serve SSR objectives. Throughout, the argument is the same: AI can only serve SSR where its use is conditioned on three interlocking safeguards: Independent oversight mechanisms, accountability frameworks, and binding legal frameworks. Where these safeguards do not exist, AI should be treated not as a reform tool but as a reform risk, a risk to the very populations SSR is meant to protect, and one that practitioners, donors and international partners have a responsibility to name and address before engagement proceeds.
AI in Security Sector Governance: What Is Already Happening
The formal integration of AI into SSG processes remains limited or, at a minimum, not yet visible in the public record. However, the conditions under which such integration could expand are already troubling. For instance, state authority across authoritarian and conflict-affected settings remains fragmented, tightly controlled, weakened, or structurally incapable of meeting citizens’ basic needs. Where it cannot advance, the structural conditions enabling instability are left to compound. AI-enabled tools, nominally designed to improve efficiency and coordination within security institutions, risk producing the opposite effect when deployed in environments where oversight is weak or absent. Under authoritarian governments, rather than strengthening governance, these tools provide new instruments to consolidate control, suppress dissent and evade accountability, outcomes that deepen, rather than resolve, existing power asymmetries within SSG.
What already exists, for instance in the Middle East and North Africa, is a pattern of digital repression, spanning both longstanding surveillance and hacking tools and an expanding layer of AI-enabled analytics that reveals precisely how such tools function when introduced into SSG more broadly, and how these practices in turn undermine the pillars and even the possibility of SSR. ) Authoritarian governments across the region are already expanding digital surveillance infrastructure, increasingly layered with AI-enabled analytics used by their security institutions in ways that directly contradict the principles SSG claims for itself, such as civilian oversight, transparency and legal accountability, let alone SSR’s aspirations. These are not scattered, ad hoc uses. Two patterns occur across the region. The first is transnational digital repression that follows exiled critics across borders. Transnational repression itself is not new, but AI-assisted data processing changes its reach: Automated triage and cross-referencing of intercepted communications let a handful of operators track far more exiled critics, at once, than manual surveillance ever allowed.
The second is domestic: Content-based AI-enabled surveillance that polices what citizens say and do at home. For instance, commercial spyware such as Pegasus, deployed by the United Arab Emirates and Saudi Arabia, has been used to track and silence dissidents both at home and abroad, with activists having their devices compromised and facing detention, physical violence, or, in the most extreme cases, assassination upon return. It is worth being precise here: Spyware such as Pegasus is fundamentally a hacking tool built on exploit chains, not an AI system in itself, but the data it harvests is increasingly sorted, cross-referenced, and prioritized by AI-driven analytics, which is why this paper treats it as part of the same broader landscape of AI-adjacent security technology rather than as evidence of AI proper. The same software was used by Saudi operatives in October 2018 to target the family members of murdered Saudi journalist Jamal Khashoggi, both before and after his death. Bahrain shows a similar trajectory: Between 2010 and 2012 it used FinFisher spyware to compromise and target journalists, lawyers, opposition politicians and pro-democracy activists, and Citizen Lab, a digital-security research group at the University of Toronto’s Munk School, later confirmed that, between June 2020 and February 2021, Bahrain deployed a zero-click iMessage exploit known as “FORCEDENTRY” to install Pegasus spyware on the devices of nine activists at home and abroad. The 2021 Pegasus Project, a collaborative investigation led by Amnesty International’s Security Lab and a consortium of international media outlets, further found that Moroccan authorities had used NSO Group’s spyware, NSO Group being the Israeli firm that develops Pegasus, to target dozens of journalists, activists and human rights defenders.
Facial recognition, biometric identification and predictive data analytics extend this same logic domestically. In Egypt, authorities use AI to monitor social media for signs of dissent, analyzing keywords, hashtags and patterns of online activity to predict, and in some cases preempt, potential protests. Across both patterns, these tools operate entirely outside judicial oversight or legal accountability, in direct violation of the rule of law and human rights protections that SSR seeks to build. This growing capacity for digital repression has forced many civil society organizations to relocate their operations abroad to avoid harassment, detention, or physical violence, including targeted killings in the most extreme cases. As a result, domestic civic space is contracting and local actors are steadily losing the ability to shape governance processes. This, in turn, erodes the nationally led character on which legitimate SSR depends and severs the state–society relationship that SSR ultimately works to rebuild.
These practices are not incidental; they are underwritten by legal architecture that gives them cover. Libya’s Law No. 5/2022, an anti-cybercrime measure, grants government authorities broad and vaguely defined powers to remove digital content deemed offensive to Libyan culture and values, with no independent oversight mechanism to prevent arbitrary or politically motivated enforcement, powers that extend readily to AI-assisted content monitoring and takedown. In Tunisia, data-privacy law does not distinguish between biometric and personal data, an ambiguity that allows governments and private entities to deploy AI-enabled biometric identification for surveillance and control without legal consequence. Egypt’s Cybercrime Law of 2018 goes further still: It treats any account with over 5,000 followers as a media outlet eligible for censorship, permits warrantless searches of personal devices, and allows AI-flagged social media accounts to be blocked without judicial authorization on the grounds of vaguely defined “false information”, a provision that has directly enabled increased penalties and harassment of journalists and activists. Together, these laws show that legislation can formalize AI-enabled repression as readily as it can constrain it: Where institutional checks are already weak, vaguely worded legal mandates become instruments through which AI is licensed to do harm, rather than safeguard against it. This is precisely the absence of the third safeguard this paper argues for: Binding legal frameworks that constrain AI use rather than legitimize its abuse.
None of this is AI being deployed within SSR. It is AI being deployed into SSG’s existing machinery, in ways that actively work against the preconditions SSR would need to take hold. What is absent, across every case above, is any framework capable of ensuring that AI, if ever introduced into SSR processes, serves reform rather than subverts it.
From Governance to Reform: How AI in SSG Forecloses SSR
Dr. Grazvydas Jasutis, an international security expert with extensive field experience advising on security sector reform, describes this as a fundamental “black box” problem: No sufficiently comprehensive accountability framework has yet emerged across the diverse contexts in which AI is being incorporated into SSG, a challenge that extends well beyond the security sector but carries particular stakes here. Without clear, comprehensive and legally grounded frameworks, AI risks reinforcing structural inequalities and corruption, undermining human rights, and further eroding what limited public trust remains in law enforcement and state institutions. This observation points to a structural problem that transcends any single case: For SSR, accountability, the second of this paper’s three safeguards, is not merely a procedural requirement but the mechanism through which institutions establish legitimacy in the eyes of the populations they are meant to serve.
When AI systems operate as black boxes, producing outputs whose logic cannot be examined, challenged, or explained, they undermine that legitimacy at its foundation. AI also carries limitations that complicate SSR – specifically, ones that no oversight architecture, however robust, can fully resolve. A Security Sector Reform Officer at the United Nations explains what AI fundamentally cannot understand in conflict-sensitive environments: The political dynamics, historical grievances, personal rivalries, social identities, and cultural nuances that SSR must navigate to rebuild trust and make context-specific judgments. Another related risk is the overreliance on algorithmic outputs, a documented phenomenon known as automation bias, in which human operators defer to automated systems even when the output is wrong or contradicts their own judgment. Objective courses of action cannot be drawn from AI without adapting them to context on the ground. In Libya, the Balkans, and various peacekeeping missions, the informal power of armed group commanders, deep historical grievances, and the institutional memory attached to particular community structures can determine whether a reform process succeeds or collapses. These dynamics cannot be readily translated into data points AI systems can process unless they are explicitly encoded in the algorithm, which is often impossible. This is not a limitation that better training data or more sophisticated algorithms can overcome. It reflects a fundamental difference between the relational knowledge SSR demands and the pattern-recognition capacity AI currently provides.
This overreliance on algorithmic outputs is most dangerous in precisely the environments where its appeal is greatest: Fragile, data-poor contexts where the promise of “data-certainty” is strongest and the capacity for human verification is weakest. In such settings, an AI system’s output can substitute for, rather than support, the judgment that effective reform requires. For practitioners, recognizing what AI can and cannot tell them, while keeping human judgment, local knowledge, and political accountability at the center of reform processes, is therefore essential. This distinction highlights the difference between AI supporting SSR and AI displacing it.
The Promise of AI for SSR: Relocating the Agency
None of this forecloses a role for AI in SSR, but it relocates the agency. The relevant question is not how states and their leaders might better deploy AI. On the contrary, the states most active in adopting these tools show little political will for the democratic oversight, transparency, or accountability that SSR requires. The more useful question is how SSR practitioners, civil society actors, donors, and international partners can leverage AI themselves to advance reform processes, even as the states around them use the same technology to entrench control. Concretely, that agency cuts both ways: The same practitioners, civil society actors, donors and international partners who push for SSR’s own preconditions are equally placed to insist on the safeguards AI itself requires – pressing for independent oversight bodies before AI tools are procured, building accountability into how those tools are deployed, and conditioning funding or partnership on binding legal limits governing their use. Security institutions in fragile contexts carry the accumulated weight of decades of impunity, unreformed power and the subordination of civilians to those who hold the means of coercion. Whether AI narrows or widens that gap depends not on the technology, but on whether it ends up in the hands of practitioners pushing for reform or the states resisting it.
The most credible positive case is also the most instructive for understanding the conditions under which AI can genuinely serve SSR objectives, precisely because AI did not make security-sector decisions at all. During the UN political roadmap process in Libya, AI was used to support multi-day structured dialogue processes: Organizing information into clear formats, generating references to relevant materials, translating documents, and synthesizing large volumes of data to facilitate more efficient dissemination. Recordings were transcribed and translated into Arabic, significantly improving accessibility. What would have required days of processing became a matter of minutes, freeing advisors and dialogue facilitators to focus on what AI cannot replace: Relationship- and trust-building, navigating competing interests, and forging consensus among Libyan actors. In a process where timing and momentum were everything, that efficiency dividend was more than logistical; it helped keep a fragile political process on track. What makes the Libya example relevant to SSR is not its scale but its design logic: AI supported human judgment in service of SSR’s relational core rather than replacing it.
A related but distinct example shows how AI can strengthen SSR practitioners’ work. Selma, a chatbot currently being developed by the UN and described as the world’s first SSR-focused AI tool, draws on roughly two decades of Secretary-General reports across country contexts to answer questions about what has been done in a given setting, what has worked, and what has not, synthesizing institutional memory that would otherwise require extensive manual research to reconstruct. This is a genuine contribution to SSR since, rather than relying on individual recall or fragmented archives, practitioners gain rapid, structured access to the accumulated record of how a security sector has actually functioned over time. Selma 2.0 expands this considerably, drawing on a corpus spanning 17 countries over 20 years, including humanitarian reports, periodic UN SSR documentation, and panel-of-experts assessments. Crucially, Selma is designed not to replace expert judgment but to feed it: The intended workflow requires that an SSR expert first validate the factual accuracy of any output, then apply their own judgment, informed by current UN plans and political context, before any decision is made. Selma is restricted to a curated, uploaded document base, with no open-web research permitted. This is a safeguard ensuring output remains traceable to a known and bounded evidence set. Selma exemplifies the same design logic as the Libya example, strengthening the foundation on which a particular security context is understood, in order to give SSR practitioners a firmer, faster starting point from which to exercise the political and technical judgment that reform actually requires, without ever making the reform decision itself. That said, the tool inherits the limits of its sources; Secretary-General reports are institutional and diplomatically curated rather than empirical, ground-truthed accounts. Selma’s outputs therefore risk reproducing the interpretive and political leanings already embedded in that reporting rather than correcting for them. The safeguard, again, is the expert in the loop, not the tool itself, a working instance of the independent oversight this paper treats as the first of its three preconditions.
At the operational level, AI also demonstrates real capacity to enhance public safety delivery in bounded, technically specific domains, offering the clearest counterexample in this paper of AI contributing to improvements in SSG that could be incorporated into responsible reform design. AI platforms are increasingly used in crowded locations to provide real-time insight into crowd density and movement, enabling interventions that deliver measurable gains in public safety. The Saudi Data and AI Authority’s “Baseer” platform is the clearest regional example: Deployed to monitor crowd density and movement at the Grand Mosque during the Hajj pilgrimage, it analyzes real-time data to detect unusual activity and enable timely intervention, helping standardize operational procedures, reduce human error, and improve the allocation of security resources.
Baseer is instructive precisely because it is bounded: It operates in a defined physical space, toward a narrow and measurable safety objective, with human officials retaining authority over any intervention it flags. This example, however, must be treated with caution; the same surveillance and crowd-monitoring capabilities that improve safety at a religious pilgrimage can be repurposed for political control just as easily elsewhere. The tool does not determine its own use. The institutional environment and the oversight frameworks that govern it do, which is precisely why the three safeguards this paper argues for cannot be treated as optional. For SSR practitioners, the implication of both examples is the same: AI is a powerful tool, but a fundamentally limited one. The choice is to learn its limits and work within them, or risk falling behind states already deploying it for control while forfeiting the reform opportunities it could otherwise support.
Conclusion
The question facing practitioners, donors, and international partners is not whether AI belongs in security sector governance and reform, but under what conditions it can serve reform rather than subvert it.
This paper has argued that those conditions are three and interlocking: Independent oversight mechanisms to check how AI is deployed within security institutions; accountability frameworks that assign clear responsibility when AI causes harm; and binding legal norms that constrain rather than formalize repression. In a region and context where AI is already being weaponized against populations SSR is meant to protect, their absence is not a gap to be managed but a warning to be heeded. In contexts where these preconditions do not exist, introducing AI into reform processes therefore exploits a governance void rather than filling one. This is not to say that AI should not be used. Responsibly bounded, operating in genuine support of human judgment, within defined parameters and toward transparency rather than control, AI can meaningfully advance the goals of institutional reform. The evidence from Libya and the Hajj pilgrimage demonstrates that the technology is not the problem, but the conditions of its deployment are.
The broader lesson is harder to escape: The infrastructure of repression is already in place. What remains absent is any framework capable of ensuring that AI, when introduced into SSG, serves the people it is supposed to protect. Until that framework exists, the burden falls on those who engage in pushing for meaningful SSR to name the risk plainly, to condition their support accordingly, and to resist the institutional temptation to treat technological capability as a substitute for political will. Reform cannot be automated, but it can be undermined by those who try.
The views represented in this paper are those of the author(s) and do not necessarily reflect the views of the Arab Reform Initiative, its staff, or its board.
