Artificial Intelligence at the Crossroads of Gender, Peace and Security: A Still-Missing Link

The Women, Peace and Security agenda has been slow to address the growing use of artificial intelligence in conflict and security. Gender considerations remain largely absent from AI governance, national security policies and military applications unless purposely introduced. Making them a routine part of these systems will require necessary changes to how AI is designed, evaluated and regulated.

Introduction

October 2025 marked the twenty-fifth anniversary of the Women, Peace and Security (WPS) agenda. When United Nations Security Council Resolution 1325 was adopted in 2000, the digital revolution was still in its infancy; today, digital spaces are embedded in both everyday life and the practices of conflict and peace [1]. Over the same period, artificial intelligence (AI) has moved from technical curiosity to a structuring force in the economy, in governance and in warfare, inheriting and in some cases amplifying the gender inequalities of the societies that build it [2,3,4]. Its effects on gender equality and on peace are mixed and highly dependent on context, shifting the analytical question toward the institutional conditions that shape those effects and the actors who set them.

This article argues that three areas, research on gender bias in AI, the WPS agenda, and the international governance of AI, identify common weaknesses in how gender considerations appear in institutional practices. These are reflected in training data that treat men as the default user, in national action plans that mention “cyber” without analysing its gendered effects, and in language models that leave women out of conflict analysis unless prompted to include them. This exhibited pattern equates to “default omission”. The following sections trace “default omission” through the design of AI systems (I–II), digital violence and the WPS agenda (III), military uses of AI (IV) and conflict prevention (V), before outlining recommendations (VI).

I. Default settings: how AI reproduces gender inequality

Gender bias in AI is widespread across sectors and is well documented. An analysis of around 133 biased AI systems deployed between 1988 and 2021 found that 44.2 per cent exhibited gender bias and 25.7 per cent exhibited both gender and racial bias [5]. Generative models show the same pattern. Of 666 images produced by DALL-E 2 for 37 occupations, 59.4 per cent were strongly stereotyped, compared with 35 per cent in equivalent studies of human perceptions [4]; across nine text-to-image generators, images of lawyers, doctors, engineers and scientists depicted men in 76 per cent of cases and women in only 8 per cent [6]. Taken together, these studies indicate that AI systems amplify existing stereotypes as well as reproducing them [4,6].

Gender bias in AI-based decision-making is a socio-technical problem: it arises at the design, implementation, and institutional levels and cannot be corrected by adjusting algorithms alone [7,8]. These gender biases can enter AI through several mechanisms, including historically imbalanced datasets, language-processing techniques that reproduce occupational stereotypes, and development teams in which women make up only about 30 per cent of the AI workforce combine to make the male user the referential [2,3,7]. Empirical research on how organisations can manage such bias nonetheless remains thin, which leaves equality dependent on deliberate design choices that few developers are required to make [5,7,9].

The gender digital divide feeds bias into these systems at the data stage. In low-income countries, only around 20 per cent of women use the internet [2]; in Africa, women were 12 per cent less likely than men to be online in 2024 (31 per cent against 43 per cent) [10]. Lower smartphone ownership and digital literacy among women limit both their participation in the digital economy and their presence in training data [3]. Their underrepresentation in STEM and  in the design of these systems risks embedding gender prejudice in algorithms that increasingly decide on hiring, credit and access to services [9].

Once deployed, these design choices produce measurable harms that fall unevenly across groups of women. Diagnostic tools trained predominantly on male data perform less reliably for women, and commercial facial recognition systems have shown error rates below 1 per cent for lighter-skinned men but up to 35 per cent for darker-skinned women, carrying discrimination into surveillance, law enforcement and recruitment [3,8,9]. Because errors concentrate where gender intersects with race, class or disability, any assessment of whether a system performs accurately has to be intersectional [5,6,9].

II. Opportunities and risks for gender equality

When gender equality is written into their objectives, AI applications can widen women’s access to education, health care, financial services and justice, particularly in underserved regions, where a substantial share of the research on Sustainable Development Goal (SDG) 5 finds a predominantly positive effect [3]. Documented applications include platforms for reporting street harassment, chatbots providing sexual and reproductive health information, machine-learning tools that help target interventions against child marriage and female genital mutilation, and legal assistants for survivors of violence [3]. Other proposed uses include reducing bias in candidate screening, automating domestic tasks and drawing on large datasets for evidence-based advocacy [9]. In every case, the gains depend on equality being an explicit design goal.

The same technologies can reproduce and sometimes deepen existing inequalities when gender is not considered an integral part of design objectives. Automation threatens sectors with high female employment, such as administration, customer service and retail, with no guarantee of retraining [3,9]. Furthermore, voice assistants that default to female voices and are designed to be accommodating reinforce stereotypes of subservience [3,8]. Generative AI has added a more critical harm: an estimated 96 percent of deepfakes contain non-consensual sexual content, 99 percent of which targets women, with documented effects on safety, self-image and democratic participation [3,11]. Overall, AI could advance five of the nine SDG 5 targets while posing particular risks to those on non-discrimination and enabling technologies [3]. Which of these outcomes prevails depends on decisions about design, deployment and regulation, and it is in regulation that the WPS agenda becomes directly relevant.

III. The missing link: cybersecurity and technology-facilitated gender-based violence

The WPS agenda, the main international framework linking gender to security, has been slow to respond as gendered threats have moved online. Technology-facilitated gender-based violence (TFGBV) is now common: 38 per cent of women surveyed across 51 countries report having experienced online violence and 85 per cent report having experienced or witnessed it [11], while regional estimates range from 16 to 58 per cent [12]. These figures matter for WPS because this violence targets the women whose participation the agenda is meant to secure. Seventy-three per cent of women journalists report online violence in the course of their work [12]; women negotiators have received rape and death threats online; and in Myanmar, women activists who were doxxed were subsequently arrested by the junta [11]. Gendered disinformation, increasingly produced with generative tools, frames women politicians as untrustworthy, incompetent or too emotional to hold office, and state-aligned actors have used it to silence opposition voices [13,14]. Online abuse therefore works as a mechanism of exclusion from public and political life, and it can escalate into offline harm [11,15].

National action plans (NAPs), the main instrument for implementing the agenda, have taken up cyber issues more readily than they have analysed them. In 2021, only two of the 98 NAPs then in force addressed cyber threats [15]. A 2025 review of 37 national and regional plans found broader but still shallow engagement: most refer to cyber issues only minimally, eight show significant engagement across pillars, and only five explicitly name TFGBV [1]. The two samples are not directly comparable, but together they suggest that cyber has entered WPS planning faster than the analysis needed to make it operational. Several plans follow an “add cyber and stir” approach, mentioning cybersecurity without examining how it intersects with gender and security, and most fail to connect their cyber provisions to national cybersecurity strategies, disinformation responses or early warning systems [1]. This is default omission at the policy level: the term is present, the analysis is not.

The gaps are concentrated in the pillars and actors most relevant to remedy. The relief and recovery pillar is almost entirely absent from cyber provisions, existing digital rights frameworks are rarely invoked, and the private sector, which hosts and amplifies much of this violence, is engaged in none of the plans reviewed [1]. This omission runs against calls for international coalitions to press technology companies to apply consistent moderation and enforcement standards across jurisdictions [11]. It also mirrors the marginal place of gender in international cyber diplomacy, where women accounted for only about 20 per cent of the experts nominated by governments to the UN Groups of Governmental Experts on ICTs [15].

The plans that perform best ground their cyber provisions in broad consultation and rely on partnerships with real implementation capacity. Colombia’s plan bases its cyber provisions on a six-month national consultation with more than 1,500 Indigenous, Afro-Colombian, LBT and other women, while ASEAN’s Regional Plan of Action, supported by UN Women, connects early warning, action against cybercrime and women’s participation in responses to digital threats [1]. Yet Colombian civil society organisations have criticised the lack of follow-up on the plan [1], and NATO, one of the few institutions whose WPS framework now names TFGBV [1,16], has not yet shown how that commitment applies to the AI systems its members deploy [17]. The quality of consultation opens the way to effective plans; sustained follow-up and resources determine whether they are implemented.

IV. When AI enters the battlefield: the WPS compliance gap

Military uses of AI carry default omission into decisions on the use of force, at a time when multilateral governance of these uses remains stalled. After more than a decade of discussions under the Convention on Certain Conventional Weapons, states converged in 2025 on the notion of “context-appropriate human control and judgment”, but still disagreed on whether autonomous weapon systems should be qualified as “lethal” and on whether international humanitarian law requires additional rules [18]. Diverging approaches among major powers deepen the impasse. The European Union is implementing its AI Act, China promotes multilateral capacity-building through its 2025 Global AI Governance Action Plan, and the United States, whose July 2025 AI Action Plan seeks “unquestioned and unchallenged global technological dominance”, voted against the December 2025 General Assembly resolution on lethal autonomous weapons systems [18]. In SIPRI’s account of these processes, gender is largely absent; bias mitigation appears only as one item in a list of principles proposed by states for military AI [18].

Some security institutions have begun to name these risks without yet applying WPS standards to the systems they use. NATO’s 2024 WPS Policy refers explicitly to gendered disinformation, TFGBV and the gender dimensions of AI [16], and its revised AI strategy lists technology-facilitated gender-based violence among AI-enabled threats and includes bias mitigation among its principles of responsible use [19]. Meanwhile, AI tools already inform targeting and intelligence in conflicts including those in Ukraine, Gaza, Yemen and Syria, yet there is no indication that the WPS commitments of the states involved have been extended to these systems [17]. Tests of three leading large language models across 13 conflict scenarios show that when WPS language is absent from prompts, as it is in real field situation reports, scores on WPS integration fall from 0.65 to 0.08, a drop of nearly 90 per cent [17]. The same models performed well once given system prompts based on WPS principles and access to national action plans [17]. The shortfall therefore lies in how systems are configured and procured, which makes it correctable through contracts and standards.

Research on military AI anticipated this pattern several years ago. Training data drawn from past conflicts can turn contingent assumptions into fixed rules, for example by over-identifying military-aged men as combatants while overlooking women’s roles as fighters, civilians or peacebuilders [8]. Gender analysis therefore concerns men as well: when sex and age serve as proxies for combatant status, the protection owed to civilians is eroded [8,20]. Gender-based audits, independent testing against demographic benchmarks and evaluation teams that include gender expertise were recommended in 2021 [8]; the competence gap measured in 2026 indicates that these measures have not been adopted at scale. When a decision-support system leaves women out of its analysis, the cost falls on populations that states have formally committed to protect under Resolution 1325 [8,17].

V. From reaction to prevention: governance, trust and local ownership

The prevention pillar of the WPS agenda provides an existing framework for moving AI governance beyond the mitigation of harms after they occur. Current approaches to AI governance remain fragmented, largely voluntary and oriented toward innovation or after-the-fact mitigation, and no framework systematically integrates gender analysis, conflict-prevention research and the WPS agenda across the life cycle of AI systems [21]. Proposals to fill this gap rely on existing instruments: in the United States, the Women, Peace and Security Act of 2017 and the Elie Wiesel Genocide and Atrocities Prevention Act of 2018 could be used to require gender-responsive, conflict-informed risk assessments of AI used in national security [21]. Multilaterally, the Global Digital Compact, adopted in 2024 as part of the Pact for the Future, lists gender equality among its principles and commits states to address gender-based violence that occurs through or is amplified by technology [2,22]. The bodies created in its wake, the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance launched in 2025 [18], are therefore natural venues for introducing WPS standards.

AI can support prevention only if the data it relies on and the institutions that use it are trusted by the communities concerned. Biased or manipulated information can distort early warning, AI is most useful when it supports analysts’ political judgement without replacing it, and responsible governance frameworks are a precondition for institutional trust [23]. These conditions have a gender dimension that is rarely spelt out. Early warning systems fed by data in which women are underrepresented are likely to miss signals that WPS research treats as indicators of instability, such as rising online misogyny or threats against women activists [11,21]. Some WPS advocates go further and treat gender inequality in technology as itself an early warning sign of political and social instability [14].

Humanitarian assistance and disaster relief, where armed forces increasingly use AI, show the same requirement in a different setting. AI supports forecasting, needs mapping and the targeting of aid, but without sex-disaggregated data, women in hard-to-reach areas may not appear in the information on which relief is planned [24]; the gender gaps in internet access and device ownership documented in Southeast Asia, Africa and low-income countries compound this risk [2,10,24]. UN Women’s Bridges for Peace project, supported by Canada, works with the armed forces of Indonesia, Malaysia, the Philippines, Timor-Leste and Viet Nam to use gender-responsive AI to identify gendered risks and strengthen early warning before crises occur [24]. Such initiatives link the prevention pillar of the WPS agenda to relief and recovery, the pillar that national action plans most often leave out of their cyber provisions [1].

African institutions and youth initiatives show both what preventive AI can achieve and the conditions it requires. The African Union Peace and Security Council has moved from early techno-optimism toward what has been described as “AI realism”: at its 1339th meeting in April 2026, it emphasised transparency, accountability and inclusivity, called for women and youth to be included in AI governance and for AI systems to be developed in African languages, even as concerns over autonomous weapons and drone warfare grow [25]. Young Africans already use digital tools for peacebuilding. Ushahidi was created in Kenya after the 2007–2008 post-election violence, iVerify supports fact-checking in Zambia, Sierra Leone and Liberia, and young researchers have built hate-speech lexicons in local languages in Cameroon, Ethiopia, Kenya and South Sudan [10]. These initiatives face structural limits. Only 38 per cent of Africans used the internet in 2024, against 68 per cent globally; only 49 of more than 2,000 African languages are available on Google Translate; and content moderation failures in Amharic and Tigrinya allowed harmful content to spread during the 2020–2022 war in Ethiopia [10]. Connectivity and language coverage determine whose warning signs an AI system can detect, which makes them central to the inclusiveness and legitimacy of preventive AI [10,25].

VI. Toward coherent integration: recommendations

Correcting default omission requires making gender explicit at each stage where AI systems are designed, procured, governed and evaluated. Each recommendation below identifies the actors responsible, the mechanism and an indicator against which progress can be measured.

Make gender impact assessment a condition of deployment.

Developers and public authorities deploying AI in high-risk areas such as employment, credit, health and policing should carry out and publish gender and intersectional impact assessments before deployment, report error rates disaggregated by sex and, where possible, by race and age, and repeat the assessment after major updates [3,5,9]. In the European Union, these assessments can be built into the risk-management and fundamental-rights assessments that the AI Act already requires for high-risk systems, avoiding a parallel procedure [26]. Because the bias is socio-technical, these assessments should be paired with organisational measures, such as teams that combine data scientists and social scientists and regular internal fairness audits [7]. Indicator: the share of high-risk systems with published, disaggregated performance data.

Turn cyber references in national action plans into operational commitments.

Governments drafting or revising NAPs should name and define TFGBV explicitly; attach a budget line, a lead agency and measurable targets to cyber commitments under every pillar, including relief and recovery (for instance, survivor support services and the preservation of digital evidence for prosecution); establish a standing coordination mechanism between WPS focal points, national cybersecurity agencies and justice ministries; and require platform companies operating in the country to offer multilingual reporting and redress channels, and report on how quickly they respond to gendered abuse [1,11]. Cyber commitments in NAPs should also be cross-referenced in national cybersecurity strategies [1]. Consultations should follow the Colombian model of structured national consultation with diverse women’s groups, and each plan should provide for an independent mid-term review [1]. Indicator: the number of NAPs with costed, measurable cyber commitments in more than one pillar.

Write WPS standards into defence AI procurement.

Ministries of defence and international organisations with WPS obligations, including the United Nations, NATO and the European Union, should require vendors to meet WPS performance thresholds on standardised benchmarks before contracts are awarded, configure decision-support systems with access to national action plans and WPS guidance, and test them on sparse, realistic prompts that contain no WPS cues [17]. This testing should be complemented by gender-based audits conducted by independent teams with gender expertise [8]. AI tools used for military recruitment, assessment and promotion require the same scrutiny: with women making up on average only 13.9 per cent of Allied armed forces personnel in 2024 [27], models trained on historical personnel data are likely to reproduce the pattern documented in civilian recruitment algorithms [7]. States should also raise gender analysis in the Convention on Certain Conventional Weapons, in the Geneva informal exchanges on military AI mandated by General Assembly resolution 80/58, and in the REAIM process [18]. Indicator: the proportion of defence AI contracts containing verifiable WPS performance clauses.

Build gender and local languages into AI for early warning and humanitarian response.

Early warning systems, including the AU Continental Early Warning System, should incorporate gender-specific indicators, such as online misogyny and threats against women activists, and be designed to operate in local languages and through low-tech channels such as SMS and radio [10,11]. Armed forces and humanitarian agencies using AI for disaster relief should collect and use sex-disaggregated data as a standard requirement [24]. The AU Commission and Regional Economic Communities should establish independent AI review panels that combine youth, gender, human rights and data science expertise, and formalise advisory roles for young women and men in early warning [10,25]. Donors should fund interdisciplinary research linking technical development, gender analysis and conflict prevention, together with connectivity and digital literacy programmes targeted at rural and conflict-affected areas [10,21]. Indicator: the share of early warning indicators that are sex-disaggregated or gender-specific.

Use the Global Digital Compact follow-up to connect the agendas.

Member States and UN Women should ensure that the Independent International Scientific Panel on AI includes WPS and gender expertise and that the Global Dialogue on AI Governance places TFGBV and gender bias on its agenda, in line with the Compact’s commitments on gender equality [18,22]. Indicator: the inclusion of gender analysis in the Panel’s first assessment report.

References

1. Baekgaard K. The missing link: cybersecurity and technology-facilitated gender-based violence in the Women, Peace and Security agenda. How national action plans address (and overlook) digital threats. [place unknown]: Association for Progressive Communications; 2025 Oct. ISBN: 978-92-95113-77-0.
2. UN Women. Artificial intelligence and gender equality [Internet]. New York: UN Women; 2024 [cited 2026 Oct 4]. Available from: https://www.unwomen.org/en/articles/explainer/artificial-intelligence-and-gender-equality
3. Stecker L, Boecker M, Kowol U, Raniga T. AI and gender: the ambivalent role of AI in advancing gender equality for SDG 5. J Hum Rights Soc Work. 2026. doi:10.1007/s41134-025-00432-5
4. García-Ull FJ, Melero-Lázaro M. Gender stereotypes in AI-generated images. Prof Inf. 2023;32(5):e320505. doi:10.3145/epi.2023.sep.05
5. Smith G, Rustagi I. When good algorithms go sexist: why and how to advance AI gender equity. Stanford Soc Innov Rev [Internet]. 2021 Mar 31 [cited 2026 Oct 4]. Available from: https://ssir.org/articles/entry/when_good_algorithms_go_sexist_why_and_how_to_advance_ai_gender_equity
6. Gorska AM, Jemielniak D. The invisible women: uncovering gender bias in AI-generated images of professionals. Fem Media Stud. 2023;23(8):4370–5. doi:10.1080/14680777.2023.2263659
7. Nadeem A, Marjanovic O, Abedin B. Gender bias in AI-based decision-making systems: a systematic literature review. Australas J Inf Syst. 2022;26. doi:10.3127/ajis.v26i0.3835
8. Chandler K. Does military AI have gender? Understanding bias and promoting ethical approaches in military applications of AI [Internet]. Geneva: United Nations Institute for Disarmament Research; 2021 [cited 2026 Oct 4]. Available from: https://unidir.org/publication/does-military-ai-have-gender-understanding-bias-and-promoting-ethical-approaches
9. Chen Y. Gender inequality in the age of AI: predictions, perspectives, and policy recommendations [unpublished essay]. ORCID: 0000-0002-8457-6757.
10. Liaga EA. Advancing youth-led peacebuilding through digital technologies in Africa [Internet]. Pretoria: Institute for Security Studies; 2025 Nov 4 [cited 2026 Oct 4]. (ISS Policy Brief; no. 203). Available from: https://issafrica.org/research/policy-briefs/advancing-youth-led-peacebuilding-through-digital-technologies-in-africa
11. Baekgaard K. Technology-facilitated gender-based violence: an emerging issue in women, peace and security [Internet]. Washington (DC): Georgetown Institute for Women, Peace and Security; 2024 [cited 2026 Oct 4]. Available from: https://giwps.georgetown.edu/wp-content/uploads/2025/11/GIWPS-TFGBV-Report-2025.pdf
12. UN Women. FAQs: digital abuse, trolling, stalking, and other forms of technology-facilitated violence against women and girls [Internet]. New York: UN Women; 2025 Nov 13 [cited 2026 Oct 4]. Available from: https://www.unwomen.org/en/what-we-do/ending-violence-against-women/faqs/tech-facilitated-gender-based-violence
13. Di Meco L, Wilfore K. Gendered disinformation is a national security problem [Internet]. Washington (DC): Brookings Institution; 2021 Mar 8 [cited 2026 Oct 4]. Available from: https://www.brookings.edu/articles/gendered-disinformation-is-a-national-security-problem/
14. Our Secure Future. WPS message guide: a gender perspective on AI risks to national security. Washington (DC): Our Secure Future; [date unknown].
15. Sharland L, Goussac N, Currey E, Feely G, O’Connor S. System update: towards a women, peace and cybersecurity agenda [Internet]. Geneva: United Nations Institute for Disarmament Research; 2021 [cited 2026 Oct 4]. Available from: https://unidir.org/wp-content/uploads/2023/05/UNIDIR_System_Update.pdf
16. North Atlantic Treaty Organization. NATO policy on women, peace and security [Internet]. Brussels: NATO; 2024 Jul 10 [cited 2026 Oct 4]. Available from: https://www.nato.int/cps/en/natohq/news_227243.htm
17. Whelan M, Frant J. Women, peace and security frameworks must apply to defense AI [Internet]. Tech Policy Press; 2026 Apr 21 [cited 2026 Oct 4]. Available from: https://www.techpolicy.press/women-peace-and-security-frameworks-must-apply-to-defense-ai/
18. Bruun L, Palayer J, Boulanin V. Artificial intelligence and international peace and security. In: SIPRI yearbook 2026: armaments, disarmament and international security [Internet]. Oxford: Oxford University Press; 2026 [cited 2026 Oct 4]. p. 437–59. Available from: https://www.sipri.org/sites/default/files/YB26%2014%20AI%20Governance.pdf
19. North Atlantic Treaty Organization. Summary of NATO’s revised artificial intelligence (AI) strategy [Internet]. Brussels: NATO; 2024 Jul 10 [cited 2026 Oct 4]. Available from: https://www.nato.int/cps/en/natohq/official_texts_227237.htm
20. Acheson R. Gender and bias: what does gender have to do with killer robots? [Internet]. Stop Killer Robots; 2021 [cited 2026 Oct 4]. Available from: https://www.stopkillerrobots.org/wp-content/uploads/2021/09/Gender-and-Bias.pdf
21. McGill K. From reaction to prevention: WPS and artificial intelligence [Internet]. Washington (DC): New Lines Institute for Strategy and Policy; 2026 Aug 19 [cited 2026 Oct 4]. Available from: https://newlinesinstitute.org/gender-peace-security/from-reaction-to-prevention-wps-and-artificial-intelligence/
22. United Nations General Assembly. The Pact for the Future. Annex I: Global Digital Compact. New York: United Nations; 2024 Sep 22. UN Doc. A/RES/79/1.
23. Song MJ, Cucinotta G. Building trust in data and AI for conflict prevention: three lessons from the Data for Peace Conference 2026 [Internet]. Turin: United Nations System Staff College; 2026 Jul 15 [cited 2026 Oct 4]. Available from: https://www.unssc.org/news-and-insights/blog/building-trust-data-and-ai-conflict-prevention-three-lessons-data-peace
24. Nair T. Towards gender-responsive AI in humanitarian assistance and disaster relief operations. Singapore: S. Rajaratnam School of International Studies, Nanyang Technological University; 2026 Aug 13. (RSIS Commentary; no. 175).
25. Dersso SA, Habu P. From optimism to AI realism: the African Union Peace and Security Council on AI and peace and security [Internet]. Addis Ababa: Amani Africa; 2026 Jun 10 [cited 2026 Oct 4]. Available from: https://amaniafrica-et.org/from-optimism-to-ai-realism-the-african-union-peace-and-security-council-on-ai-and-peace-and-security/
26. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Off J Eur Union. 2024 Jul 12;L 2024/1689.
27. NATO International Military Staff, Office of the Gender Advisor. Summary of the national reports of NATO member and partner nations to the NATO Committee on Gender Perspectives 2023–2024 [Internet]. Brussels: NATO; 2025 [cited 2026 Oct 4]. Available from: https://www.nato.int/content/dam/nato/webready/documents/genad/2023-2024-ncgp-summary.pdf