EU AI Act Military Gap: Europe’s Regulatory Dilemma 

The EU AI Act is widely praised as the world’s most comprehensive artificial intelligence regulation. Yet, buried in Article 2(3) is a single word—“exclusively”—whose implications remain insufficiently unpacked. In practice, it exempts AI systems developed or used solely for military or national security purposes from the Act’s scope¹. However, in an era where technologies routinely operate across civilian and security domains, the boundary implied by “exclusively” is far from clear-cut. Dual-use systems, hybrid applications, and cross-border data flows blur this distinction, creating a significant ambiguity at the heart of the regulation. The result is not merely a technical loophole, but a potential democratic deficit within Europe’s AI governance framework.

What the EU AI Act Actually Excludes 

The AI Act defines high-risk systems, mandates independent audits, and demands algorithmic transparency. These are meaningful protections. But a system certified as military or national-security-related falls outside the boundary.

On paper, Article 2(3) is a limited technical exception. In practice, it creates a loophole: an autonomous target-recognition system can be certified as military, thereby bypassing every audit requirement, and then be redeployed at borders or in humanitarian missions — with no mechanism tracking that transition. The Court of Justice of the EU insists that such exceptions must be interpreted restrictively. The Treaty on European Union protects the “exclusive responsibility of Member States” in national security matters². A friction arises because EU law seeks to narrowly constrain exceptions, while simultaneously deferring to Member States’ broad and self-defined authority over national security. 

However, the deeper problem is political, not technical. As long as defence ministries evaluate themselves through internal ethics committees, as do France³,  the European model of fundamental rights rests on a so-called “gentlemen’s agreement” since the system relies on the expectation that states will voluntarily act ethically—even when no effective external accountability forces them to do so. And gentlemen’s agreements do not survive in times of urgency.

Why “Military” and “Civilian” AI Cannot Be Separated

The idea that military AI constitutes a separate, distinguishable category from other AI is one of the biggest misconceptions in European policy. The AI computer vision on Ukrainian drones that is used to detect targets is the same technology Amazon uses in its warehouses.  Similarly, the language models analysing intercepted communications are identical to those processing customer service calls.

For example, Elon Musk’s Starlink, a private civilian infrastructure, operated by SpaceX, became the backbone of Ukrainian military communications⁴.

A system may be partly developed under civilian criteria, funded through Horizon Europe, EU’s key funding programme, and partly under defence criteria. For this reason, when deployed operationally, no single authority has a complete view. International Humanitarian Law was not designed to resolve situations that straddle defence and civilian contexts ⁵.

The European Defence Fund (EDF) explicitly prohibits financing fully autonomous weapons research⁶. Yet it funds AI components, perception modules, autonomous navigation, decision-support algorithms used in weapons and indistinguishable from those in a fully autonomous lethal system⁷. The line between the whole and its parts determines legal accountability. Nobody is drawing it.

Almost no AI system relevant to defence is exclusively military. The audit chain is fragmented from the outset and so far it seems that nobody has a complete view of the whole. 

Lethal Autonomous Weapons: The Accountability Gap No Law Has Closed

This raises a timely and tough question: If a dual-use algorithm that was developed for civilian use but adapted for defence and deployed in a crisis operation that is not formally a war commits a lethal error, who is responsible?

The commander who authorised it? The programmer of the underlying civilian model? The company that built the training dataset labelled for commercial use? In the words of researchers working directly in this field: the chain of accountability becomes literally untraceable.

The European Parliament has demanded “meaningful human control” over all weapons systems⁸. The EDF Regulation prohibits fully autonomous weapons research⁹ but in December 2024, a UN General Assembly resolution on lethal autonomous weapons systems (LAWS) received 166 votes in favour while the United States, Russia, China, and several EU Member States resisted legally binding restrictions¹⁰.

The Geneva Conventions require distinction between combatants and civilians. NATO and the UN insist machines “should never autonomously make life or death decisions” ¹¹. But the existing framework relies heavily on state interpretation and self-regulation, and when decision-making chains span civilian developers, military integrators, and commanders across multiple jurisdictions, significant accountability gaps remain¹².

Europe’s Late Awakening: How the Peace Illusion Left a Technology Gap

Before February 2022, the dominant European political consensus held that large-scale conventional war on the continent was unthinkable. Germany’s constitutional debt brake made significant defence spending almost impossible. For years, Poland and the Baltic states had warned about Russian intentions, but nobody in Brussels or Berlin listened seriously.¹³

Russia’s full-scale invasion of Ukraine changed everything, not just in budget terms, but conceptually. Europe found itself a decade behind in military AI integration. The United States had been running Project Maven since 2017¹⁴. China formalised its civil-military AI fusion doctrine as state policy in 2015¹⁵. Europe had sophisticated regulatory frameworks and twenty-seven fragmented sovereignties. And its delay comes from fragmentation, a regulatory-first approach, and a persistent separation between civilian and military innovation, which together prevented the rapid operational integration of AI seen in more centralized systems like the United States and China.

As Dempsey¹⁶ notes, the Ukraine conflict demonstrated that industrial-scale drone warfare and AI-enabled electronic warfare have permanently altered the operational landscape. European defence planners who once treated autonomous systems as a distant future problem now confront their live deployment on the continent’s eastern flank.

 Europe Has the Rules. But What About the Regulator?

The European Defence Agency coordinates but does not supervise. It has no coercive power. There is no European equivalent of the Pentagon’s Chief Digital and Artificial Intelligence Office to oversee military AI from development through to deployment.

The result is that military AI standards  represent the lowest common denominator that is acceptable to all twenty-seven Member States. Grey zones proliferate; we see spaces where AI advances with limited oversight, where algorithmic systems profile and stigmatise vulnerable communities, and where discrimination is obscured behind a veneer of technical neutrality.¹⁷

What would a supranational AI defence regulator require?

Europe needs a supranational AI Military Chief Director role: a centralised European figure capable of coordinating criteria, auditing dual-use systems regardless of their funding origin, and establishing clear limits on AI in defence and security. Without it, the gap between stated values and operational reality will only widen.

AI Accountability and Democracy: The Habermasian Test

Addressing Europe’s regulatory vacuum does not imply technological Luddism. It is a demand that complexity be managed with the tools proper to democratic societies: transparency, parliamentary oversight, independent audits, and real accountability.

Habermas’s theory of communicative action is instructive here. He highlighted that this type of action refers to a type of social interaction in which individuals coordinate their actions through reasoned dialogue aimed at mutual understanding, rather than through force, manipulation, or self-interest.

Thus, democratic legitimacy derives from public deliberation — from citizens’ capacity to contest and revise the norms governing their lives. Therefore, algorithmic systems making opaque, consequential decisions  are constitutively illegitimate. This means that extending communicative action to source code is to demand that AI systems deployed in public functions be, in principle, publicly accessible and contestable.

The EU AI Act is a valuable starting point. But while Article 2(3) remains unchanged, while “exclusively” functions as a universal shield against accountability, the world’s most sophisticated regulatory framework has a war-sized hole at its centre.

Democracy must be compatible with technology, but this will only happen if we insist that technology be accountable to democracy. This is not a utopian dream but rather the minimum condition for the European project to continue making sense.

References

Thanks to Markus Spiske from Unsplash for the Header Image.

1. European Commission. (2021). Regulation (EU) 2021/697 establishing the European Defence Fund. Official Journal of the European Union, L 170, 149–177. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32021R0697
2. Treaty on European Union (Consolidated Version). (2012). Official Journal of the European Union, C 326, 13–390. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:12012M/TXT
3. Boulanin, V., & Verbruggen, M. (2017). Mapping the development of autonomy in weapon systems. SIPRI. https://www.sipri.org/publications/2017/other-publications/mapping-development-autonomy-weapon-systems
4. Abels, B. (2024). Private infrastructure and the transformation of warfare: Starlink in the war in Ukraine. European Journal of International Relations
5. International Committee of the Red Cross. (2021). ICRC position on autonomous weapon systems. https://www.icrc.org/en/document/icrc-position-autonomous-weapon-systems
6. European Commission. (2021). Regulation (EU) 2021/697 establishing the European Defence Fund and repealing Regulation (EU) 2018/1092. Official Journal of the European Union.
7. Broeders, D., Adamson, L., & Creemers, R. (2023). Governing military AI in Europe: Regulatory challenges and institutional gaps. European Journal of International Security, 8(2), 145–163. https://doi.org/10.1017/eis.2023.14
8. European Parliament. (2023, September 12). Resolution on autonomous weapons systems. https://www.europarl.europa.eu/doceo/document/TA-9-2023-0313_EN.html
9. European Commission. (2021). Regulation (EU) 2021/697 establishing the European Defence Fund. Official Journal of the European Union, L 170, 149–177. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32021R0697
10. United Nations General Assembly. (2024, December 16). Resolution 79/52: Lethal autonomous weapons systems. https://digitallibrary.un.org/record/4058977
11. NATO. (2021). Principles of responsible use of artificial intelligence in defence. https://www.nato.int/cps/en/natohq/official_texts_187617.htm
12. Roff, H. M. (2016). The strategic robot problem: Lethal autonomous weapons in war. Journal of Military Ethics, 13(3), 211–227.
13. HCSS (2025)
    The Hague Centre for Strategic Studies. (2025). [Title of the report]. The Hague Centre for Strategic Studies.
14. Pellerin, C. (2017, July 21). Project Maven to deploy computer algorithms to war zone by year’s end. U.S. Department of Defense. https://www.defense.gov/News/News-Stories/Article/Article/1254719/
15. Kania, E. B. (2019). Chinese military innovation in artificial intelligence. Center for a New American Security. https://www.cnas.org/publications/testimony/chinese-military-innovation-in-artificial-intelligence
16. Dempsey, J. (2023). Lessons from Ukraine: The transformation of European defence. Carnegie Europe. https://carnegieeurope.eu/strategiceurope/89743
17. Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin’s Press.