CyrilBirks

The ethical governance of AI in defence confronts a paradox: principles proliferate while implementation falters. This systematic review of 1085 publications exposes why frameworks flounder where ethical governance matters most. We identify seven barriers—governance and structural, conceptual, strategic, operational, relational and cultural, technical and data, and resource constraints—that operate as an interconnected system. Governance and structural barriers dominate the literature. This prominence may reflect their visibility rather than their primacy; technical, operational, and cultural dynamics remain harder to trace but are equally consequential. The challenges created by these seven barriers are severe but not insurmountable. We propose three strategic interventions: establishing authoritative, interoperable governance architectures that transcend organisational and national boundaries; building institutional capacity through dedicated resources and interdisciplinary expertise; and integrating ethical governance as core capability throughout the AI lifecycle rather than as compliance retrofit. This systemic analysis advances scholarship beyond principle enumeration and towards understanding the institutional, structural, and technical conditions under which ethical governance of AI in the defence domain can acquire operational traction in contested strategic environments.

@article{toffert2026seven,
  author  = {Antonia-Felicia Toffert and Katharina Klotz and Huw Roberts and Cyril Birks and Mariarosaria Taddeo},
  title   = {{Seven barriers to the ethical governance of artificial intelligence in defence}},
  journal = {AI \& SOCIETY},
  year    = {2026},
  doi     = {10.1007/s00146-026-03118-2},
  url     = {https://doi.org/10.1007/s00146-026-03118-2}
}

Ethical governance of artificial intelligence (AI) in defence must be legitimate under democratic standards and effective under operational pressure. The values it should serve are largely settled, as reflected in Just War Theory, International Humanitarian Law, and a convergent body of AI ethics. What is missing are the procedures to bind those principles to authoritative decision, institutional responsibility, and operational constraint. The deficit is therefore procedural, not substantive. In this article, we focus on ethical governance as a problem of policy design. Any such design must address three non-substitutable problems, those of knowledge, of authority and accountability, and of justification, each dominant at a different stage of the policy process: knowledge at inception, justification at deliberation, authority and accountability at operationalisation. We argue that defence does not create new problems but intensifies these, through recurring conditions such as operational tempo, secrecy, and coalition fragmentation that raise their difficulty without altering their kind. From this we build a stage-differentiated framework anchored by three instruments, a Non-Delegable Normative Decisions Register that records the substantive limits no procedure may override, diagonal-tiered transparency and accountability structures that allocate authority and extend scrutiny beyond the chain of command, and an Operational Requirements and Constraints Specification that converts those commitments into auditable parameters. The framework secures legitimacy through compliance with its procedures and makes its effectiveness assessable, so failures of either kind become visible and answerable.

@misc{toffert2026policy,
  author       = {Antonia-Felicia Toffert and Cyril Birks and Mariarosaria Taddeo},
  title        = {{A Policy Design Framework for Effective Ethical Governance of Artificial Intelligence in Defence}},
  howpublished = {SSRN preprint},
  year         = {2026},
  doi          = {10.2139/ssrn.6874566},
  url          = {https://doi.org/10.2139/ssrn.6874566}
}