Skip to content
OGERIA — Observatory of Global Evidence on Risks in AISynthesis report · 2026 ed.
Updated 2 Oct 2026

Chapter 01 · Risk map

Military and autonomous weapons

Autonomous weapons without meaningful human control

Systems that identify, select and engage a target without a person deciding that specific target at that specific moment.

Severity
Severe
Horizon
Already happening
Evidence
Observed
Consensus
Medium

The word “autonomous” names three different things and only one of them is this risk: that the system selects the target. The 128 states at Geneva excluded from the definition those systems where personnel determine the specific target and the timing of the attack [208]Report of the 2024-2025-2026 sessions of the Group of Governmental Experts on Emerging Technologies in the Area of Lethal Autonomous Weapons Systems (CCW/GGE.1/2026/CRP.1/Rev.1)Group of Governmental Experts on Emerging Technologies in the Area of Lethal Autonomous Weapons Systems · 2026 · official documentView source ↗Accessed on 9 September 2026.

What is deployed at scale in Ukraine does not cross that line. Kateryna Bondar concludes that modern uncrewed systems there are still not fully autonomous: what exists is last-mile navigation, which takes over after the operator has confirmed the target and raises the hit rate from 10-20% to 70-80% [299]Ukraine's Future Vision and Current Capabilities for Waging AI-Enabled Autonomous WarfareBondar, Kateryna · 2025 · reportView source ↗Accessed on 9 September 2026. The case that is genuinely on the boundary is the Molniya drone that killed three civilians in Zaporizhzhia on 6 July 2026: the operators set the destination and the area, and once on site the system is reported to have picked the specific object on its own [385]A Russian drone chose its own target and killed three. The code that told it what to look for was readableEuromaidan Press · 2026 · pressView source ↗archived copy onlyAccessed on 9 September 2026. The most immediate risk may not be that one, but the one described by the investigations into Habsora and Lavender: automatic target generation with about twenty seconds of human review, and 100 targets in a day where there used to be 50 a year [670]'Lavender': The AI machine directing Israel's bombing spree in GazaAbraham, Yuval · 2024 · pressView source ↗archived copy onlyAccessed on 9 September 2026 [668]'A mass assassination factory': Inside Israel's calculated bombing of GazaAbraham, Yuval · 2023 · pressView source ↗archived copy onlyAccessed on 9 September 2026.

What this does not demonstrate. There is no UN report on fully autonomous drones. There is a statement by High Commissioner Volker Türk of 7 September 2026 saying he is horrified by “reports that” such drones did this —not a verification— and which in the same sentence adds that Ukraine has reportedly also field-tested them [814]Global Update to the 63rd session of the Human Rights CouncilTürk, Volker · 2026 · official documentView source ↗Accessed on 9 September 2026. The attribution of autonomy in the Molniya case is forensic inference, not telemetry. And current US policy already allows weapons with the human out of the loop provided human judgement was exercised earlier, in design and deployment [290]Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for CongressSayler, Kelley M. · 2026 · official documentView source ↗archived copy onlyAccessed on 9 September 2026: the threshold in dispute is not technological but definitional.

Chain of materialisation

  1. PreconditionObserved

    Current doctrine already allows humans out of the loop

    The Congressional Research Service clarifies that the requirement of human judgement over the use of force does not mean these systems operate with a human in the loop. The Geneva characterisation excludes from the definition systems where personnel determine the specific target and the timing of the attack.

    Precedents: The CCW group of experts closes without a negotiating mandate

  2. TriggerObserved

    A system picks the specific object within the area it was assigned

    On 6 July 2026 a Molniya drone killed three civilians in Zaporizhzhia. Human operators set the destination and the area before launch; once on site, the onboard system reportedly identified and selected the specific object -most likely propane tanks it had been trained to recognise- on its own.

    Precedents: A Molniya drone kills three civilians in Zaporizhzhia · The UN High Commissioner calls for banning weapons that kill without human intervention

    Observed and demonstrated evidence ends here. What follows is projection.

  3. CascadeProjected

    From terminal selection to a swarm that coordinates

    Bondar is explicit: fully realised swarms -where drones communicate, decide and adapt together- have not been deployed, and Ukraine has not operated true swarms in combat. The US Congress assesses that swarming would probably require further advances in AI or networked communications. That remains true for deployment, but stopped being true for the software: in 2026 Anthropic detected and disabled a non-state team that had built -with Claude Code- shared swarm memory and fault-tolerant coordination logic for an autonomous kamikaze drone swarm, with target selection (including a "person" class) and detonation orders with no human in the loop, tested with real hardware. Anthropic rates it at a maturity level validated in simulation, not operated.

    Precedents: Anthropic's threat report documents dual-use biology and state distillation

  4. ImpactSpeculative

    Attacks with nobody to attribute the decision to

    The expert group report states that international humanitarian law imposes obligations on States, parties and individuals, not on machines, and that responsibility cannot be transferred to machines. How that responsibility is distributed in practice is unresolved in every jurisdiction.

See on the map →Report a mistake in this entry →

Sources

  1. [299] Ukraine's Future Vision and Current Capabilities for Waging AI-Enabled Autonomous Warfare · CSIS Wadhwani AI Center 2025
  2. [385] A Russian drone chose its own target and killed three. The code that told it what to look for was readable · Euromaidan Press 2026 archived copy only
  3. [814] Global Update to the 63rd session of the Human Rights Council · Oficina del Alto Comisionado de las Naciones Unidas para los Derechos Humanos 2026
  4. [208] Report of the 2024-2025-2026 sessions of the Group of Governmental Experts on Emerging Technologies in the Area of Lethal Autonomous Weapons Systems (CCW/GGE.1/2026/CRP.1/Rev.1) · Naciones Unidas — Convención sobre Ciertas Armas Convencionales 2026
  5. [290] Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for Congress · Congressional Research Service 2026 archived copy only
  6. [291] DOD Replicator Initiative: Background and Issues for Congress · Congressional Research Service 2026 archived copy only
  7. [526] UN Talks on Killer Robots Ends with Calls for Negotiations Growing · Human Rights Watch 2026
  8. [670] 'Lavender': The AI machine directing Israel's bombing spree in Gaza · +972 Magazine / Local Call 2024 archived copy only
  9. [668] 'A mass assassination factory': Inside Israel's calculated bombing of Gaza · +972 Magazine / Local Call 2023 archived copy only
  10. [298] Russia Probably Has Not Used AI-Enabled Weapons in Ukraine, but That Could Change · Center for Strategic and International Studies 2022
  11. [65] Law and Ethics for Autonomous Weapon Systems: Why a Ban Won't Work and How the Laws of War Can · Hoover Institution, Stanford University 2013

Ask OGERIA

It answers only with what the observatory publishes and can be wrong: check the entries it cites. Your questions are sent to an AI model, so don't write personal data. More in the privacy policy.

Up to 500 characters.

Support OGERIA on Ko-fi

The payment is processed by Ko-fi, not by this site. Open on ko-fi.com