By Tom Morgan, Founder & Editor, AInvasion · Coverage: Defense Technology & AI Policy Last updated: October 6, 2026 · Next update: after the CCW Seventh Review Conference (November 16–20, 2026)
How we source this: every serious claim is linked to a primary document, a named institution, and a date. Claims that rest on one outlet, on anonymous sources, or on an advocacy tracker are labeled in the text. The claim ledger is at the end.
Contents
- The short answer
- How to read the confidence labels
- Where “human judgment” breaks down
- The replicator is gone. DAWG and the Collaborative Combat Aircraft replaced it
- Ukraine did not choose autonomy. Jamming pushed it
- Geneva: a text, not a treaty
- Checklist for evaluating military-AI claims
- FAQ
- Claim ledger and sources
1. The short answer
- No treaty bans autonomous weapons today. On September 5, 2026, the CCW expert group (GGE) adopted by consensus a non-binding set of elements for a possible instrument (Reuters, Sept 5).
- U.S. policy does not require a human to approve every engagement. DoD Directive 3000.09 (effective January 25, 2023) requires that systems let commanders and operators exercise “appropriate levels of human judgment over the use of force” (Section 1.2.a). That is a design standard, not a fixed review rule.
- In Ukraine, autonomy is spreading because of electronic warfare. CSIS found that autonomous navigation raised reported strike success from roughly 10–20 percent to roughly 70–80 percent, while humans kept the engagement decision (CSIS, March 6, 2025).
- The next decision point is Geneva, November 16–20, 2026. States must decide whether and how to continue the work.
2. How to read the confidence labels
| Label | Meaning |
|---|---|
| ESTABLISHED | Supported by a primary document: policy text, official release, budget record, UN record |
| PROBABLE | Credible reporting that is single-source, anonymous-source, or not independently verified |
| DISPUTED | The party concerned contests the reporting; both positions are given. |
The labels measure source strength, not editorial opinion on the underlying debate.
3. Where “human judgment” breaks down
Most coverage treats “human in the loop” as binary: a person approves the strike, or a machine does. The U.S. directive and real systems are more granular.
What the directive says and excludes. DoD Directive 3000.09 requires systems to be designed so commanders and operators can exercise “appropriate levels of human judgment over the use of force.” It defines an autonomous weapon system as one that, once activated, can select and engage targets without further operator intervention. It excludes autonomous cyberspace capabilities from its scope, and it exempts certain categories from senior review, including operator-supervised systems that defend installations, platforms, or deployed unmanned vehicles against time-critical or saturation attacks (Sections 1.1.b and 1.2.d). [ESTABLISHED: U.S. DoD, January 25, 2023]
Editorial analysis: the definitions matter more than the slogans. Under the directive’s own glossary, a weapon that uses autonomy for tracking, prioritizing, or terminal guidance onto a target an operator has selected is semi-autonomous, not autonomous. Many claims about “autonomous drones” describe exactly that, and the policy treatment differs.
The clearest public case of thin oversight. An investigation by +972 Magazine and Local Call, published April 3, 2024, reported that the Israeli military used an AI system called Lavender to mark suspected militants in Gaza. According to six Israeli intelligence officers, the system marked as many as 37,000 Palestinians at its peak, a figure attributed to four of them. One source said officers often spent about 20 seconds per target, mainly to check that the target was male. Sources also said internal checks put the system’s error rate at roughly 10 percent, based on a manual check of a sample of several hundred targets (+972 Magazine). [PROBABLE: anonymous-source investigative reporting; figures not independently audited]
The IDF disputes this. In statements quoted in the same investigation, the IDF denied the existence of such a kill list and said its systems are auxiliary tools, with an independent analyst examination required to verify that targets are legitimate. [DISPUTED]
Why it matters. If the reporting is accurate, it describes what critics call rubber-stamp autonomy: a system that is formally supervised but built so that meaningful review is impractical at scale. That is a documented critique, not a settled legal finding.
Apply it yourself. If a vendor or agency claims “human oversight” but cannot tell you the average review time per decision, the claim cannot be verified.
4. The replicator is gone. DAWG and the Collaborative Combat Aircraft replaced it.
The promise. Deputy Secretary of Defense Kathleen Hicks announced Replicator on August 28, 2023, with a goal to field attritable autonomous systems at a scale of multiple thousands in multiple domains within 18 to 24 months (DefenseScoop, Aug 28, 2023; CSET analysis). [ESTABLISHED]
The shortfall. The Congressional Research Service reported that although Replicator was meant to field thousands of uncrewed systems by summer 2025, one former defense official said only hundreds had been fielded by then (CRS IF12611, versions of August 25 and September 19, 2025). This is CRS relaying a former official’s account, not a CRS count. [PROBABLE]
The handover. The Wall Street Journal reported on September 26, 2025, that Replicator was being moved to a new Defense Autonomous Warfare Group (DAWG) under Special Operations Command, after setbacks that included difficulty buying software able to command large numbers of different drones (summarized by Responsible Statecraft). In 2026, Defense Trade Press describes DAWG as having absorbed Replicator (Breaking Defense, April 2026; Task & Purpose). [PROBABLE: reported by news outlets]
The money. DAWG received $225.9 million in fiscal 2026. The FY2027 request seeks up to $54.6 billion for it in research and development funding (DefenseScoop, April 21, 2026; DefenseScoop, April 29, 2026). Breaking Defense reports that about $1 billion sits in the base budget and $53.6 billion depends on a future reconciliation bill (Breaking Defense). [ESTABLISHED as a request; not appropriated]
The command. Defense Secretary Pete Hegseth told the House Armed Services Committee on April 29, 2026, that a sub-unified command for autonomous warfare would be announced shortly (DefenseScoop). [ESTABLISHED: congressional testimony]
Collaborative Combat Aircraft (CCA). The Air Force announced on June 18, 2026, that it awarded production contracts for Increment 1 to General Atomics (FQ-42) and Anduril (FQ-44), four months ahead of schedule. Mission-autonomy software contracts went to a pool of six vendors (Anduril, General Atomics, Lockheed Martin, Northrop Grumman, RTX Collins Aerospace, and Shield AI), with production options for Anduril, RTX Collins, and Shield AI. The Air Force plans to select a primary autonomy provider by summer 2027 (U.S. Air Force, AFLCMC; Military Times, June 18, 2026). [ESTABLISHED]
| Program | Core focus | Scale / budget signal | Status (Oct 2026) |
|---|---|---|---|
| Replicator (2023–2025) | Attritable systems across domains | Goal: multiple thousands; reported: hundreds (CRS, citing a former official) | Absorbed into DAWG |
| DAWG (2025–present) | Institutionalized autonomous-systems acquisition | FY26: $225.9M; FY27 request: up to $54.6B, mostly via reconciliation | Active; request pending in Congress |
| CCA (Air Force) | Autonomous wingmen for crewed aircraft | Increment 1 production awarded June 18, 2026 | Autonomy competition runs to summer 2027 |
5. Ukraine did not choose autonomy. Jamming pushed it.
The best counter to “autonomy is a policy choice states can defer” is Ukraine, where autonomous navigation answers a physics problem.
The data. A CSIS report by Kateryna Bondar (March 6, 2025) found that drones with autonomous navigation raise the target-engagement success rate from around 10 to 20 percent to around 70 to 80 percent because they no longer depend on constant manual control and stable communications, both vulnerable to electronic warfare. Frontline operators interviewed by CSIS put typical FPV mission success at only about 10 to 15 percent and 30 to 50 percent for highly skilled operators (CSIS). [ESTABLISHED as a CSIS finding based on interviews; treat the figures as estimates, not audited statistics]
What CSIS did not find. The same report said full autonomy, as the U.S. military defines it, was not yet present on the Ukrainian battlefield, and that engagement decisions remained with humans (CSIS). That was the position in early 2025.
What has been claimed since?
- Serhii “Flash” Beskrestnov, an adviser to Ukraine’s defense minister, wrote in June 2026 that target prioritization, selection, and autonomous engagement decisions are the near future for strike drones. He added that autonomous target-selection systems are still in early combat testing in both Ukraine and Russia (Ukrainska Pravda, June 24, 2026). [PROBABLE: his own public statement on Telegram, reported by a national outlet]
- A report that surfaced publicly on June 10, 2026, claimed that Ukrainian drones in a “Terminator mode” test killed Russian soldiers without human oversight near Chasiv Yar. According to Small Wars Journal, the claim originates with Ukrainian defense-industry figure Alexander Kokhanovskyy; the test reportedly took place around mid-2024, and Kyiv has not confirmed it (Small Wars Journal, August 17, 2026). [PROBABLE: single-source claim, unconfirmed by the Ukrainian government]
Editorial analysis: precision is not the selling point. Ukrainian forces plan around a “reasonable radius,” not a pinpoint, because autonomous navigation mainly lowers the skill and communications demands on operators, according to CSIS. That is a different risk profile from the “surgical AI strike” narrative common in defense marketing.
| Framework | Human-control standard | Enforcement | Status (Oct 2026) |
|---|---|---|---|
| U.S. DoD Directive 3000.09 | “Appropriate levels of human judgment”; senior review for most autonomous weapons | Internal DoD review; waivers available | Binding on U.S. forces; no external audit |
| CCW GGE text (Sept 5, 2026) | Non-binding elements of a possible instrument | None; adopted by consensus without legal force | Goes to the Seventh Review Conference |
| Ukraine battlefield practice | Autonomy for navigation and terminal guidance; engagement decisions human (per CSIS, 2025) | Operational necessity | Claims of more autonomy since, unconfirmed |
| Israel (IDF statements) | Analyst examination required; disputed by +972 sources | Internal military process | No independent audit |
6. Geneva: a text, not a treaty
The deadline. In October 2023, UN Secretary-General António Guterres and ICRC President Mirjana Spoljaric jointly called on states to launch negotiations of a new legally binding instrument on autonomous weapons and conclude them by 2026 (UN News, October 2023). They renewed the call on August 25, 2026 (UN Geneva newsroom). [ESTABLISHED]
What happened on September 5. At the GGE’s final session (August 31 to September 4, running overnight into September 5), the 128 High Contracting Parties to the CCW reached consensus on a non-binding set of elements. Reuters had reported on September 4 that the United States and Russia proposed late changes and deletions to the draft (Defense News/Reuters, Sept 4; Reuters, Sept 5; Opinio Juris). [ESTABLISHED that consensus was reached; the text is non-binding]
How much support exists. Automated Decision Research, a Stop Killer Robots initiative, counts 76 states (70 of them CCW parties) that have backed moving to negotiations on the basis of the GGE text, and says only 12 states have explicitly opposed a legally binding instrument (ADR, Sept 4, 2026). That includes a September 2025 joint statement delivered by Brazil on behalf of 42 countries. [PROBABLE: advocacy tracker] At the UN General Assembly’s First Committee, a resolution on autonomous weapons passed on November 6, 2025 with 156 votes in favor and five against (the United States, Israel, Belarus, Russia and North Korea), and the General Assembly adopted it as resolution 80/57 on December 1, 2025 (RSIS; UN resolution record). [ESTABLISHED]
What comes next. The Seventh Review Conference meets in Geneva on November 16–20, 2026, as decided by the CCW’s 2025 meeting of High Contracting Parties (UN document CCW/MSP/2025/CRP.1). The GGE’s mandate ends this year, and states must decide whether and how to continue the work (ADR). The CCW works by consensus, so a single state can block a decision (Just Security backgrounder). [ESTABLISHED]
Editorial analysis (our view, not a sourced finding): the real contest is no longer whether autonomy gets rules. It is whether the rules are binding, and who writes the definitions that decide what counts as “autonomous.”
7. Checklist for evaluating military-AI claims
For analysts, journalists and procurement staff, not for building or deploying systems.
| # | Ask | Why it matters |
|---|---|---|
| 1 | What is the average human review time per AI-flagged decision? | Seconds per decision at scale suggests formal, not real, oversight |
| 2 | Is the error rate self-reported or independently audited? | Operator-reported figures are directional, not verified |
| 3 | Which policy actually applies: 3000.09, a vendor policy, or none? | “AI-powered” and “policy-compliant” are different claims |
| 4 | Is it navigation, terminal guidance, target selection, or engagement autonomy? | Under 3000.09’s glossary, terminal guidance on an operator-selected target is semi-autonomous |
| 5 | What did the Review Conference decide: mandate, extension, or lapse? | Determines whether any external accountability track exists through 2027 |
8. FAQ
Does the U.S. military have fully autonomous weapons that can kill without human approval? Not as a declared policy. DoD Directive 3000.09 requires “appropriate levels of human judgment over the use of force,” which is a design and review standard, not approval of every engagement. Operator-supervised defensive systems against time-critical or saturation attacks are exempt from the senior-review requirement, and autonomous cyber capabilities fall outside the directive entirely.
What happened to Replicator? It fell short of its August 2025 goal, with only hundreds of systems reportedly fielded, according to CRS citing a former defense official. Its work moved to the Defense Autonomous Warfare Group, which has a FY2027 request of up to $54.6 billion, most of it dependent on a reconciliation bill.
Did the UN ban autonomous weapons in 2026? No. On September 5, 2026, the CCW expert group agreed a non-binding text. The Seventh Review Conference (November 16–20, 2026) decides what happens next.
What is “meaningful human control,” and why is it disputed? It is a proposed standard that humans keep real, informed authority over lethal decisions, not just formal sign-off. It is disputed because states have not agreed on measurable criteria, such as review time, information provided, or ability to override, that would let outsiders verify it.
9. Claim ledger and sources
| Claim | Source | Institution | Date | Type |
|---|---|---|---|---|
| “Appropriate levels of human judgment”; cyber excluded; senior-review exemptions | DoDD 3000.09, §1.1.b, 1.2.a, 1.2.d | U.S. DoD | Jan 25, 2023 | Primary document |
| Lavender: 37,000 marked; ~20 sec review; ~10% error | +972 Magazine / Local Call | +972 Magazine | Apr 3, 2024 | Investigation; anonymous sources; disputed by IDF |
| Replicator announced; “multiple thousands” in 18–24 months | DefenseScoop | DoD (Hicks speech) | Aug 28, 2023 | Reporting of official speech |
| Only “hundreds” fielded by summer 2025 | CRS IF12611 | Congressional Research Service | Aug 25, 2025 and later versions | Official report citing a former official |
| Replicator moved to DAWG | Responsible Statecraft on WSJ; Breaking Defense | WSJ; Breaking Defense | Sept 26, 2025; Apr 2026 | Reporting |
| DAWG FY26 $225.9M; FY27 request up to $54.6B | DefenseScoop | DoD budget documents, via DefenseScoop | Apr 21, 2026 | Request, not appropriation |
| $1B base / $53.6B reconciliation split | Breaking Defense | Breaking Defense | Apr 2026 | Reporting |
| Sub-unified command for autonomous warfare to be announced | DefenseScoop | Secretary of Defense testimony | Apr 29, 2026 | Congressional testimony |
| CCA Increment 1 production; six-vendor autonomy pool; selection by summer 2027 | AFLCMC | U.S. Air Force | Jun 18, 2026 | Official release |
| 70–80% vs 10–20% success; human oversight of engagement | CSIS, Bondar | CSIS | Mar 6, 2025 | Report; interview-based estimates |
| Autonomous target selection “near future” | Ukrainska Pravda | Beskrestnov (Telegram) | Jun 24, 2026 | Public statement |
| “Terminator mode” test near Chasiv Yar | Small Wars Journal | SWJ (G. Mako) | Aug 17, 2026 | Single-source claim; unconfirmed |
| UN/ICRC call for binding instrument by 2026 | UN News | UN / ICRC | Oct 2023 | Official appeal |
| Renewed call | UN Geneva newsroom | UN / ICRC | Aug 25, 2026 | Official appeal |
| GGE consensus on non-binding text; late U.S./Russian changes | Defense News/Reuters; Reuters, Sept 5 | Reuters | Sept 4–5, 2026 | Reporting |
| 76 states back negotiations; 42-state statement | ADR / Stop Killer Robots | Automated Decision Research | Sept 4, 2026 | Advocacy tracker |
| UNGA First Committee 156–5; resolution 80/57 | RSIS; UN record | UN General Assembly | Nov 6 and Dec 1, 2025 | Official record and analysis |
| Review Conference, Nov 16–20, 2026 | CCW/MSP/2025/CRP.1 | CCW High Contracting Parties | Nov 12, 2025 | Primary document |
Notes on method
Sources are ranked: primary documents and official records first, then original investigations and datasets, then reputable secondary reporting, then advocacy trackers and aggregators. Anything from the last two groups that carries a number is labeled in the text. Dates reflect the document or publication date.