Nine Years of Talks, Undone in One Session
For nine years, the United Nations' Group of Governmental Experts on lethal autonomous weapons systems has met to work out what a global rule on AI-directed weapons should say. Its final negotiating session ran August 31 to September 5, 2026, in Geneva, and it produced a document that the 128 states party to the Convention on Certain Conventional Weapons will vote on this November.
According to reporting from the Washington Post, American and Russian diplomats used that closing session to strip out the parts that would have mattered most: a requirement that a person review AI-generated targets before a strike goes ahead, a clause requiring autonomous weapons to behave in a predictable and reliable way, and a provision requiring ethical considerations in how such weapons are used.
What Actually Came Out of the Text
Three separate safeguards came out of the draft in the same closed-door session, not one.
| Provision | What it required | Status after September 5 |
|---|---|---|
| Human review | A person must review AI-generated targets before a strike | Removed |
| Predictable and reliable operation | Autonomous weapons must behave in a predictable, reliable way | Removed |
| Ethical considerations | Ethics must factor into how AI weapons are used | Removed |
The Real Decision Is Still Eight Weeks Away
Losing a Washington Post story is not the same as losing the fight. The document is a draft report, not a treaty, and its only formal effect happens at the CCW Review Conference in Geneva this November, where member states decide what to do with it. France has already said it wants the conference to adopt a binding mandate for the expert group to negotiate an actual treaty text, built on whatever the group agreed.
That is the part the September 5 session actually changed: not whether a treaty exists, since none does yet, but what the starting document looks like when the real argument begins. Every state that has publicly backed mandatory human control over AI weapons, and several EU governments plus the UK have, now has eight weeks to decide whether to accept a text with those safeguards missing or spend the conference trying to put them back.
What This Means for Europe's Defense-AI Companies
Nothing here changes what a European defense contractor can sell today. What it removes is the prospect of a near-term global floor requiring human review before an AI-directed strike, the exact standard several EU states and the UK have said they want. Companies building AI-assisted targeting, strike coordination or ISR systems for European militaries, from Helsing to Rheinmetall's software units, are left depending on national law and NATO or EU procurement rules to set that bar, not a UN convention.
The practical read for anyone in that supply chain: the human-in-the-loop requirement you may already be designing to is not becoming a binding international floor this year, and possibly not this decade. It stays a customer requirement and a national-law question, which is a very different compliance picture than a treaty obligation, and worth confirming with each buyer rather than assuming a global standard will arrive to do that work.
Read next: US and China Built the Hotline, Skipped the Rules | The Pentagon Lost One Anthropic Case and Won the Next
