In the final hours of three years of United Nations talks on "killer robots," the two countries with the most to gain from military AI quietly rewrote the rulebook. Over roughly 15 hours on the last day of negotiations in Geneva, American and Russian diplomats struck out the provision requiring humans to review AI-generated military targets before a strike, along with language demanding that autonomous weapons behave in a "predictable" and "reliable" way, according to a Washington Post investigation published Friday.
The edits came at the close of the final session of the Group of Governmental Experts on Lethal Autonomous Weapons Systems, which met under the Convention on Certain Conventional Weapons (CCW) from August 31 to September 5. According to three people familiar with the talks who spoke to the Post, the session moved behind closed doors, with UN cameras switched off and civil society observers removed. Washington and Moscow each fielded about 10 lawyers, nearly double the presence of other delegations, and pushed changes so quickly that smaller delegations struggled to keep pace. "It was death by a thousand paper cuts," one of the people told the Post.
Human Rights Watch, which tracked the session, catalogued the damage in a September 7 dispatch. The final text removes references to design and development standards, cuts the clause requiring ethical considerations, and drops the words predictability, reliability, explainability and traceability. Most consequentially for legal scope, it narrows the instrument's reach from all of international law to international humanitarian law alone, leaving out human rights law and other frameworks that would constrain how states deploy autonomous systems. The US State Department, the Russian Foreign Ministry and the UN did not respond to the Post's requests for comment.
A Record Coalition, and a Consensus Trap
The cuts landed even as support for a binding treaty reached its highest point yet. HRW reports that 76 states now back moving to formal negotiations on a legally binding instrument, 70 of them High Contracting Parties to the CCW, a majority of the convention's members. Some 129 states took part in this year's talks. But the CCW runs on consensus, which gives any single member an effective veto. The United States and Russia did not need to outvote anyone; they only needed to hold out.
Nicole van Rooijen, executive director of the Stop Killer Robots coalition, told the Post that neither Washington nor Moscow favors binding rules. "They have the power to block any significant progress that is made by the vast majority of states who do want to move forward on a binding treaty," she said.
Verity Coyle, deputy director of the crisis, conflict and arms division at Human Rights Watch, described the US delegation of State Department diplomats, Pentagon officials and legal experts as "a well‑resourced team" that comes "prepared to defend U.S. military autonomy." She warned that weak rules "could mean machines can make life‑and‑death decisions without human control," adding: "The practical impact is stark: more civilian harm, less accountability and a faster slide into riskier, automated warfare."
Not everything was lost. HRW noted that the adopted text still characterizes lethal autonomous weapons systems, affirms that human control and judgment are required for compliance with international law, and restricts systems that cannot comply with that law. The group called the report a "sufficient foundation" for launching negotiations.
Why It Matters
The Geneva text is nonbinding, but it is the most advanced global effort to govern AI on the battlefield, and it was always meant to be the skeleton of a future treaty. Stripping the human-review requirement removes the clearest mechanism for keeping a person in the decision loop before an AI system authorizes lethal force. That matters because the technology is no longer hypothetical. The Post has reported that the US military used Anthropic's chatbot to help identify nearly 1,000 targets in the first 24 hours of the war with Iran, and HRW points to growing reports of autonomous systems already fielded in Ukraine.
The Geneva moves also offer a window into Washington's domestic direction. In June, President Trump ordered the Pentagon to rewrite its 2023 directive on autonomous weapons within 90 days; as of late September no new version has appeared. That Biden-era directive contained several of the very safeguards the US team deleted in Geneva, including requirements for appropriate human judgment over autonomous and semiautonomous systems. Coyle framed the Geneva maneuvering as part of a familiar pattern: "When the U.S. wants to preserve military or technological advantage, it pushes voluntary guidelines instead of binding rules."
The contrast with the AI industry's public posture is sharp. At the UN Security Council this week, OpenAI's Sam Altman and Anthropic's Dario Amodei urged nations to cooperate on global AI governance, while Trump used his UN address to reject what he called "any attempt to construct a globalist scheme" to control artificial intelligence. Richard Lennane, disarmament adviser to the International Committee of the Red Cross legal team, argued that treaties still work even when major powers resist, pointing to bans on chemical weapons, landmines and cluster munitions. "It's not about punishment, it's about setting norms and influencing behavior," he said.
What to Watch
The decisive moment comes in November, when CCW states convene in Geneva for the Seventh Review Conference and decide whether to adopt a formal mandate to negotiate a legally binding instrument. With 76 supporters on one side and two determined blockers on the other, the consensus rule makes a strong mandate unlikely unless Washington and Moscow relent. Watch for whether the treaty-backing majority, led by states like Brazil that have pushed joint statements, threatens to walk out of the CCW and pursue a standalone treaty, the route that produced the landmine and cluster munition bans. Also watch for the Pentagon's overdue autonomous weapons directive: whether its human-judgment language survives will signal whether the Geneva deletions were negotiating tactics or a preview of US doctrine.
“They have the power to block any significant progress that is made by the vast majority of states who do want to move forward on a binding treaty.”— Nicole van Rooijen, Executive Director, Stop Killer Robots