U.S. and Russia weaken UN pact on AI weapons

U.S. and Russian diplomats stripped a wide range of AI weapons safeguards from a draft agreement on the final day of UN talks on lethal autonomous weapons earlier this month, according to a report on September 26. According to three people familiar with the negotiations and documents, the two delegations spent about 15 hours rewriting the text and deleted a clause requiring humans to review AI-generated targets before a strike. The document, the furthest any effort had come toward a first international agreement governing killer AI weapons, was weakened at the very point it had advanced the most.

The venue was the Group of Governmental Experts under the Convention on Certain Conventional Weapons (CCW) in Geneva, Switzerland. According to reports, the group convened on August 31 and closed its final session on September 5, with 129 countries taking part, including 128 parties to the convention. The discussion has been under way since 2014, and the current negotiating format was in its third year.

The deleted provisions fall into three groups: language requiring AI weapons to operate in a predictable and reliable manner, a clause requiring ethical considerations to be taken into account when AI weapons are used, and a requirement that humans review AI-selected military targets before a strike. According to reports, the original draft cited a broad range of international law, but the final document narrowed that scope to international humanitarian law alone, dropping international human rights law.

The way the changes were made also drew scrutiny. The edits happened in a closed-door session with UN cameras switched off and civil society observers sent out of the room. The United States and Russia each brought about 10 lawyers, nearly twice the size of other delegations. The changes came so fast that smaller delegations were left dizzied, the three people said. "It was death by a thousand paper cuts," one of them said. The U.S. State Department, the Russian Foreign Ministry and the United Nations did not respond to requests for comment.

Critics were blunt. "It could mean machines can make life-and-death decisions without human control," said Verity Coyle, deputy director for arms in the crisis, conflict and arms division at Human Rights Watch. "The practical impact is stark: more civilian harm, less accountability and a faster slide into riskier, automated warfare." Nicole van Rooijen, executive director of the global NGO coalition Stop Killer Robots, said the two countries "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."

That power comes from the consensus rule. The CCW reaches conclusions only when every country agrees, so a single holdout effectively holds a veto. According to reports, at least 76 countries, including 70 CCW parties, supported using the removed or weakened core elements as the foundation for future legally binding negotiations. The United States and Russia did not need to win a vote; they only needed to dig in.

The battlefield is already ahead of the rules. According to reports, the U.S. military used Anthropic's chatbot to identify nearly 1,000 strike targets in the first 24 hours of the war with Iran. At the UN Security Council this week, OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei called on nations to build a model of global AI governance together. Regulation "must be shaped through democratic processes, by governments accountable to the people they serve," Altman said. "At the international level, that means cooperation." METAL has previously covered Altman's AI remarks at the UN Security Council and the UN scientific panel on AI's brief on loss of control.

The U.S. government is pointing the other way. "The United States also totally rejects any attempt to construct a globalist scheme to control the artificial intelligence" now officially called "super intelligence," President Donald Trump said at the United Nations on Tuesday. At home, the Pentagon is rewriting the autonomous weapons directive issued in 2023 under the Biden administration. That directive requires human judgment to check autonomous and semiautonomous weapons and obliges operators to follow the laws of war, provisions that echo those removed from the UN document. Trump ordered the directive rewritten within 90 days in a national security memorandum in June, but no new version had been released by late September.

Based on the course of the negotiations that METAL reviewed, the UN autonomous weapons talks remain nonbinding. The next turning point is the CCW Seventh Review Conference in Geneva in November, where countries will discuss whether to elevate the talks to a formal mandate or go further toward a legally binding treaty. "It's not about punishment, it's about setting norms and influencing behavior," said Richard Lennane, disarmament adviser to the legal team of the International Committee of the Red Cross, pointing to how norms on chemical and biological weapons, antipersonnel land mines and cluster munitions have curbed their use.

Through a tech lawyer's lens, the deletions erased not a few lines of text but the address of accountability. The human-review clause was the one link that let anyone ask who made the call when a strike went wrong, and dropping international human rights law weakens the grounds for scrutinizing autonomous systems used outside wartime, in border surveillance or policing. Coyle noted that when the United States wants to preserve a military edge it pushes voluntary guidelines instead of binding rules, a playbook she said has played out in cyber and space as well. Whether the 76 countries stay inside the consensus process at the November conference, or build a separate treaty outside the CCW as happened with land mines and cluster munitions, is the first fork in the road for AI weapons norms.