More than 130 civil society organisations and over 160 individuals, many of them current or former employees of Google, OpenAI and Google DeepMind, have signed a joint statement demanding that technology companies and governments halt the use of artificial intelligence in military kill chains, warning that AI-driven warfare is eroding legal accountability and enabling mass killing at unprecedented speed.
The statement, published by Access Now and co-signed by Amnesty International, the Association for Progressive Communications, Article 36, the Campaign to Stop Killer Robots and dozens of other digital rights, peace and human rights groups, says AI systems embedded in military targeting are accelerating the scale of assaults in ways that risk violating international humanitarian law, international criminal law and human rights law.
The signatories call on tech companies to refrain from signing or fulfilling contracts with military agencies or armed groups implicated in violations of international law, and to stop selling, transferring or exporting AI decision-support systems used for kill-chain targeting and human surveillance, including target-generation tools and remote biometric surveillance.
They also want companies to withhold non-lethal AI systems, including large language models, from military decision-making until genuine human oversight, accountability and transparency become possible. States, the statement says, should halt the use of AI tools including LLMs in military targeting and disclose how such technology is currently being used in hostilities.
The statement names several companies whose products, according to media investigations, have played a role in recent conflicts. It cites reporting that Anthropic's Claude and Palantir's Maven Smart System supported United States and Israeli military planning during strikes on Iran, and notes that OpenAI recently agreed to supply AI services to the US Department of Defense while Google, Microsoft and Amazon have long-standing cloud, data and "warfighting" infrastructure contracts with the Pentagon.
It says nearly 2,000 targets were struck in Iran within 48 hours of the campaign beginning, with reporting attributing much of that speed to AI-assisted target generation, and cites accounts of civilian and infrastructure casualties that followed.
The statement draws a direct line to Israel's use of AI-assisted targeting systems in Gaza, referencing tools reported in the press as Lavender, Gospel and Where's Daddy, and cites a UN commission finding that genocide has been committed in the territory, alongside reporting that Palestinian deaths have surpassed 72,000.
It argues that framing algorithmic recommendations as objective allows human responsibility for lethal decisions to be diluted and obscured, and says Microsoft, Google, Palantir and other firms may have enabled Israeli military access to the data infrastructure underpinning these operations.
The organisations reject the idea that "human in the loop" arrangements or built-in technical safeguards can currently prevent the harms of AI-accelerated warfare, arguing that speed and scale make meaningful human review impossible in practice and risk reducing it to a rubber stamp.
They quote Anthropic's own public statement that "today, frontier AI systems are simply not reliable enough" (accessnow) to power fully autonomous weapons, while arguing this caution has not stopped major firms from deepening military partnerships. The statement also notes that Anthropic has been in a reported standoff with the US government over use restrictions, that Google and Amazon have reportedly suspended parts of their terms of service in contracts with Israel, and that more than 560 Google employees signed an open letter in April 2026 urging the company's leadership to refuse government use of its technology in classified military operations.
The statement discloses that Amnesty International wrote separately to OpenAI and Anthropic seeking clarity on their human rights policies regarding military use of generative AI; only OpenAI had responded at the time of publication.
Comments
Post a Comment
NOTE: Hateful, abusive comments won't be published. -- Editor