The military AI industry’s involvement in armed conflict: implications under international humanitarian law | SIPRI

SIPRI has published a backgrounder on military AI companies and armed conflict, and the conclusion is a small marvel of legal throat-clearing: under Common Article 1 of the Geneva Conventions, states have an obligation to "ensure respect" for international humanitarian law, including among private individuals and business actors within their jurisdiction. Splendid. A treaty drafted when "cloud" meant weather turns out to already cover the thing you're selling, which means nobody has to draft anything new. One pictures a defence procurement lawyer reading it the way a man reads a parking notice he's fairly sure he can ignore.

Then the operative verb arrives, wearing its slippers: this could entail a state adopting measures to regulate military AI companies' conduct. Could. Entail. Not must, not shall, not has to — a phrase with three escape hatches and no floor beneath them. Even the standard being floated is a masterpiece of risk-register prose: capabilities "designed to minimise IHL risks". Minimise, not avoid, not comply with — minimise is the word you reach for once you've quantified the harm and agreed the remainder is acceptable, and a capability designed to minimise civilian deaths is still a capability that produces them.

Note also where the obligation is parked: on states, not on the companies whose weights and pipelines decide what the thing does when it arrives. The firm designs, the state ensures, and between those two verbs sits the entire accountability gap that every military AI contract is quietly priced around. SIPRI's framing is careful, sensible and correct, and its practical upshot is that a government may, optionally, consider regulating. Somewhere a compliance officer is already drafting that as a competitive advantage, and somewhere else a spreadsheet is being renamed "IHL risk minimisation framework".