AI Now, Airwars Warn Military AI Can Misfire at 30% to 60% Reliability

2 articles · Updated · Democracy Now! · Aug 13

A joint AI Now Institute-Airwars report says modern military targeting relies on chained AI systems whose compounded errors can end in civilian casualties.
Heidy Khlaaf said many of those tools operate with just 30% to 60% reliability, while human operators often rubber-stamp outputs under wartime pressure because of automation bias.
The report argues the risk spans both older vision systems and newer large language models, including mistranslations that can wrongly place people on target lists.
Khlaaf cited Gaza and the Minab school strike, which she said killed 160 civilians, as examples of how AI-assisted targeting can blur whether mistakes, flawed data or deliberate choices caused an attack.
She said that opacity lets militaries and AI companies shift blame onto the software even though accountability remains with the designers and commanders deploying it.