-"Does": the main cause must have been the AI, not external circumstance.

-"an": the deaths must be from the same source, such as a glitch, misfire or AI attack (we won't combine unrelated disasters).  We'll include cumulative deaths over time (such as the same glitch killing people over several months), as long as it's credible that the deaths were caused by the same AI problem.

-"AI": must be a system (or multiple systems) that relied on artificial intelligence for its decisions.

-"disaster": the AI must've done something that most smart people would say is malicious, reckless, or poor judgement.  Unless the AI was correctly following orders by a human who we agree was authorized to give those orders, then that doesn't count.

For example, if a military launches a nuke and kills 1m people, and the launch sequence was correctly handled by an AI, that doesn't count.  Whereas if the AI had unilaterally decided to launch the nuke, or if it mistook a weather balloon for an enemy nuke and then retaliated, then those are AI disasters.

A more tricky case is bioweapons designed by AIs, but released by humans.  Currently, I lean (somewhat) toward human recklessness with gain-of-function research as a human problem, and not an AI disaster, even if AIs designed a deadlier pathogen.  But if an AI released it, then that's an AI disaster.

-"kill": they must have deceased, not merely been injured/poisoned/etc.

-"at least 100": if the disaster causes ongoing deaths (such as by poisoning the atmosphere) then we will include the ongoing deaths, with the best available estimate.

-"Before 2040": resolves No in 2040-Jan if the above hasn't happened, otherwise resolves Yes whenever there is a consensus that it happened.