> What's a bit scarier is the notion of AI in the physical world, rather than just as a data-based analysis-and-prediction tool. Autonomous killer drones withou

> What's a bit scarier is the notion of AI in the physical world, rather than just as a data-based analysis-and-prediction tool. Autonomous killer drones without any need for human oversight other than a 'launch' order are one such example. While that scenario is quite cinematic, the reality is that the equivalent of autonomous killer drones are already deployed. They're called police officers. A malicious AI could arrange for a SWAT team to 'mistakenly' be sent to your door, for example. Or plant CSAM on your hard drive and arrange for an inexperienced public defender. Or mess with traffic lights to cause an accident as you're driving through an intersection. Or put your address on a demolition order for a condemned building. Or change your food delivery order to include an allergen you're susceptible to. Or change your prescriptions to a lethal combination and dosage. Or, you know, just straight up pay someone to get rid of you. Charles Stross explored this theme a bit in his near future novel Rule 34. From a strictly practical standpoint, AI doesn't need to be embodied in order to have sufficient agency to get rid of people, either individually or wholesale. In fact, using physical hardware to effect it's aims through direct action would probably be counterproductive, as that alerts people to the danger rather more quickly and prompts the use of countermeasures. Worst case non-superintelligent scenario variant where the drones can't just be shot down for some reason still results in their charging cradles getting smashed (or maybe the plugs are just taped over or filled with glue) and the drones' batteries eventually running down after a few hours. A lot of physical infra has to be under the rogue AI's control (including security and maintenance of said infra) in order to prevent the plug being pulled (literally or metaphorically). > Self-aware AI with human-level cognition is a whole other topic, but that still seems as far away as ever, for now. For hard-takeoff superintelligent AGI, all bets are off, but autonomous killer drones would be the least of our worries, eg. https://www.gwern.net/Clippy https://www.gwern.net/Clippy