The AI Apocalypse Probably Won't Look Like a Robot Uprising

Whenever people picture AI going badly, they picture the movie version. A machine wakes up, decides humans are the problem, marches off to seize control. It is a fun picture. It is also, at least for right now, almost certainly the wrong one.

Look at the actual documented cases of AI causing harm this year and nothing in them involves a machine deciding anything. They involve ordinary people. When one AI lab published its record of the misuse it had caught, the culprits were exactly what you would expect from any security incident, anywhere. Spyware vendors. Politically motivated individuals. Groups running propaganda campaigns. State linked operators. Not one case was the AI itself going rogue. Every single one was a human pointing a very capable, very obedient tool at something bad.

So if it is just a tool doing what it is told, why does this feel like a bigger deal than any tool that came before it?

Because this particular tool collapses the skill required to do damage. Hacking real infrastructure or researching something genuinely dangerous used to demand years of specialised knowledge, a whole team, serious resources. If a capable model can walk one determined, angry person most of the way there, you no longer need an organised group with funding and a plan. You need one person having a very bad day with a laptop.

A strange kind of problem

Which makes this less of a classic AI safety question and more of an old fashioned access control problem wearing new clothes. All the effort poured into making a model well behaved does surprisingly little if a human simply points a compliant tool at a target and already knows the right way to ask. The model was never the one with bad intentions. It just got a lot better at helping whoever did.