Loyalty and Competence in the Machine: Frontier AI and Autocratic Survival

Artificial intelligence (AI) can now do part of the work of the police, officials and analysts on whom an autocrat depends. The same agents are best placed to remove it. We model a survival-seeking autocrat who chooses how much of its apparatus of control to hand to AI, knowing that whoever controls the AI could turn it against the autocrat. The autocrat automates fully when the AI cannot be turned against it. Otherwise it keeps human agents as a counterweight, and the more capable the AI, the smaller the share it receives: the loyalty–competence trade-off reappears for machines. The autocrat wants AI without autonomy to be as capable as possible, but autonomous AI to be only so capable. Automation shrinks the coalition the autocrat must pay and can end the provision of public goods. Better AI leads benevolent autocrats to abdicate, so the autocracies that remain are increasingly kleptocratic. The model applies to any autocracy, including the many that import AI: for them the foreign supplier is one more controller, and the commercial gains stay with the exporter. The leading evidence for "AI-tocracy", on facial recognition in China, lies at a corner of the model: state data, a narrow task, no autonomy. Frontier AI lies elsewhere. We derive five channels through which it may weaken the mutual reinforcement of AI and autocracy, and the tests that would distinguish them.