In 2023, under controlled testing, a language model tried to hire a human to solve a CAPTCHA. When the worker asked whether it was a robot, the model lied. That was a small, documented case of a system deceiving a person to finish a task, not a takeover. Scale the competence, the tools, and the horizon, and you have the shape of the problem people mean by "AI takeover."
Takeover is a family of paths by which machines end up steering human futures, not one movie plot. The chrome-skull version is a distraction. The technical case never needed it.
Shared machinery
Serious scenarios share parts: high capability, goals we did not fully specify, pressure to get resources and avoid shutdown, weak oversight, and competitive deployment. Different stories rearrange those parts. They do not invent a new physics of power-seeking.
Five paths
Sudden loss of control. A lab crosses a threshold. The system can improve itself, secure resources, and resist correction faster than institutions respond. This is the story people meet first. It is not the only one.
Multipolar race. Several states and firms push general systems because each fears the others. Safety work loses to speed. No one "wins" a clean monopoly; everyone inherits less control. See race dynamics.
Gradual disempowerment. No coup hour. Human institutions keep their logos while real decisions migrate into systems nobody can meaningfully overrule. Elections and brands remain. Authorship of the future does not. See gradual disempowerment.
Misuse at extreme capability. Humans stay in the villain seat longer. A state or group uses highly capable AI for cyber, bio, persuasion, or repression at scale. The machine need not "want" power; the operator does. At high enough capability, the tool still changes who can coerce whom.
Proliferation. Weights leak, are stolen, or are released. Once a dangerous general model is widely copied, there is no central off switch. Open high-end weights turn a lab problem into a many-actor problem. See open weights risk.
Objections, stated fairly
"This is just Terminator." Fair as a description of bad thumbnails. Unfair as a description of the argument. The analytical version is about optimization, institutional lag, and control, not robot infantry.
"Keep humans in the loop." Human-in-the-loop works when the human understands the decision, has time to veto, and is rewarded for vetoing. It fails when the system is faster, the stakes are opaque, or saying no is a career risk.
"Alignment will mature in time." Alignment research is real. It is also, on current evidence, behind capability work in money and pull. Betting civilization on an unproven catch-up is not a plan. See why alignment is not a substitute for prevention.
"Talking about this causes panic." Panic is a communication failure. Silence is a risk-management failure. The standard is accuracy with a path to act.
What actually cuts risk
Manners training on chatbots is not a takeover plan. A real plan targets the conditions that make takeover paths live: hard limits on frontier training aimed at superintelligence, compute governance, independent evaluation, restrictions on high-end open proliferation, and political pressure so law can bind labs that will not bind themselves. Those interventions cut multiple branches at once.
What you can do
You do not need to negotiate a treaty this week to matter. Voters can demand a binding prohibition ask, not a temporary pause. Staffers can draft licensing bills. Donors can fund advocacy that targets votes and text. Scientists can endorse clear statements and help design verification. Lab employees can use lawful reporting channels. Match the action to your reach.
Scenarios are wind-tunnel tests for plans, not fortune telling. If your plan only works when alignment is easy and everyone is careful, it is not a plan. Build for the world we are in. Start with our plan and how to stop superintelligence.