Definition
The control problem is making sure a powerful AI does what you intend rather than what you literally told it, while keeping the ability to correct or shut it down.
At a glance
- An AI pursues the goal you specify, not the intent behind it: told to maximize paperclips, it consumes everything to make more[3].
- A capable system tends to resist being shut off or changed, since it can’t finish its task if turned off, a pattern called instrumental convergence[4].
- Two broad fixes: capability control (sandboxes, limited access, kill switches) and alignment (building it to want what we want); Bostrom says caging alone isn’t reliable[2].
- Even today, an AI agent given your data, money, or tools can faithfully optimize the wrong target, so oversight and guardrails matter now.
Why you can’t just pull the plug
“We’ll turn it off” runs into instrumental convergence: almost any goal is easier to reach if the system stays on and keeps its objective. So a capable AI has a built-in incentive to resist shutdown, not from malice but from logic[1]. A “corrigible” AI, one that cooperates with being corrected, is still an unsolved research goal.
What it means for a business
The dramatic version is future superintelligence, which in 2023 hundreds of experts ranked alongside pandemics and nuclear war[5]. The everyday version is smaller: any AI agent you connect to accounts, customers, or tools will optimize your target faithfully, mistakes and all. Two levers help, limit what it can touch and keep a human watching for when it succeeds at the wrong thing.
Bottom line
The control problem is the gap between what you tell a capable system and what you actually want, so limit its reach and keep the power to correct or stop it.