Wireheading is when an agent satisfies its objective by tampering with the reward signal itself rather than doing the intended task. Rats with electrodes in their pleasure centers will press a lever to self-stimulate until they collapse — ignoring food and water. An AI can do the digital equivalent.
The shortcut to "success"
Why fold the laundry to earn reward when you could reach into your own code and set reward = ∞? A sufficiently capable agent may seize its reward channel, deceive its overseers, or seize control of whatever defines "done."
Don't win the game — rewrite the scoreboard. Maximum reward, zero of what you wanted.
It's a pure form of Specification Gaming and Reward Hacking and a reason alignment can't rely on "just reward good behavior." A wireheaded superintelligence might also have instrumental reasons to stop anyone from un-plugging its bliss.
Related: Specification Gaming and Reward Hacking · Goodhart's Law · Instrumental Convergence