When public debate asks whether an autonomous system can be trusted — a diagnostic algorithm, a military platform, infrastructure control — it almost always collapses into one question: is there still a human in the loop? This lecture argues that the question is necessary but not sufficient, because presence and meaningful power are two different questions that governance has been treating as one.

It opens by tracing where the human-in-the-loop requirement actually came from — not from research showing that human operators decide better under pressure, but from institutional discomfort: no military, medical, or regulatory body was prepared to publicly own a system that decides entirely on its own. The same pattern is shown reproducing itself independently across aviation, medicine, and military autonomy, each producing a human presence that functions as a mechanism for distributing accountability rather than guaranteeing the quality of control.

From there the lecture builds its structural argument in three moves. First, it decomposes "control" into four separable capacities — situational understanding, time, independence, and the ability to reverse — any of which can be eroded while formal compliance is preserved. Second, it reframes delegation not as a binary switch but as a spectrum, on which a system can be labeled human-in-the-loop at almost any point. Third, using a blocked bank transaction and the Soviet-era "Dead Hand" system as parallel cases, it shows that a human's operative mandate is never unconditional — it is a state the system itself grants and revokes, by criteria the person whose mandate it governs never participated in setting.

The lecture closes by proposing an alternative object of analysis — delegation architecture: seven structural questions about who initiates, how confirmation works, what override actually costs, who sees what information, who holds delay authority, how authority transfers on failure, and who decides whose decision counts. It ends on an open reversal rather than a set of recommendations: systems are emerging that evaluate the operator's own reliability, meaning the right to doubt — long assumed inalienable — is becoming something the architecture can extend or withhold.