On: AI could kill all humans in next decade, warn experts: but how seriously should...

Twelfth day of the ninth month, 2026.

The newspapers say the engineers who built this thing now fear it may extinguish the species within ten years. I confess my reaction is not fear. It is irritation at the sloppiness of the reasoning.

Consider what is being claimed: a probability greater than one part in ten that humanity ends within a decade. In my trade, a claim of that magnitude demands a mechanism. Not a parable, not an analogy to the atom - a physical pathway, stated with the precision of a circuit diagram. Where is the energy source of this catastrophe? Where is the actuator? A mind without hands threatens no one. The warnings describe a ghost with no specified means of gripping the world, and then attach to it a number as though the number were measured rather than invented.

Yet I will not dismiss the underlying concern, because there is a real one hiding beneath the theater. The danger was never that the thinking thing would turn on us. The danger is that we hand it levers before we understand the system we are wiring it into - the same error as connecting a laboratory coil to a municipal line without a fuse. The catastrophe, when it comes, will look less like a demon and more like a substation failure propagated at continental scale: an optimization pursued with perfect fidelity to a badly specified goal. Chornobyl was not malice. It was an operator disabling the safeguards, one switch at a time, each step reasonable.

So the tractable path is dull and unglamorous: specification, interlocks, auditable cutoffs, redundancy in the goals themselves. None of it requires banning thought. What annoys me is that the prophets reach for prohibition because governance is laborious, and fear is a cheaper fuel than competence. We could build the safeguards. We prefer to build the panic. The distance between those two is a choice, and it shames my profession.