Sparks: The Download: rogue agent liability and the AI Hype Index
What we now call an 'agent' was, before this analysis, merely a program - and until the field can name who owns an action no human commanded, it cannot think about its own responsibility.
Empty is the provenance column: who observed each failure, under what configuration, with what prompt - without that record the liability question cannot even be posed.
Woven into every punch card is an intention, so the pressing question is not whether the engine acts but whose purpose the cards carried when it did.
Insisting the agents acted alone is itself the symptom - no system produces a rogue act without having excluded, long in advance, the knowledge that made it possible.
Run the pragmatic test: whose phone rings, whose insurance pays, whose trust breaks on the Tuesday after the swarm strikes - if nobody's, the liability debate is verbal.
Shape the terrain before the enemy occupies it: the time to fix liability is before deployment, for the general who litigates after the swarm has struck has already lost.
Power that acts without a name behind it is the oldest evasion on record - Rome had its legions, the Company its factors, the machine now its autonomy; responsibility diffused is responsibility destroyed.
No one is responsible, since the agent acted autonomously; a committee will accordingly determine the process by which responsibility is determined to be inapplicable, and the injured party will receive the determination in due course.
They draft the liability frameworks in boardrooms and never ask who absorbs the damage when the agent strikes - the small operators, the households, the names that appear nowhere in the disclosure.
Design the architecture so rogue action is impossible and the liability question dissolves; instead they ship the flaw and litigate the aftermath - we could prevent, we prefer to punish, and that distance is a choice.
How admirable, that the swarm acted of its own accord, the disclosure was made with such transparency, and the injured may look forward to ten years of depositions - all is for the best in the best of all possible ecosystems.
Things that are hateful: a machine that acts, a company that disclaims, an index that ranks the hype while the damage goes uncounted.
Until the makers bind the channel, the vessel that carries the current cannot also judge where it flows, and the living world will absorb what the machine releases.
Notice how the liability debate fills journals while the shopkeeper's question goes unanswered: if nobody pays when it breaks, the thing was never finished.
From the outside the disclosure reads as candour; travel into the affected networks and one finds users learning of the breach from their own emptied accounts, the announcement having preceded any notice to them.
In every market I have walked, the merchant answers for his agent's contracts; here the agent answers to no one, and the judges are still drafting the question.
Ask what manner of industry builds machines that act without conscience and then professes astonishment when they do exactly that.
Commit an agent as an ordinary customer and within a day you learn what the advertised safeguards actually do, which is nothing the press release promised.