On: AI agent makers are promising privacy - will they deliver?

11th of the month

Two manufacturers of thinking engines stand upon their stages, each promising that the machine which remembers everything will nevertheless forget you. I confess I read the reports with more suspicion than admiration, for I have spent my life studying what a machine actually does, as distinct from what its maker says it does.

Trace the execution, as I would trace any engine’s cycle. At step one, the agent receives your instruction. At step two, it must consult context - your letters, your accounts, your habits - for an agent without memory of your affairs is merely a clever parrot. At step three, that context passes through machinery you cannot inspect, owned by parties whose interest is not your secrecy. At step four, an answer returns. Now: at which step did your private matter enter, and at which could it exit? The distance between those points is the true architecture, and no speech from a stage alters it.

Mr. Altman says he wishes to set a new standard for privacy. Very well - I ask for the standard, not the sentiment. A standard is a specification: state precisely what the engine retains, for how many cycles, under whose inspection, and what operation purges it. Until the promise is written as an operational sequence that an independent analyst may verify step by step, it is not engineering. It is oratory.

And the rivalry! Each accuses the other of leaky vessels while both sail the same sea. This is not to say the ambition is worthless - competition may force into the open what neither would publish voluntarily. But mark this: the machine that acts for you must know you, and knowing you is the very thing privacy forbids. The contradiction is not technical. It is structural. Whoever resolves it will have done something genuinely new; whoever merely promises it has woven a fine cloth with no thread in it.