when a person or an agent uses a model, the model can learn who they are — and once a model knows, that knowledge is hard to take back. this lab is about giving the model everything it needs to be useful, and nothing it needs to identify you.
the problem
identity leaks into models by default. a name, an account, a tenant — these travel in the request and the context, and can be surfaced again later, logged, or absorbed. the fix is not to stop using models; it is to stop handing them your identity.
deconstruction
the model does not need to know who you are
separate who is asking from what the model sees. the model needs the request; it does not need the person. so we give it a stand-in.
a pseudonym is an identity with the answers removed
a pseudonym lets the system know who you are while telling the model nothing. the mapping between the stand-in and the real person lives only on the trusted side, and is never shared.
ephemeral — dies with the session
a pseudonym that lasts one session cannot be accumulated, correlated, or replayed. once the session ends, the stand-in means nothing.
single-use — worthless even if it leaks
a one-time pseudonym cannot be reused to reach back to a person. even a leaked stand-in is spent.
the model, not the system
the real identity still exists — for accountability, for audit, for enforcement. it just never crosses the boundary into the model. pseudonymous can mean anonymous — from the model’s point of view, that is exactly what it does.
tensions
- utility versus privacy — the model needs enough context to be useful, never enough to identify.
- accountability versus anonymity — you stay accountable; the model stays blind.
- correlation — even pseudonyms can be strung together if they persist, which is why they are ephemeral and single-use.
where this lands
a boundary where identity is translated into a stand-in before it touches the model, and the stand-in is discarded the moment the session ends.