Personalization is often described as a bargain: let the product observe enough and it will become more useful.
The bargain is poorly specified. Useful for which purpose? Which observations are necessary? How long do they persist? Can the person see, correct, or remove what the product believes? Does the experience still work when the answer is no?
Attention does not require exhaustive observation.
Prefer declared context
A person can often provide a better signal than a hidden inference. They can choose a goal, save a preference, identify the subject they are practicing, or select the kind of experience they want today.
Declared context is not perfect, but it is legible. It gives the person a way to shape the interaction directly and a clear object to change later.
Behavioral signals should have a concrete product purpose, not a vague promise that more data may become useful someday.
Separate experience from acquisition
Private product activity does not belong in a public marketing analytics stream. Conversation content, learner work, reflection writing, scores, uploads, and sensitive identifiers should not become campaign parameters or page-level telemetry.
Measurement can answer bounded questions with closed, non-sensitive events. The product can improve without exporting the substance of a person’s work.
This separation is both a privacy boundary and a design discipline. It forces the team to name what it actually needs to learn.
Make the profile contestable
If an experience keeps a preference or working context, the person should be able to understand it. Inferences should not quietly harden into identity.
Correction and deletion are not edge cases. They are how the product acknowledges that people change, systems misunderstand, and a useful choice in one context may be unwelcome in another.
The colleague-or-spy test
A good colleague pays attention to the work you are doing together. A spy gathers broadly because the information may become useful later.
Can the product explain why it needs each signal? Is the scope proportional to the experience? Can the person decline without losing the core value? Can they see and revise what persists?
Personalization should make the product more responsive without making the person more exposed.