Tolerance for inference depends on the setting; the same inference is judged differently across products

What it is

A music app estimates a mood tag from late-night plays and most people call it convenient. The same sequence, on a hiring platform or a shared family album, is judged an offence. Context-dependent inference tolerance means acceptance of one kind of inference moves with the product’s promise, relational stakes, and whether exit exists. There is no cross-product “we may estimate this.”

Can estimate, estimates accurately, and the user accepts are three different facts. Acceptance is bound to the setting.

Why it happens

People interpret data use through the product contract. A music app’s contract is “find the next track”; inference is read as service. A hiring app’s contract is “show what I can do”; inference is read as scrutiny. Same technique, different contract, different moral sort. Relational stakes move too: if the inference might be seen by family, employer, or classmates, tolerance drops even when the tag is harmless in solitary use.

Products also paper every setting with “this is personalization,” skipping a contract that should have been renegotiated. A pipe accepted in music is then plugged into health, school, or work. The clash is not that users are fickle. The contract was swapped.

Studying it

Describe the same inference (mood, household structure, health lean) inside different product frames: entertainment, shopping, hiring, health, education, work. Measure acceptance, surprise, demand to be asked first. Independent variables: product frame, whether the inference is displayed, whether an exit exists, relational stakes (might others see it). Dependent variables: acceptability ratings, intent to exit, judging the inference as “service” versus “surveillance.”

Do not report “users do not actually mind inference” from a single-product sample. That contract does not explain another. Cross-product contrast is the design of this question.

Where it stops holding

Legally barred sensitive classes do not become lawful because a setting is more casual; tolerance studies cannot paint over a red line. “Acceptance” from children and dependent users is not adult consent. The same product can have internal scene cuts (work profile vs private profile); cut by profile, not by app icon. This entry explains why the same inference is judged differently. It does not replace limits on sensitive classes, and it does not replace splitting inference opt-out from collection opt-out.

Applying it

• Write an inference-purpose list per product contract: tags allowed in entertainment must not, by default, enter hiring, health, or work surfaces on the same account.
• Re-inform when the surface changes. Do not assume a music “allow” still holds at work.
• Check: take one inference about to ship and ask acceptability in one sentence on three different surfaces. If it is acceptable on only one, ship it only there — not as a platform default.

Related

• Same group: L6.06.1 Behavioural inference can reach information the user never disclosed · L6.06.2 Displaying an inference is itself a disclosure · L6.06.3 Inferences in sensitive categories need extra constraints · L6.06.4 Combinations of non-sensitive behaviours can yield sensitive attributes; item-wise compliance is not whole-system compliance · L6.06.5 Inferences on shared devices and shared screens disclose to third parties · L6.06.6 Wrong inferences also harm, and the misclassified often have nowhere to appeal · L6.06.8 Opting out of inference and opting out of data collection are different controls
• Nearby: L6.05 Turning Personalization Off · L5.09 Overtrust and Trust Collapse · L1.02 Expressing Capability Boundaries
• Search terms: context-dependent inference tolerance · contextual acceptability · product-specific inference