Users should be able to see which signals the system currently used to draw which inference

What it is

If an implicit system has already formed an inference—“you are stressed,” “you are leaving”—the user should see two things: which kind of signal it used, and which state sentence it drew. Inspectability is not a weight plot from inside the model. It is a runtime grounds–conclusion pair. Without visible grounds, later correction has no object.

Why it happens

Implicit input has no user-signed command log; memory must be supplied by the system. Grounds need vocabulary people can check: heart rate rose, the wrist shook, you neared the door—not “the intelligence engine.” The conclusion must line up with what the interface already did, or inspection becomes a second copy deck. Inspectability is also time-limited: an inference that lives 200 ms cannot be inspected; a short playback buffer is required. Full explainable ML is not the premise—kind-level disclosure (“EDA, not the Ethernet camera”) comes before features.

Studying it

Right after an inference, ask what the system thinks happened and on what basis. Factors: grounds shown or not, kind versus raw waveform. Outcomes: attribution accuracy, rate of blaming oneself. Contrast “conclusion only” with “conclusion plus grounds.” Laboratory research sensors already hint that “the body is being measured”; background inference in the field is harder to inspect, and tests should run on real skin.

Where it stops holding

Safety and anti-abuse settings may be unable to show all grounds (to block gaming the detector). Kind-level disclosure is usually still possible. Raw physiological waveforms are unreadable for most people; dumping an ECG is not inspectability. Children or people with cognitive impairment need shorter sentences, not more charts. Inspection itself discloses sensing and may need to be deferrable in social settings.

Applying it

• Put “grounds” next to the current inference: signal kind plus one conclusion sentence, raw streams collapsed by default.
• Keep a few tens of seconds of replay after the inference ends, so it does not vanish.
• Ban “AI thinks” in grounds copy; write a checkable measurement.
• Verify by asking unbriefed users why it did that after an inference; if they cannot name the signal kind, inspectability failed.

Related

• Same group: C9.13.2 A correction control must appear with the inference result, not buried in settings · C9.13.3 A correction should carry to later similar inferences, or users will keep fixing the same error · C9.13.4 Implicit inferences that cannot offer a correction path should not change user-perceptible UI behavior
• Adjacent: C9.03 Heart Rate and Electrodermal Activity · C9.05 Implicit Interaction
• Search: inference inspectability · why this inference · physiological disclosure