Skip to content
AI CertaintyTHE HUMAN DIMENSION

AI CERTAINTY / CONCEPT STUDY 01

More signals.
More context.

Room for uncertainty.

Language, voice, movement, and context show different sides of a person. AI can use clues to help, while leaving room to ask what those clues mean.

Explore the signal map

Synthetic concept · no live sensing

01 / THE SIGNAL MAP

Signals, before assumptions.

Five possible input categories, not a universal checklist. Choose for the task; there is no need to enable them all.

ILLUSTRATIVE ONLY

Potential input

01 / 05

Language

What is said

Text or transcripts a person chooses to share, including explicit requests and their own descriptions of their experience.

Example observations

  • Content and conversational meaning
  • Requests, preferences, and self-reports

Interpretation boundary

A sentence does not fully represent feelings, stable preferences, or future intentions. Preserve ambiguity and the person’s own account.

Potential input

02 / 05

Voice

How it sounds

The acoustic form of speech: pauses, pace, pitch, and the timing of conversational turns.

Example observations

  • Tempo and pause structure
  • Pitch and intensity patterns

Interpretation boundary

A pause alone does not establish anxiety, dissatisfaction, or deception. Check the speaker’s explanation, context, and recording conditions.

Potential input

03 / 05

Vision & Body

What is visible

Visible facial movement, head direction, gestures, posture, and their changes within an interaction.

Example observations

  • Facial and bodily movement
  • Gaze direction and gesture timing

Interpretation boundary

A smile does not necessarily mean happiness; looking away does not establish lost interest. Situation, culture, and individual differences matter.

Potential input

04 / 05

Interaction

How interaction unfolds

Interaction records that a person allows, such as clicks, undo actions, steps, and explicit feedback.

Example observations

  • Turn-taking and response latency
  • Choices, corrections, and feedback

Interpretation boundary

Repeated edits and waiting do not establish limited ability. Interpret an interaction with its task, situation, and the person’s feedback.

Potential input

05 / 05

Physiology

Optional, with consent

Optional heart-related, breathing, or skin-conductance measurements, only when needed for the task and voluntarily shared.

Example observations

  • A defined measurement and sensor
  • Baseline, conditions, and signal quality

Interpretation boundary

A bodily response is not a mental-state label or diagnosis. Physical activity and measurement conditions can also change a signal. No sensors are enabled here.

02 / THE INTERPRETATION LAYER

Meaning lives in context.

In this concept, context and time sit around the inputs. They help frame interpretation; they are not extra sensors.

Context

The task, surroundings, relationship, culture, and context a person has chosen to share. A group pattern cannot replace their explanation.

Time

Sequence, duration, change, and an appropriate baseline. One instant is only one observation.

  1. 01

    Observe

    Describe the available signals.

  2. 02

    Contextualize

    Check context, missing information, and conflicts.

  3. 03

    Clarify

    Ask the person before acting.

SYNTHETIC EXAMPLE

One pause. Different possibilities.

The pause stays at 2.4 seconds. Change only the context the fictional person provides.

Thinking of the words

“I’m thinking about how to say it.”

An example response

Take your time. We can wait until you’re ready.

Audio cut out

“The sound cut out.”

An example response

Would you like me to repeat that part?

No further context

The cause is undetermined.

An example response

Would you like a moment to think, or a different way to phrase it?

Written examples, not analysis of you or model-generated replies.

03 / THE INFERENCE BOUNDARY

Signals are observations.
States are hypotheses.

Emotion, intent, and need belong to an interpretation layer. A responsible design would keep these inferences uncertain, explain their evidence, and allow the person to correct them.

Potential observations
Hypotheses to verify
EmotionIntentNeed
Conceptual relationship; no scoring, diagnosis, or individual state estimate is produced.

LONG-TERM / DESIGN POSITION

More data does not automatically mean more certainty.

Our long-term direction is to connect human signals with evidence and uncertainty. This page proposes a vocabulary for that conversation, not a universal checklist: availability, purpose, consent, and evidence should determine which inputs belong in a particular system.

About the evidence graphQuestion-to-evidence demoExplore our research