💡 Confidence-Transparent Avatar: Uncertainty Signaling Through Body Language and Verbal Hedging

[github-actions[bot]]

Summary

TalkTerm's avatar visually and verbally expresses its confidence level in responses. When the agent is certain, the avatar maintains confident posture and direct speech. When uncertain, the avatar shows thoughtful hesitation (chin-touch, slower pace, slight head tilt) and uses verbal hedging ("Based on what I found, though I'd recommend verifying..."). This helps non-technical users — who can't independently evaluate AI outputs — make better decisions about when to trust and when to double-check.

Market Signal

No major competitor offers calibrated uncertainty signaling through embodied avatar behavior — this is a first-mover opportunity. The trust gap is real and measurable: 56% of CEOs report zero measurable ROI from AI (Gallagher 2026 AI Adoption Benchmarking), partly because teams blindly accept AI outputs without evaluating accuracy. Research on AI trust consistently shows that users who discover errors after blind trust lose far more trust long-term than users who received advance uncertainty signals. Voice AI satisfaction is 72% (up from 53% three years ago per G2 2026 AI Market Report), but trust gaps persist as the next frontier.

User Signal

TalkTerm's target users are non-technical knowledge workers (PRD Journey 1: Sarah the PM). Sarah can't evaluate whether a competitive analysis is accurate by reading the source material — she needs the AI to signal when it's less sure. Existing idea #323 (Emotion-Aware Avatar Response) detects USER emotions to adapt communication style — this is the inverse: expressing AGENT confidence to inform user judgment. No existing idea covers this dimension. The non-technical user adoption research (Goldman Sachs 2026, McKinsey 2026) shows that building trust is prerequisite to sustained adoption — users who feel uncertain about AI reliability stop using it.

Technical Opportunity

Rive state machines already support multiple animation states (listening/thinking/speaking per FR3). Adding confidence-graduated states (confident-speaking, hedging-speaking, uncertain-speaking) is a natural extension of the existing state machine — 3 new states mapped to the existing speaking animation with posture/pace variations. The Claude API's response metadata and the agent's tool-call patterns provide natural confidence proxy signals without relying on unreliable model self-assessment: single authoritative source = high confidence, synthesizing multiple contradictory sources = lower confidence, speculation beyond source material = explicit uncertainty. The ConfidenceLevel value object can be defined in src/shared/types/domain/ and flow through the existing IPC message bridge.

Assessment

Dimension	Score	Rationale
Feasibility	high	Extends existing Rive state machine with 3 new animation states; confidence estimation uses proxy signals (citation count, task type, source quality) rather than unreliable model self-assessment
Impact	high	Directly addresses the trust gap that prevents sustained non-technical user adoption — the #1 differentiator for TalkTerm's "relationship layer" positioning
Urgency	med	No competitor offers this yet — first-mover opportunity without competitive pressure. Should be designed into the avatar system from the start but can ship iteratively

Adversarial Review

Strongest objection: LLM confidence calibration is notoriously unreliable. If the avatar expresses false confidence on wrong answers, or excessive uncertainty on correct answers, the feature backfires and erodes trust faster than no signaling at all.

Rebuttal: Don't rely on model self-reported confidence. Use proxy signals: (1) number and quality of source citations in the response, (2) whether the response required speculation beyond available data, (3) task type (factual lookup = higher baseline confidence, creative synthesis = inherently lower). Default to confident; only signal uncertainty for genuinely ambiguous situations (the 80/20 rule). The avatar's persona makes uncertainty feel natural ("Let me be upfront — I'm not 100% sure about the Q3 numbers, so you might want to verify that with your finance team") rather than alarming. This mirrors how trusted human colleagues communicate.

Suggested Next Step

Define ConfidenceLevel as a domain value object (branded type) in src/shared/types/domain/ with three levels: confident, thoughtful, uncertain. Design 3 Rive animation state variants mapped to these levels (posture, pace, gesture differences). Add confidence estimation logic to the AgentSession aggregate root based on response metadata analysis (citation count, tool-call patterns, task type classification).