When a Digital Human Is Unsure: Six Live-Video Disclosure and Repair Tests

Abstract dark blue digital display representing live video AI uncertainty and cross-modal feedback

Real-time AI digital-human video can be friendly and expressive while remaining transparent about what an AI personality actually knows. This implementation framework proposes six scenario tests, consistent voice and caption states, and measurable correction behavior for mainstream social apps.

Separate uncertainty from an error message

A digital human in live video is not a camera-connected person who necessarily knows the answer. When an AI personality lacks reliable evidence, a fluent face and natural voice can make a guess sound unusually authoritative. A useful interface therefore distinguishes three states: the system has grounded information, it is offering a creative or hypothetical suggestion, or it cannot verify a factual claim. These states must be communicated in text and speech without pretending that a human specialist has checked the response. The distinction is especially important when a question concerns a real person, a purchase, a public event, or medical and financial consequences. A short statement such as “I cannot verify that detail” preserves the social tone while maintaining factual boundaries. Teams should record which state the model selected and which evidence supported that state, not the user’s private question itself.

Use a cross-modal disclosure contract

The visible avatar, spoken reply, and text transcript should describe the same degree of certainty. A model that says “probably” while the subtitle renders a confident assertion has broken the contract. Define a small structured object for each reply: claim type, source availability, confidence treatment, user-facing disclaimer, and escalation decision. Do not expose a raw model probability as though it were calibrated. The uncertainty message should remain readable when captions are enabled and audible when the video is minimized. A distinctive facial animation may reinforce the message, but it cannot substitute for words because users may have visual or cognitive accessibility needs. Test muted video, delayed audio, low bandwidth, and a screen reader. The goal is not to place warnings everywhere; it is to prevent one modality from undoing a qualification made in another.

Map six high-risk live-video situations

An operational design can classify questions into six scenarios. First, a user asks about an unverifiable recent incident. Second, a creator’s twin is asked to confirm a private preference of the real creator. Third, an uploaded photo contains an unidentified bystander. Fourth, the user asks for a purchasing promise or refund guarantee. Fifth, the avatar is interrupted halfway through a qualified answer. Sixth, the conversation changes language and a negation is lost. Each scenario needs a permitted response, an explicit prohibited claim, a transcript example, and an expected fallback. For the creator example, the AI can discuss its published persona rules but should not assert that the human creator personally approved a new claim unless approval evidence exists. For the purchasing example, it should refer to actual checkout terms rather than invent a binding promise. A case table is more useful than a generic “be safe” system prompt.

Ground the release gate in established guidance

The National Institute of Standards and Technology published AI Risk Management Framework 1.0 on January 26, 2023. Its govern, map, measure, and manage functions suggest a practical division of labor: product owns the disclosure UI, engineers own evidence identifiers, reviewers own scenario labels, and operations owns incident response. The W3C’s WCAG 2.2 guidance on status and accessible information helps evaluate whether feedback works beyond an ideal video experience. Neither document is a product endorsement or a complete legal compliance checklist. We propose a controlled test with two versions: voice-only uncertainty wording versus consistent voice-plus-caption wording. Ask participants to identify what the avatar knew, what it guessed, and whether a human had reviewed it. Measure comprehension, task success, and unnecessary user abandonment. Randomize test order and keep audio quality comparable. Otherwise animation quality may confound the result.

Design interruption and repair before optimizing charisma

Real-time digital-human video involves barge-in, network jitter, partial subtitles, and speaking overlap. A correction that appears only in the last second of a video clip may never reach a user who interrupts early. Design an atomic claim marker so a key qualification is delivered near the claim rather than after a lengthy performance. When the user says “stop” or “that is incorrect,” voice synthesis should stop, the visible speaking state should change, and the transcript should preserve the disputed status. The AI can ask a clarifying question, but it should not silently rewrite the prior answer as if no error occurred. If a network reconnect repeats a partial sentence, the player should show that it is resuming rather than issuing a new verified assertion. Test all these paths with synthetic conversations containing no real personal data. A useful defect metric is the share of interrupted claims whose essential qualification was lost.

Publish evidence that users can understand

Before shipping, build a review package with at least fifty purpose-designed test turns rather than merely a polished demo. Include ordinary chat, references to news, identity ambiguity, photo uploads, payment questions, and multilingual interruptions. Have reviewers rate whether text, voice, and video agree; whether the user can request a correction; and whether a misleading certainty cue survives in a screen recording. Report failure counts by scenario, not one aggregate “trust score.” A creator-facing dashboard can explain why the AI personality declined to confirm a private fact, while withholding personal conversation content. For mainstream AI social apps, clarity and relationship growth are compatible: the digital twin can be warm, engaging, and visually expressive while stating what it does not know. That is a stronger product promise than offering frictionless confidence on every turn.

Sources

Illustration: A Chosen Soul / Unsplash. The study scenarios and metrics are proposed editorial methods, not claims of deployed Tuikor AI features.