Skip to content

[SIGNAL] Unreliable rationales when audio or video is missing #1

@filipsardjoski

Description

@filipsardjoski

Preflight checklist

  • I searched existing signal feedback issues
  • I am not including raw video or audio in this issue

Signal type

Multiple signals

Feedback type

Rationale does not match observable cues

Time range in video (seconds)

No response

What was observable in the video?

The input was missing one modality: there was either no usable audio track or no usable video (for example, a silent video, a black/covered camera, or audio-only input).

Observable behavior in the remaining modality was minimal.

Despite this, the API still produced detailed rationales that referenced cues that were not actually present (for instance, describing facial expressions when no video was available, or vocal prosody when no audio was available).

What did the API return?

What should have been detected instead?

When a modality is missing or effectively empty, rationales should be conservative: clearly scoped to what can be inferred from the available modality, explicitly note uncertainty, or avoid modality-specific claims. Alternatively, the API could surface a structured warning that analysis is video-only or audio-only so clients can down-weight or hide rationales.

Video context

Other

Additional context

No response

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type
No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions