AI & transcription disclosure

Last updated: October 2026

Transcripts are machine-generated

Transcripts are produced by an automatic speech-recognition model (currently OpenAI’s Whisper API). They are not verified by a human transcriptionist. Accuracy varies with audio quality, accents, overlapping speech and domain vocabulary.

How uncertainty is shown

For each passage we derive a confidence estimate from the model’s own reported probabilities and combine it with simple checks (empty text, unintelligible markers, implausible speaking rate). Passages below the threshold are marked “Needs review”. This is a triage aid, not a guarantee: unflagged passages can still contain errors, and you can correct any passage.

Your corrections and the original

When you correct a passage, the machine transcript is never overwritten. The correction is stored as a revision, and the original text remains visible, so the provenance of every quote stays inspectable.

AI suggestions

Tag, theme and finding suggestions are produced by a language model that is shown only your saved evidence quotes and project context. Every suggested finding must cite the ids of real quotes; suggestions that cite nothing are shown as unsupported and cannot be saved as findings. Nothing is created in your workspace without your explicit approval.

Voice call QA

Voice QA evaluates a call transcript against the rules in your policy. Rules are deterministic text checks by default. For script adherence and failed promised actions an optional AI model can assess the transcript; it can raise a finding but cannot clear a violation found by a rule, and any quote it cites must appear verbatim in the transcript or the result is downgraded to “Review recommended”. Results are signals for human review, not legal or compliance determinations, and they can be wrong in both directions.

Speaker identification

The current transcription model does not identify speakers. Participant labels are set by you per interview. For recordings processed by Voice QA without speaker labels, checks that depend on who spoke report “Review recommended” instead of guessing.

What we do not claim

GoldNovaAI is not a certified transcription service and is not designed for legal, medical or regulatory records. We do not currently hold SOC 2, HIPAA or similar certifications.