Ethical AI

As Remembered uses AI to help families preserve oral history — not to replace judgment, memory, or consent. This page explains how models are used, where humans stay in control, and what we refuse to do with your recordings.

Human in the loop

Every As Remembered project has a Director — a family member who reviews what the system proposes. AI does not silently merge people, resolve contradictions, or publish stories on its own.

  • Linking suggests when two different people might share the same relationship event across sessions; the Director marks suggestions Connected or Unrelated
  • Conflicts flag when extracted facts contradict; the Director chooses which account to treat as canon
  • Merges of duplicate people are always manual
  • Generated narratives are drafts for review, not authoritative records

Provenance over invention

Extractions trace back to source material. Relationship claims in our ledger reference the raw quote from the transcript. When the system creates linker “mirror” rows to connect partners in a story, those mirrors are indexes for search — not new facts attributed to someone who did not say them.

If you cannot verify a claim in the UI, you should treat it as provisional until you confirm it against the recording or transcript.

What we send to third-party AI services

Today, processing works like this:

  • Deepgram receives audio bytes for speaker-diarized transcription
  • Google Gemini receives transcript text and Series Bible context for entity extraction, linking analysis, and narrative generation

These vendors process data under their own policies. We do not sell your recordings or transcripts to data brokers. Details are in our Privacy Policy.

What we do not do

  • We do not use your private recordings to train public foundation models
  • We do not auto-publish family content to the web or The Archive without opt-in
  • We do not present AI-generated prose as a verbatim transcript
  • We do not allow one family’s data to be visible to another account

Known limitations

Oral history is hard. Models and heuristics make mistakes. Examples we design around:

  • Diarization errors — wrong speaker attribution in noisy or overlapping audio
  • Narrator aliases — “Mom” and “Dad” are roles, not always separate people to link across sessions
  • Possessive phrases — “Steve’s father” must not be treated as a distinct person named Steve
  • Date contradictions — the same event recalled as 1985 vs 1986 belongs in Conflicts, not silent overwrite
  • Generative tone — narrative presets range from strict transcription style to more creative “Showrunner” prose; stricter presets hew closer to source quotes

We maintain automated tests and golden evaluation corpora to catch regressions in linking behavior. Beta users help us find edge cases real families encounter.

Messy capture, structured retrieval

We expect uneven, redundant extraction across long projects. The long-term product vision is rich capture with clever retrieval — filter by name, year, place, or relationship — not perfection on session one. Families should not have to manually curate every entity before the archive feels valuable.

Your responsibilities

Obtain consent from people you record where your jurisdiction requires it. Review AI outputs before sharing them outside your family. Use generated documents as aids to storytelling, not as legal or genealogical proof without verification.

Feedback

If you see harmful or surprising model behavior in your project, tell us at hello@astoldby.app. Ethical AI is an ongoing practice, not a checkbox.