A box of photographs, cassette tapes, or scanned documents is raw material rather than a described archive. AI tools can draft transcriptions and captions, but speed and accuracy vary by media and tool. The workflow must keep machine output separate from confirmed historical claims.
This is a candidate method in which an approved tool drafts transcription or organisation metadata and a person records consent, provenance, and confidence. The schema can make an unconfirmed date visible, but it cannot prevent someone or another system from later dropping that qualifier.
The three things every archive item needs
For each photo, recording, letter, or document you add to the archive, track three things separately, because conflating them is exactly how family archives drift from documented history into confident folklore:
- Permission and audience: who appears in the item, what uses have they agreed to, and who may see it? For this family workflow, record agreement for a specific item or clearly identified batch. That is a conservative family practice, not a statement that every private archive has identical legal requirements.
- Provenance: where did this item come from: whose collection, roughly when was it acquired or created, and who is the source of any information about it?
- Confidence: is each claim about the item (date, names, location, event) confirmed by a reliable source, or an estimate, and if an estimate, how was it arrived at?
Step 1: Transcribe and describe, with confidence tagged from the start
Before a recording or photograph enters a tool, agree on this particular use with the people involved. Permission to keep a photo in a private album does not establish permission to send it to a transcription or image service. Check who receives the file, account access, retention, training use and deletion limits. If the proposed processing is not acceptable, transcribe manually or use a suitably controlled local tool. Local software still needs review of network connections, backups and shared-device access.
When you use a model to transcribe an audio recording or draft a description of a photo, ask it to separate what it can read or hear directly from anything it is inferring.
If original audio is supplied through an approved transcription path, transcribe it and mark uncertain words [unclear] with timestamps. Do not invent missing words.
If only an existing text transcript is supplied, format that text without claiming to hear audio or detect omitted speech. Preserve its uncertainty markers and flag textual inconsistencies as text-only observations.
Keep the original recording and compare disputed words by listening to it before accepting a corrected transcript.
Here is a description of what's visible in this photo: [describe what
you can see, or the output of an image-description tool]. Help me
draft a caption using only what's stated. Separately, list anything
you would need external confirmation for (a date from an inscription or a location named in the supplied notes), clearly labeled as
"estimate, needs confirmation," not stated as fact.
A model asked to identify people, places, or dates in old photographs can produce a plausible-sounding guess with no visible signal that it is guessing, especially for clothing-based date estimates or generic-looking locations. Never enter a model-generated guess as confirmed fact. Record a living relative’s recollection as named, dated testimony with its basis; it is provenance, not independent confirmation by itself. Upgrade a claim only under a defined evidence rule (for example, corroboration by a dated original or independent sources) and preserve conflicting testimony rather than forcing agreement.
Face and voice processing create extra risk beyond captions. Auto-tagging faces in photo apps, uploading recordings for transcription, or “enhancing” old audio can create additional copies that may be reused. The FTC warns that scammers can clone a relative’s voice from audio posted online (FTC consumer guidance). Prefer local tools when practical; if you use a cloud transcription or vision service, check retention and training settings first, keep original media offline, and do not provide a person’s voice sample without a clear need and informed consent.
Step 2: Record permission for a defined set and use
Agree on the audience and uses before sharing a set involving living people. A batch can be practical when its contents and purpose are clear; separate any item someone wants handled differently. A general “family history project” description is too vague to cover public posting, face tagging, voice processing and private storage at once.
Help me draft a short, plain-language consent question to ask a family
member about a specific set of photos or recordings they appear in,
covering whether they're comfortable being included in a private family
archive, and separately, whether they're comfortable with any of it
being shared more widely (e.g. with extended family, or publicly).
Keep it a genuine question, not a leading one.
Record their actual answer per item or per identifiable group of items, and revisit if someone’s comfort changes; a “yes” given once is not permanent consent for every future use of the archive, especially if the archive’s audience or purpose changes later. When consent is withdrawn, act on it: remove or restrict the item, redact the person from shared views, and note the change in the metadata. “The right not to be included” is operational only if removal has a clear owner and date.
Step 3: Children’s data gets a stricter default
Photos, recordings, and stories involving children need a stricter default than adults: assume restricted access (private family archive only, not shared publicly or with a wider circle) unless the responsible adult has a lawful, child-centred basis to share, and avoid uploading identifiable children’s media to any AI service whose retention and training policies you have not checked. Seek the child’s views in an age-appropriate way and revisit earlier decisions as they mature. UNICEF’s current guidance on AI and children prioritises safety, privacy, transparency, accountability, and children’s best interests. See what ChatGPT remembers, sees, and shares before uploading anything involving a child to a general-purpose tool.
Guardianship, consent, image rights, data protection, and children’s evolving capacity vary by jurisdiction and circumstances. Obtain qualified local review before creating a public, institutional, school, care, or disputed-family archive involving children.
Step 4: Build the metadata schema
Use the family archive metadata schema to keep permissions, sources and claims separate. One item can contain both a confirmed inscription and an unresolved date. A single overall confidence score would hide that difference.
Here is an invented example, not a recovered family history:
| Item/claim | Recorded evidence | Status and next action |
|---|---|---|
| P014, original photograph | Scan from album A; physical print retained | Record scan location and backup |
| Event year: 1958 | Relative A’s recollection, recorded 12 September | Testimony; basis is a remembered move |
| Event year: 1962 | Relative B’s recollection, recorded 14 September | Conflicting testimony; basis is a remembered wedding |
| Caption | “Family gathering; year disputed between two recollections” | Preserve both accounts until another source resolves them |
| Colourized derivative P014-C | Generated from P014 | Label as altered; colours are not historical evidence |
| Audience | Named relatives agreed to a private album | No public posting or cloud processing authorised by that agreement |
An AI caption draft that chooses 1962 because it sounds more specific fails this record. So does “circa 1960” if it conceals the disagreement. Check the exported caption as well as the metadata: qualifiers must survive copying into an album or slideshow.
Use AI to format already permitted descriptions or draft questions for relatives. Do not ask it to identify an unknown person or turn clothing into a confident date. Resolve claims through original documents and attributable testimony, preserving uncertainty when the sources do not decide them.
Do not let restoration or colorization quietly become invention
AI photo restoration and colorization tools may make some damage less distracting, and they can also add detail that never existed: an invented pattern on clothing, guessed eye colour, or a changed facial feature. Keep the original scan alongside every derivative and label the derivative clearly as AI-processed rather than more accurate. The US National Archives recommends keeping physical originals after digitisation and making backup copies; adapt formats, storage, and retention to your archive. This is the same provenance discipline covered in provenance and watermarking.
The escalation rule
If a family disagreement arises about what should be included, who should have access, or whether a specific story is even accurate, that is a family decision, not a metadata problem; the schema can hold “disputed” as a confidence status, but it cannot resolve the dispute. If the disagreement involves a deceased relative’s story that conflicts across different family members’ memories, record both versions with their sources rather than picking one to enter as the official account.
Start with one bounded, consented set
Pick one shoebox, one folder, or one recording to start with. Build the metadata table using the family archive metadata schema, agree on processing and access with the people involved, and check that every exported caption preserves its source and uncertainty. Revisit disputed claims when new evidence appears, rather than asking a model to break the tie.



