Free AI Genealogy free · no login · no account · nothing stored

The record reader that never makes things up

The AI genealogy tool that shows its receipts.

Paste a record — a census row, a marriage entry, a ship manifest, an obituary, a page of old German or Latin handwriting. It reads it, extracts it, translates it, and tells you where to look next. Every fact points at the exact words it came from. What it cannot read, it flags. What it cannot know, it says.

Read a record See what it refuses to do

The Vol. 5c, Page 473 problem

In June 2026, on a RootsChat thread that drew 615 replies, a researcher told how a well-meaning relative ran her family question through a major AI’s deep-research mode. It announced “a critical breakthrough”: a birth registration, Wells district of Somerset, September quarter 1896 — citing Birth Index Vol. 5c, Page 473. She checked the actual index. No such entry exists. Confronted, the AI admitted the reference was “completely fabricated.”

RootsChat, “A cautionary tale re. the perils of AI in genealogy,” June 2026

Every design decision on this page exists so that sentence cannot appear in its output — not because the model is more honest, but because the output is checked: every fact must point at verbatim words in your record, and “where to look next” can only cite sources this server actually searched seconds ago.

The tool

No record handy? Try a sample (all fictional):
…or drop a photo or scan of a record here (JPG/PNG)
It will be transcribed first — unreadable words flagged — so you can check it before analysis.
Free · 15 records a day, 5 image transcriptions · nothing stored

The ledger will fill here.

The contract

What it refuses to do

  1. Will not fill in a word it cannot read. Unclear means unclear — flagged as [unclear], never guessed.
  2. Will not invent a citation. No volume, no page, no archive, no URL that did not come from somewhere real.
  3. Will not call a hypothesis a fact. A shared surname is a question, not a relative.
  4. Will not silently “fix” the record. A wrong-looking age stays wrong-looking — and gets flagged, because the error is evidence.
  5. Will not import assumptions. A draft card is not service. A neighbor is not a cousin. A passenger is not yet an immigrant.
  6. Will not pretend the record answers a question it does not. The “what this record does not say” list is the feature.
  7. Will not pad. Three facts is three facts.
  8. Will not modernize your ancestor’s spelling.
  9. Will not cite a source this server did not check — and tells you plainly which sources you must search yourself.
  10. Will not remember you. No account, no history, no tree. Your record is read and discarded.

The mechanics

How the honesty is enforced

Verbatim anchors

Every fact carries the exact words it came from. After the model answers, this server checks each anchor against your record’s text. Anchors that fail are demoted publicly — flagged amber in your results — not quietly dropped.

Live lookups

“Where to look next” may only cite the honest link-out list — genuinely free archives with no API, labeled “you run this search yourself” — or Chronicling America’s ~21M free newspaper pages, and that search runs live from your own browser, so this server never even sees the hits.

Uncertainty as output

[unclear] flags, confidence grades, and an explicit cannot-determine list. The model is instructed that “I cannot tell you” is always a valid, complete answer.

Reading by DeepSeek V4 Flash; handwriting transcription by Gemini 3.5 Flash. Free because reading is cheap and the honesty is the product. Your record goes to the reading model and nowhere else — nothing is stored, logged, or kept. There is no account because there is nothing to keep.