AI Content Policy
FlightFinder uses language models to turn official accident records into readable summaries at a scale no single editor could reach — roughly a quarter of a million occurrences drawn from more than a hundred investigation authorities. This page states plainly where models are used, what they are never allowed to do, and which automated gates their output has to clear before it is stored or indexed.
Where models are used
- Plain-language accident summaries. Generated from the official narrative and the structured record for that occurrence.
- Structured extraction. Pulling fields (phase of flight, occurrence category, occupant counts) out of report prose into a fixed taxonomy.
- Directory and digest text. Short descriptive passages on aircraft, airline, and airport pages, derived from data we already hold.
Long-form data stories are written by a person and carry a named byline. Statistics inside them are computed from the database, not produced by a model.
What models are never allowed to do
- Introduce a number that is not in the source. Every digit in a generated article is checked against an allowed set built from the resolved record plus the digit tokens appearing verbatim in the official narrative. Ungrounded numbers fail the gate.
- Assign a cause. Generated text is scanned for probable-cause phrasing and rejected when it appears. Cause language reaches a page only from a published official finding.
- Invent a field value. Extraction is validated server-side against fixed taxonomies; a value outside the taxonomy is refused outright rather than stored.
- Carry a fabricated byline. Machine output is published under the collective FlightFinder Editorial, never under an invented person.
- Speak in its own voice. Model preamble, self-commentary, and refusals are detected and stripped; a refusal is treated as no output, not as content.
Why the gates are deterministic
The checks above are ordinary code, not a second model asked to grade the first. A generation worker measured at roughly a fifty-percent hallucination rate is still safe to run behind a gate that can only ever pass numbers present in the source — and that property is what we rely on, rather than the model behaving well. The server-side validator is authoritative even when the worker performs its own checks.
What this policy does not claim
We do not claim that a person reads every machine-generated summary; at this volume that would not be true. Bulk generated pages are governed by the automated gates described here and by the numeric floors in our Editorial Standards — output that fails them is not stored, or is stored but kept out of the index. Human review is applied to data stories and curated digests before they are marked indexable.
We also do not generate photorealistic imagery of real people, aircraft, or accident scenes. Photography is credited to its source; see Attributions.
Precedence
Where a generated summary and the official report disagree, the official report is correct. The summary is corrected or removed. Every accident page links its source report so any reader can check the machine against the original.
Reporting a problem
If you find a generated passage that misstates a record, email [email protected] with the page URL. Factual corrections are verified against the upstream dataset and applied to the source of the error, so the same fault does not regenerate.
See also: Editorial Standards · Methodology · Attributions · About · Contact