Most discussion of AI safety in software is about what a model should refuse to say. A more practical question, for the kind of systems most of us actually build, is narrower: what is the machine allowed to assert on its own?
We had to answer it concretely on LitScape, a literary atlas that maps the places Nigerian novelists named across seventy years of writing. Automated extraction reads the corpus and proposes that a novel referenced a place. That proposal, published unchallenged, would be a scholarly claim — and a wrong one would be a citation someone else builds on.
Make the rule structural, not cultural
The obvious approach is a review process: extract, then have someone check before publishing. That works until the day it is inconvenient, and processes that depend on nobody ever being in a hurry are not really controls.
So the constraint went into the data model instead. Every machine-produced link is written as unverified and surfaces in the interface as a candidate, visibly distinct from a claim. Only the researcher can promote it. There is no code path where an extraction becomes a citation on its own — not a discouraged one, an absent one.
Draw the uncertainty
The second rule was about honesty in presentation. Fiction rarely supplies coordinates. A novel says a character crossed the lagoon, not that they were at 6.45°N, 3.40°E.
The tempting move is to place a neat pin and let the interface imply a precision the source never had. Instead every place carries a confidence level drawn as an actual circle on the map — 500 metres when the text is exact, two kilometres when the reference is symbolic. The reader sees how firmly a place is known, not just where it supposedly is.
Why this generalises
Neither idea is specific to literature. Any system where automated output feeds something consequential — a credit decision, a compliance flag, a medical note — faces the same question, and usually answers it with a policy rather than a schema.
A policy is a promise about behaviour. A data model is a constraint on it. When the stakes are someone else's work being wrong in public, the second is worth the extra afternoon.