How NCASE uses AI

What AI writes

When it is switched on, a language model writes the sentences of an answer from the figures NCASE found, and nothing else. A second model, from a different company, then reads each sentence against those figures; a sentence that does not pass is marked "Not verified".

The models used today are from Anthropic (writing) and OpenAI (checking), through OpenRouter.

What it does not write

Figures: every figure is inserted from the evidence ledger with its publisher, period and date; the model never types one.

The check of an uploaded file, the briefs monitors email, and answers when AI is off: these are written by code straight from the figures, and say so.

How you can tell

An answer AI wrote says so, and that a second AI model checked it; its interpretive sentences carry an "AI reading" tag. Downloaded reports say the same, and include only sentences that passed the check.

No person reviews an answer unless you ask for a review by an NCASE analyst; their name and the date are then printed on the report.

Your data

To write an answer, your question, the earlier questions of the conversation and the figures found are sent to the model provider, outside the Kingdom. Your uploaded files are not. The privacy policy has the details.