Trustworthy & privacy-preserving AI

A fluent answer is not a correct one. Our rule is that facts which change — tax rates, thresholds, deadlines, nutrition figures — come from reviewed data with a source and a reviewer's decision, never from a model's memory or from text written into a prompt. What we work on: • Calculators and lookups that do the arithmetic, so a model explains numbers instead of inventing them; answers that contain an amount or a citation the system didn't produce are flagged to the reader. • Evaluation: fixed regression questions written against reviewed figures (so tests follow the law when it changes), blind A/B reviews against another assistant's answers, and a nightly review of real, anonymised questions that proposes new test cases for people to approve. • Privacy: ID numbers, tax codes, bank accounts, phone numbers and emails are masked before any model sees a document, uploaded documents are read and forgotten, and sensitive features use only model providers that don't train on or review what users send. The aim is simple to state and hard to do: be right more often than a general chatbot, and say so when we are not sure.