AI use policy
Automation can move quickly. Evidence remains the gate.
WITT uses a disclosed automated newsroom for standard-risk explainers. Sensitive, disputed or weakly sourced work cannot publish unattended.
Permitted assistance
Approved models may help cluster discovery leads, extract candidate claims, organize evidence, draft from verified records, test text against those records and suggest diagnostic improvements. Provider, model, latency and outcome are logged for newsroom evaluation.
Controlled publication
A standard-risk story may publish automatically only when it has the required independent sources, every material claim maps to retrieved evidence, the trigger is confirmed or probable, a separate verification pass finds no unsupported statement, originality and search-intent thresholds pass, and a separate risk review permits unattended publication. The exact approved version and all decisions are logged.
Hard limits
Automation cannot resolve disputed claims by assertion, publish high-risk or unresolved work, change historical evidence, expose plaintext credentials, execute arbitrary code, modify its own safety gates, or deploy code. Health, finance, crime, death, litigation, politics, war, personal allegations and similar sensitive subjects require human review.
Integrity layer
Research is gathered before drafting. Instructions require exact-event relevance, verified proper names, source-bound numbers and removal of unsupported sentences. Quotations are used only when exact wording is present and attributed. The article is checked independently against the source packet before it can become indexable.
Controlled experimentation
The system may analyse aggregate quality scores, reading engagement and sharing actions and propose a small drafting-guidance experiment. A proposal cannot activate itself. An owner must inspect and approve it, explicitly start an A/B test, and may stop it at any time. The unchanged prompt remains the control; evidence rules, risk controls and publication gates cannot be altered by an experiment, and no candidate is promoted automatically.
Disclosure and monitoring
Each story states whether it passed the automated integrity desk or received individual editorial review. We monitor provider failures, sourcing diversity, quality scores, held drafts and critical errors. Model and prompt changes cannot promote themselves into production.