Permitted assistance
Automated tools may help classify topics, extract candidate entities, cluster coverage, structure fact packets, generate angles and drafts, and perform hostile criticism. Deterministic processing is preferred wherever it can do the job reliably.
Human control
AI output is untrusted editorial input. It cannot verify itself, lower a risk lane, activate a feature, satisfy a rights check or approve publication. Every published item requires explicit human approval of the exact revision.
Data boundaries
Raw story leads, submitter information, admin identities and unpublished confidential material are not sent to models. The public story-lead feature is disabled. We do not automate a consumer AI interface or store consumer-product credentials.
Cost and failure
Launch inference is limited by a daily application ledger below the target free allocation. Quota, schema, provider or quality failures stop safely as MANUAL_AI_REQUIRED. They do not retry indefinitely and they never trigger publication.
Disclosure
Where AI materially assisted a published article, that article carries a specific disclosure. The editor remains accountable for selection, verification, revision and publication.