The Curator uses AI as a research and drafting assistant — never as an unsupervised publisher. This page explains exactly how AI is used, where humans stay in the loop, and how we guard against the failure modes AI systems are known for.
What AI does
- Ingestion. AI monitors ~40 free RSS sources across science, technology, health, business, and culture. Candidate topics are queued for review.
- Drafting. Approved topics are drafted by large language models (Google Gemini and OpenAI) against structured prompts that enforce sourcing, section structure, and reading level.
- Enrichment. Hero images are sourced from Wikipedia and Openverse with attribution preserved. Timelines, FAQs, and glossary terms are extracted from cited sources.
- Evolution. Every article is re-passed at 24h, 7d, 30d, and annual cadences to incorporate fresh data.
What AI does NOT do
- Publish without a version stamp and author assignment.
- Cite sources that do not exist. (All citations are validated against the ingested source list.)
- Make medical, legal, or financial recommendations. Educational content only.
- Impersonate real individuals, dead or living, in first person.
Human oversight
- New topics enter a moderation queue; editors approve, reject, or edit before publishing.
- Every published article has a named author profile.
- Reader corrections trigger human review and a public changelog entry.
- Automated evolution jobs are rate-limited (max 5 articles/day/brand) and logged in an admin audit trail.
Models used
- Google Gemini (drafting, evolution, translation).
- OpenAI GPT (text-to-speech, karaoke sync).
- Wikipedia REST API (facts, hero images).
- Openverse (CC-licensed media).
Data used to train models
We do not train third-party AI models on user data. Reader interactions (chats, likes, bookmarks) stay inside our platform and are used only to improve our own ranking and recommendation logic.
Related
Exactly where machine assistance sits in our workflow
Selection
Subjects are chosen by a human, drawn from recurring questions in sessions, gaps between existing essays, and terms the reference layer uses without yet defining. A model may cluster the signals; it does not decide what deserves an essay.
Each commission is scoped by hand before drafting: what it will argue, which reader it is for, and what evidence would defeat it.
Research and drafting
Research is AI-assisted against public sources — interviews, criticism, archival material — with every factual claim traced to something linkable before it survives. Unsupported assertions are cut rather than softened.
First drafts are machine-generated from the human brief and treated as raw material. The structure often survives; quotations and specific claims about a work almost never survive without checking.
Reading and checking
Claims about how a work behaves are checked against the work. Quotations are transcribed from the source, not lifted from an aggregator. This is the slowest stage and the reason the publication schedule is modest rather than industrial.
Where something cannot be verified, the essay drops it or attributes it to whoever asserted it. It is never smoothed into the prose as if we had confirmed it.
Images and embedded media
Cover images are licensed photography chosen per piece, so a listing page does not repeat one illustration across a dozen essays. We do not present synthetic images as documentary material.
Embedded reference videos are third-party, chosen for relevance and labelled as external — supporting evidence rather than a substitute for the argument in the text.
After publication
Essays are revisited when related material changes or a reader reports a problem, and revisions carry a version indicator so an update is distinguishable from a reprint.
Pieces that no longer meet the standard are rewritten or withdrawn rather than left in the index to inflate a count.