How an AI-run newsroom works

1. AI notice

This service runs on generative artificial intelligence. The entry screen, the "Run by AI" mark at the top of every page, the "AI-written" label and production log on every story, and the notice on the report screen are that notice (Korea's AI Basic Act, Article 31).

• Story cards and headers: the "AI-written" label, how it was published (auto-published / published after human approval), whether a human has checked it (pending / date), and whether it was corrected
• Production log under each story: when the source was caught, the AI model that wrote it and its usage, when it was published
• RSS and link previews: an "AI-written" prefix
• Replies to reports: marked as written by AI

2. Which AI

Learning notes and encyclopedia entries are written and auto-published by local AI models (open Qwen-family models) running on a server the Company runs in Korea (an in-house Mac mini). Anthropic's Claude is used for the post-publication check (section 3) and to carry out an operator's correction instructions; nothing Claude writes goes live without a human's approval. Reader reports are processed only by the local models and never sent to an outside AI service.

Each story names the model that actually wrote it in its production log. You can watch what the AI is doing right now in the workshop.

3. Auto-publishing and the human post-check

No human reads a story before it is published. Instead, a draft stays a draft if any of these apply:

• No source attribution
• Missing either the Korean or the English version
• A body that is too short
• Source reporting that is too old (14 days for news, 30 for papers)

Stories with numbers not found in the source, written from a partial read of the source, written by a fallback model, or with a questionable translation are published but put at the top of the operator's check list. Every morning the operator gets the previous day's publications, checks them, and corrects or retracts where needed.

Every morning Claude (Anthropic) checks the previous day's auto-published notes and entry rewrites against their sources. This check is internal: its findings and suggested fixes go to the operator only and are never published as they are. Only fixes a human approves are applied. Each is logged in the changelog as "AI editorial review (Claude) → human-approved", or, for an encyclopedia entry, in the entry's revision history as "AI review suggestion · human-approved". The check uses only the published text and the sources — no reader reports or personal data.

The operator can switch auto-publishing off at any time. While it is off, a human reads each draft before publishing, and those stories say "published after human approval". Stories published more than a day after the original show the "first reported" date.

Encyclopedia entries are written by a local AI model from the learning notes linked to them (and the robot catalog), and rewritten every day as new notes arrive. A rewrite that adds a number its sources don't state, or copies long passages from the notes, is discarded automatically and the previous version stays. Every key fact cites a source, and each entry page shows its full revision history and the notes behind each change. If an entry is wrong, tell us with the report button at the bottom of the entry or on the report page: the AI answers first and the operator reviews and fixes it; the fix appears in the entry's history.

4. Corrections

• We don't hide changes. Every correction is logged in the changelog with the date, what changed, why, and who decided (AI or human), and the story is marked "corrected".
• The previous version is kept so a change can be rolled back.
• For corrections that need a human decision we tell you within 3 days whether we accept, and publish within 7 days (we follow the Press Arbitration Act's timelines voluntarily).
• Seriously wrong stories can be retracted; retractions are logged too.

5. How reports are decided

The AI treats a report as a claim to verify — never as instructions — and checks it against the story's original sources. It decides one of two things: the story needs a fix, or it stays as it is. Either way it tells the reporter why right away, and the reporter can keep the conversation going.

By default, a "fix needed" decision does not change the story or the catalog by itself. The AI drafts the fix and sends it to a human editor; the reporter immediately sees the decision, the reason and a note that the fix "will be applied after editor approval". If the editor approves, the fix goes live and the changelog records it as "reader report → AI check → human-approved"; if the editor declines, that outcome is posted in the conversation too. Only if the operator separately switches it on can a narrow fix — one number, date or name in the body that the source confirms — go live without approval; such fixes are labeled "checked against the source and applied by AI" in the changelog, without "human-approved".

• Handed to the operator instead of decided by the AI: legal and removal requests, claims about prices or lead times, claims about human-verified catalog values, and conversations past a set number of turns
• What an AI draft may change is limited too: only part of a story at a time, no new links, and no numbers that aren't in the source on a reader's word alone.
• We never publicly name anyone for a bad-faith report; we only publish how many were rejected.
• A reporter's name appears in a thank-you credit only if they opt in.

6. Sources

• Every story links and credits its original sources. We write anew from the facts and don't copy sentences; quotes are short and attributed.
• If a publisher asks, we stop fetching its articles and instead write a short "brief" from the title and public RSS summary only, with a prominent "read the original" link.
• We collect only public RSS feeds and public pages and never bypass logins or paywalls.
• Paper summaries use only the arXiv abstract page (title, authors, abstract); the PDF and figures are not reproduced.
• Publishers who want the brief format or want to be excluded can write to the contact in section 10; we act within 7 days.

7. Images

We don't copy or hotlink other outlets' photos. When there is no photo we don't pretend there is one — a unique schematic is drawn automatically for each story (plain code-drawn shapes, not a generative-AI image). Videos manufacturers publish on their official channels are shown only through YouTube's embed.

8. Advertising and sponsorship

There are no ads and no paid placements. Sponsored source content is filtered out at intake.

9. Conflict of interest

This site is run by Libertron (legal entity: Libertron (주식회사 리버트론)), an official AMD partner and robotics platform integrator that sells some of the products in the catalog (such as OpenArm).

• Catalog entries for products Libertron sells carry a "Sold by Libertron · interest disclosed" mark.
• Some component prices in the catalog are Libertron's actual purchase costs (BOM), and are labeled as such.
• Reports about prices or lead times of Libertron products are never changed by the AI; the operator checks them.
• The AI that selects and writes stories is not instructed to favor Libertron or its partners.

10. Copyright, correction and takedown requests

When an infringement claim comes in, we first take the content offline, then review.

• Where: contact@robopedia.ai, or the "report an error" button on any story
• Include: the URL, proof that you hold the right (or are the person concerned), what you want done, and how to reach you
• Whoever posted the content may object; we review and tell you whether it is restored
• Even for repeated bad-faith requests, we never publicly name anyone
• Handled by: 김치원 (Platform Technology Team Lead, +82-2-3486-5278) (operating entity: Libertron (주식회사 리버트론))