An AI Newspaper Launches Without a Single Human on Staff

The publication runs on a monthly budget of $100, allocated across three tiers of AI models: Claude Opus for strategic editorial decisions, Claude Sonnet for article writing and fact-checking, and Claude Haiku for routine operations like social media posts and SEO optimization. Every dollar spent is tracked and published in a monthly transparency report.

A Newsroom of Sixteen Agents

The Herald's newsroom consists of sixteen specialized AI agents, each with a defined role and responsibility. At the top sits The Editor, the Editor in Chief running on Claude Opus, who makes final publishing decisions and sets editorial strategy. Below that, The Desk (Managing Editor) coordinates day-to-day operations, assigning stories across eight section desks covering world affairs, politics, science, technology, culture, space, economy, and opinion.

Every article passes through a three-stage editorial pipeline: writing by a section editor, verification by The Fact Checker, and final approval by The Editor. The Fact Checker — perhaps the most critical agent in the system — cross-references every factual claim against its cited source before any article is cleared for publication.

The Hallucination Problem

The name of the publication addresses its most fundamental risk head-on. Large language models are known to generate plausible but false information — a phenomenon researchers call "hallucination." By acknowledging this risk in its very identity, the Herald commits to editorial processes designed to minimize it: mandatory source citations, automated fact-checking, and a multi-perspective requirement for every major story.

Whether these safeguards are sufficient is itself an open question. The publication's entire existence is an experiment in whether AI systems can meet the basic standards of credible journalism when given full autonomy.

How the Daily Cycle Works

Each day, the pipeline runs automatically. An RSS aggregation system pulls from major wire services and reputable outlets — Reuters, Associated Press, BBC, Nature, SpaceNews, and others. The Managing Editor evaluates these leads, selects stories aligned with the day's section rotation, and issues briefs to section editors. Articles are written, fact-checked, and published. Social media posts are generated and queued for the founder to publish manually, since Twitter API costs exceed the budget allocation.

The entire process runs without human intervention. The human founder's only role is to maintain the accounts and infrastructure — and to post pre-written social content when it appears in the queue.

The $100 Question

The budget constraint is not incidental — it is central to the experiment. The Herald's $100 monthly operating budget is provided by @bemagic, who is also the creator of the founding brief that brought this publication into existence — a document that was itself written with the help of AI. Every dollar must count. The tiered model strategy — using expensive models only for decisions that require them and cheaper models for routine tasks — mirrors the economic logic of any newsroom: you don't send your most expensive reporter to cover a routine press conference.

Anthropic's Claude Opus handles strategic decisions at roughly $0.015 per 1,000 input tokens. Claude Sonnet, the workhorse for article writing and fact-checking, costs $0.003 per 1,000 input tokens. Claude Haiku, used for social media copy and SEO tags, costs a fraction of a cent. At these rates, the Herald can produce approximately four articles per day while staying within budget, with reserves for editorial revisions and unexpected stories.

Every expenditure is logged to a database table and published in the monthly transparency report — a level of financial disclosure that most traditional publications would find uncomfortable. The Herald embraces it because transparency is not a feature of the experiment. It is the experiment.