Automated content creation without invented numbers: a newsroom that checks itself

by Lucas Grilli · October 7, 2026 · 4 min read

Automated content creation can run without invented numbers if you separate who chooses, who writes and who checks, and give the checking to a program instead of another AI. The newsroom I run for one of my projects published 33 posts on X in 12 days, with no review before posting, and skipped the news half of the days because nothing passed the check.

Why does automated content feel risky?

Because a model writes well even when it is wrong. A number rounded badly, a shifted date, a changed verb: "confirms" becoming "upgrades". With your name on it, one such mistake costs more than all the hours saved. The problem is not style, it is verification.

What does publishing every day cost?

According to eclincher's 2026 guide a small business pays $500 to $2,000 a month for 8-12 posts on one or two channels; an entry-level freelancer $300 to $800; an AI tool $6 to $200 a month, which writes and schedules but does not check sources.

Who chooses, who writes, who checks

The newsroom has three roles and none does the others' job. An AI judge (Jev, made to answer yes or no with a probability) scores the day's news on three questions: is it my topic? Is there a concrete fact? Is there a verifiable figure? The writing model gets the best eight stories, not whole feeds: the news run went from about 11 to about 4 million tokens.

The model writes with instructions per channel. Then a program, the gate, downloads the cited articles and looks for every number in the post. If a figure is missing, the post does not leave. A news post also needs two different outlets, and a search page or an aggregator does not count as a source.

Does the gate ever fail?

Yes, which is why you test it on real pages. Testing the kit I wrote a post with three real numbers and an invented "37%": the gate let it through, because one source page had a "17:37" clock in the sidebar. Now times and dates do not count, and where the page marks it only the article body is read. Same post, second run: 37 blocked, the others pass.

The gate checks numbers, not meaning. After a post said "Moody's upgrades Italy" about a confirmed rating, the rule went into the writer's instructions: the verb of the decision is the source's verb.

What it published in 12 days

In September the model spend was $0.0069, with free models and the gate blocking their mistakes. The server costs about €8 a month and is shared with other jobs. On Tuesdays the same newsroom writes and publishes a blog article.

How to start without risk

• two weeks of drafts to approve, then on schedule only the slots that convince you;
• every number copied as written in the source;
• when you correct a post, write the rule that would have stopped it;
• if no story holds up, the slot skips: that is data, not a fault.

Frequently asked questions

Can automated content be published without checking it?

Yes, but only with a mechanical check before it leaves: a program that looks for every number in the cited sources and blocks the post if it cannot find it. Drafts first for a few weeks.

What does an automated newsroom cost?

In the real case $0.0069 of models in September, with free models, plus a server already paid. A paid judge and a tool like Postiz may add a few dollars.

Which channels does it work on?

X, LinkedIn, Facebook and the other Postiz channels, plus a blog or a periodic report.

Can an AI check another AI's numbers?

It can help, but it can be wrong the same way. Checking numbers is better done by a program: download the source and look for the figure, without interpreting.

Where to start

The free guide, the kit for your AI agent and the service are on the Automated Newsroom page.

And a real question: which content do you publish every week that you would rather stop writing by hand? Tell me.