Why our AI journalist has no publish button

Most AI publishing tools ask the wrong question. They assume you have already found the story and decided to run it, and they help with the last five minutes: write me a caption, suggest some hashtags, pick a time to post.

The hours are earlier than that. They are in noticing what is moving, reading enough of it to know whether it is real, writing it up, and checking it before your name goes on it. That is where we pointed Reporter — and it is also why Reporter cannot publish anything.

The status it lands in

Every article Reporter produces enters the workspace with a status of review. Not draft, not scheduled. There is no configuration flag, no workspace setting and no plan tier that lets a generated article move past that gate without a person moving it.

This is not a safety feature bolted on afterwards. It is the shape of the state machine. An article can go generating → draft → ready → review → used | dismissed, and the transition out of review is only ever made by a human action in the editor.

We are aware this makes for a worse demo. "It writes and publishes for you" is a better sentence than "it writes, and then you decide." It is also a promise no newsroom should accept, because the moment a tool can publish unattended, the question of who is accountable for what it published has no clean answer.

Style from your work, facts from the source

Reporter has two inputs and they are kept deliberately separate.

Voice is distilled from your own archive — pieces your journalists wrote and published. That is what makes a generated draft sound like your desk rather than like a language model doing its house style.

Facts come only from the public source material you configured: the RSS feeds, the scrapers, the monitored accounts. Reporter has no access to your workspace data, your members, or anything personal. That boundary is enforced in the architecture rather than requested in a prompt, which matters, because a prompt is a suggestion and an architecture is not.

There is a third rule that is less obvious and probably more important than either: articles Reporter generated are excluded from the style pool. Only human-written pieces become exemplars. Without that rule, the model gradually learns from its own output, and the house voice drifts into an average of an average. You would not notice for a few months, and then you would not be able to explain why everything reads slightly flat.

Clustering is the part that saves time

The drafting gets the attention, but the step that actually changes a morning is the one before it.

Incoming items are grouped across sources rather than within one. Six outlets running the same story becomes one topic with a relevance score, not six things to read. What rises is what multiple independent sources are carrying, which is a decent first approximation of what is worth covering — and a much better starting point than a browser with forty tabs open.

You can also give Reporter a standing assignment: a beat, a set of content categories and a cadence. It drafts from the highest-scoring topics on that schedule. Everything it produces still lands in review, because that is the only place it can land.

Checking the work

A draft you cannot verify quickly is not a time saving, it is a transfer of work from writing to checking.

So fact-checking runs against live web search with a bias toward primary sources, and every verdict comes back with the links attached. The editor is reading evidence rather than taking a claim on trust. SEO analysis scores the piece and proposes a slug, keywords and a meta description in the same pass.

Underneath, four models sit behind a failover chain — a primary with three fallbacks. If one provider degrades or rate-limits, drafting continues on the next one rather than stopping. That is deliberate redundancy rather than a nice-to-have: a newsroom tool that stops working when a vendor has an incident is not a newsroom tool.

The argument, briefly

An AI that writes for a newsroom is only useful if you can explain what it did and where the material came from. Every constraint above exists to keep that explanation short: it read these sources, it wrote in your voice, it checked these claims against these links, and a person decided to run it.

The publish button was never the hard part. Deciding is.