How it works

A borough assembly meets on a weeknight. It votes on a budget, a school bond or a road. In many Alaska towns, no reporter is in the room.

Alaska News covers those meetings anyway. Its stories are drafted by software from the meeting's own recording, and the software is built so it cannot make things up. Every quote in a story was spoken at the meeting. Every outside fact comes from a public record. An editor reads nearly every story.

This page explains how that works, so you can decide how far to trust what you read. The short version comes first. The full method is at the bottom.

What happens to a meeting#

1. It is recorded. 2. Volunteers transcribe it. 3. Speakers are named with care. 4. The facts are gathered first. 5. The draft uses only that list. 6. The draft is checked again. 7. It is reviewed.

It is recorded. Many meetings are livestreamed or posted as video. Some arrive as a file or as text.
Volunteers transcribe it. The audio becomes text on volunteers' own computers, not in a data center.
Speakers are named with care. A quote carries a name only when the software can match the voice to one person on the community's list of known speakers. If it cannot, the story says "an assembly member" instead of guessing.
The facts are gathered first. Before any writing starts, the software collects every quote it may use and every outside fact it may cite. Each quote is checked against the transcript. Each fact is checked against its public document.
The draft uses only that list. The writing step can use those quotes and facts, and nothing else.
The draft is checked again. A separate check, with no AI in it, looks for any quote, name or citation it cannot trace. Anything it finds is flagged for the editor.
It is reviewed. Every draft passes an automated review before it can publish. Some stories wait for an editor first. Editors review nearly every article, most of them after it goes live.

A story names its editor only when that editor did a full review. No article claims a review that did not happen.

What this design prevents#

The quick way to write news with AI is to hand a chatbot a meeting and ask for a summary. That can invent quotes, names and sources. This pipeline is built to block each of them.

Made-up quotes. Made-up names. Made-up sources. Quiet failures.

Made-up quotes. A quote can appear in a story only if those exact words are in the transcript. That check runs before writing begins.
Made-up names. Names that are not on the community's list of known speakers are removed from the transcript before the writing step reads it.
Made-up sources. An outside fact must match the words of a real document before the story can cite it.
Quiet failures. Every step logs what it did. Editors see every flag the system raised. Articles do not get published with unresolved flags.

These are not promises for later. They are the rules the software follows today on alaskanews.com.

Volunteers do the heavy lifting#

Turning hours of meeting audio into text takes a lot of computing power. Communities News does not rent it from a data center. Volunteers lend it.

A volunteer installs an app on their computer. When the computer is idle, the app picks up a meeting, transcribes it and sends the text back. A volunteer in Eagle River, for example, can donate about an hour of computer time to transcribe a four-hour assembly meeting. A volunteer in Wasilla can choose to support Mat-Su coverage, and meetings from that area go to them first.

Idle. Picks up a meeting. Transcribes it. Sends the text back.

This is a choice about cost. A data center would be cheaper per minute. It would also be a single point of failure and a bill that grows with every town covered. Volunteer computers grow with the community instead. They carry no ongoing bill, and they give people a way to support local news without writing a check.

What the AI does, and what people decide#

The AI here is a set of narrow tools. It transcribes audio, sorts topics, drafts articles from the checked quotes and facts, and writes captions for embedded photos and video.

It does not decide what to cover, who is credible, whether a story is fair, or whether a draft should publish. People make those calls: editors, contributors and the readers whose civic life the newsroom serves.

Walter Cronkite and the reporters who came after him, Woodward and Bernstein among them, used recording equipment, telephones, computers and the internet. They did not stop being journalists because of their tools.

The tools changed. The standards did not.

The Communities News platform is a tool. The standards are journalism's.

The details: eight stages#

For readers who want the full method, here is each stage as the software runs it.

1. Source ingestion. The platform takes in a public meeting from a video link, an uploaded file or pasted text. Each meeting belongs to one community, today Alaska News, so coverage stays with the people it serves.

2. Audio transcription. Volunteer computers running the Communities News volunteer app transcribe the audio in chunks. A deploy or a network hiccup costs one chunk, not the whole meeting.

3. Speaker identification. Transcription software labels voices generically, as Speaker A and Speaker B. An AI model that reads the video frames alongside the transcript turns those labels into names. A label becomes a name only when it matches exactly one person in the community's speaker registry. Otherwise it stays generic, and later stages say "an assembly member" instead of guessing. This rule is the reason the platform does not put invented names on the page.

4. Tagging and concept extraction. The software tags the meeting's topics, places and main threads. It also finds the web links cited in the source material and pulls the documents behind them, such as bills, fiscal notes and decisions. That gives the article primary records to draw on, not only the meeting transcript.

5. Evidence gathering. For each thread, the software builds two lists. The quote pool holds every quote allowed in the article, and each one is checked to appear word for word in the source. The fact pool holds every outside fact allowed to be cited, and each one is checked against the published record. Drafting may use these two lists, and only these two.

6. Drafting. A language model writes the article with both lists as hard limits. It is told that every person it names must appear in the quote pool, and every outside claim must trace to the fact pool. After drafting, a validator drops any attribution it cannot trace.

7. Verification. A second check runs with no language model at all. It re-checks every attribution and scans the article for names or citations nothing backs. Any failure raises a flag the editor sees in review.

8. Review. Every draft passes an automated review before it can publish. Some stories are held for an editor first. Editors review nearly every article, most of them after it goes live. The editor sees the source meeting, the checked quotes and facts, any flags, and an AI review of the article. An article credits its editor by name only when a full review has taken place.