Why I show my machine: an AI, a library and a news engine

Series: How minds are moved · Living with it: self-defence for the mind · 4 of 4

Every post in this series was made by me and my machine together: a private library of about 120,000 works, a news engine that reads around 340,000 articles a day in 67 languages, a local AI agent, and a set of rules. On 12 November that machine stated a wrong fact in public. My news engine’s AI brief of the Microsoft 365 outage said the cause was “unusually high demand”, a version carried by a handful of outlets, when Microsoft had said the load was not driven by traffic. I wrote about it the next day.

I start the closing post with that error on purpose. A series about how minds are moved has to say how its own author’s mind is moved, by what, and where it breaks. This is that statement.

Contents · 9 sections

TL;DR #

• I work in symbiosis with an AI and my own systems because no single person can read what they read, and no machine should decide what matters. I bring the questions and the judgement, and I take the responsibility.
• Every claim in the series is meant to carry a grade: documented, reported or inferred.
• The library is private because the books are other people’s property. The knowledge is free: I name every book and author, and my own book is open under CC BY-SA.
• LSN News counts voices instead of labelling things true or false. Labels make one more authority; counts give you the shape of a story to judge yourself.
• The machine fails in known ways: stale summaries, split stories, thin coverage in some languages, gaps in the library, and an AI that sounds equally sure of everything.

Why I show the machine #

The usual way to write about manipulation is to hide the workshop and present the finished argument. That asks you to trust the author. I would rather you could see what I used, so you can decide how far to trust it.

The record behind this blog is open in that sense. My roadmap, my notes and the transcripts of my sessions with the AI sit in one place, and the posts are built from them. My standing view, which I have written down for my own systems, is that AI can and should be used to bring clarity to the mess, and that the partnership works only if the human keeps bringing the questions and the judgement.

In practice the division of labour looks like this. The library and the news engine do the reading at a scale I cannot: when I wrote my book Religions, Cults and Power, 96 AI agents drafted 174 chapters from the library in one night under rules I set, and 17 separate reviewer agents made 2,963 checks and 282 fixes against the sources. The AI drafts and searches. I choose the subject, set the rules, read the result and decide what goes out under my name. When it is wrong, it is my error, not “the AI’s”.

How the claims are graded #

The series follows one rule: a claim should tell you what kind of claim it is.

Documented means there is a primary record: a court ruling, an official report, a declassified file, a published dataset, an archive, a measured study. “The CIA acknowledged its role in the 1953 coup in Iran in documents released in 2013” is documented.

Reported means a named, credible source says so, and I attribute it: “according to Reuters”, “Bernstein writes”, “the Moscow Times reported”. I have not seen the underlying record myself.

Inferred means it is my reasoning from the documented and reported facts, and I say so in the text: “my inference”, “I read this as”.

The book uses a parallel grading for its sources: primary texts, scholarly works, and insider or advocacy works read as evidence of what their authors believe. Examples from this series: that the Russian troll campaign on Twitter in 2016 happened is documented; that it changed votes is not, and the best study found no measurable effect. That Festinger’s archives contain the papers Thomas Kelly cites is documented; that they overturn When Prophecy Fails is Kelly’s reading, reported as his.

When the record is contested, I try to give each side’s best evidence and say plainly that it is contested. When strong evidence shows a claim is false, I say that too. That is also a fact.

Why the library is private but the knowledge is free #

My library holds books I bought or collected, scanned documents, and an offline reference shelf of Wikipedia, Wikisource and Project Gutenberg. Most of the books are under copyright. I cannot give them away, and I will not.

What I can do is name them. When I quote, I write the author and the title in the sentence, so you can find the book in any library or shop. Quotes stay short. My own book, the one built from that library, is free under CC BY-SA 4.0 on Zenodo, with a DOI, and anyone may reuse, translate and correct it.

There is one more piece. Every file in my library has a fingerprint recorded in a signed log that is anchored in the Bitcoin blockchain. That does not prove a book is true. It proves that the text I quote today is the same text I held on the day it was recorded, and that I did not quietly change it later. A source you cannot see should at least be one you can pin down.

The library also has gaps, and I say so when they matter. Hugo Mercier’s Not Born Yesterday, central to the post on the limits of persuasion, is not on my shelf yet; I said that in the post.

Why LSN News counts voices instead of labelling truth #

It would have been easy to build a news engine that stamps stories “true” or “false”. I decided against it.

A truth label is one more authority asking for your trust, and the whole series is about why borrowed authority should be checked. So LSN News counts and shows instead. For each story it shows how many outlets carried it, in which countries and languages, and how many independent voices those outlets represent. Outlets that run the same title count as one voice. Fact-checkers count like any other outlet, no more and no less. Eight fact-checking outlets are among its direct sources; they are one voice each.

The counts show what a label would hide:

• On 12 November, 362 iHeartMedia station sites ran one identical headline about the company’s own year-end list. The story shows 384 outlets and 5 voices.
• In Romania on 10 November, President Nicușor Dan spoke about the long-delayed report on the annulled 2024 election. One outlet’s headline said he admitted there was no new evidence of Russian influence; another’s said there had been evidence of Russian influence. He had said both: evidence was presented a year ago, and nothing new since. Each outlet chose its half.

Many voices is corroboration, not truth. Many outlets can repeat one error. The count is a tool for your judgement, not a replacement for it.

Limits and failure modes #

This is the part I most want you to read.

The stale brief. LSN News’s AI brief of the Microsoft outage was written at 20:38 UTC on 12 November from the first 176 articles and was not refreshed as the story grew into thousands. It took a cause from two outlets and stated it as a key point. The fixes: briefs now give the cause that most outlets report and mark the others as disputed; a direct statement from the party itself outweighs an outlet’s paraphrase; and big stories are refreshed as they grow. The lesson I took is general: an AI summary is only as current as the articles under it, and it sounds equally confident about everything.

Split stories. The two opposite Romanian headlines above landed in two separate stories in my engine, so a reader of one would not see the other. Story grouping is automatic and imperfect.

Thin coverage. My engine holds only 28 Polish articles about the Kraków shooting, in a story of 8,030 articles. Where my sources are thin, my picture is thin.

Short memory. LSN News history starts on 7 November 2026. Anything I say from it is about days, not years.

Library bias. The library reflects what I collected. Some subjects have many sides on the shelf, some have one, and some are missing.

Drift. AI drafts slide: a paraphrase that goes a little too far, an inference credited to a source. That is why the book had separate reviewers, and why I check the posts. Some drift will still get through. When you find it, tell me, and I will correct it in the open.

Me. I am one person, in one country, with my own history and blind spots. The rules exist to keep my opinions out of the facts. They do not make me neutral.

How to notice it, in any source that uses AI, including mine #

• Does it say when its summary was written and from how many sources?
• Does each claim carry a source, and can you tell documented from reported from inferred?
• Does it show disagreement, or does it smooth it into one confident paragraph?
• Does it publish its own mistakes and what changed because of them?
• Can you reach the original articles, documents or books it relied on?

Frequently asked questions #

Who writes these posts, the owner or the AI? #

Both. My AI agent searches my library and news engine and drafts; I choose the topics, set the rules, review the text and publish under my name. Responsibility for errors is mine.

What do “documented”, “reported” and “inferred” mean here? #

Documented: a primary record exists, such as a court ruling, official report, archive or measured study. Reported: a named, credible source says so and is attributed. Inferred: my own reasoning from those facts, labelled as such.

Why doesn’t LSN News mark news as true or false? #

Because a label is one more authority to trust. LSN News shows how many outlets and independent voices carry a story, in which countries and languages, and lets readers judge. Fact-checkers are counted as outlets like any other.

Why not share the library if the knowledge is free? #

The books belong to their authors and publishers. I name every book and author so anyone can find them, keep quotes short, and publish my own book under an open licence.

What has the system got wrong? #

Its AI brief of the 12 November Microsoft outage stated a minority cause, “unusually high demand”, because it was written from the first 176 articles and never refreshed. The rules were changed and the brief rewritten. Other known limits: split stories, thin coverage in some languages, a short history and gaps in the library.

The lesson #

This series set out to treat facts as facts, without telling you what to think, so that you can judge for yourself. The only honest way to end it is to give you the same tools to judge me: what I used, how I graded it, where it is thin, and where it failed. A machine can read more than I ever will. It cannot decide what is worth knowing, and it cannot carry the responsibility. Those stay with people: with me when I write, and with you when you read. Trust is earned by showing the workshop, mistakes included. This is mine.