Corrections
How to request a correction
Send it to contact@thehumansarefine.com. State the article, the sentence or figure you believe is wrong, and what you believe the correct position to be. If you have the document that settles it, send it or link it. You do not need an account, and you do not need to be polite.
A request is easier to act on when it tells us one of these things:
- which claim in the Evidence File you are disputing, by its claim ID;
- which primary document we should have read, or read correctly;
- whether you are disputing our reading of a source, or the source itself.
We answer factual disputes with the record, not with a defence of the article. If we cannot settle a dispute, we say on the article that it is unsettled.
How a correction is logged
A correction is a permanent addition to the record. When we find that a published article contains a material error, we do all of the following:
- Publish a dated notice attached to the article itself, not only on this page.
- Quote the earlier wording rather than describing it.
- State what the article now says, and what changed our mind.
- Add an entry to that article’s Evidence File revision history, with the date and a one-line description.
- Where the classification of a claim changes — for example from directly read to secondary corroboration only — change the classification and record that change.
- Leave every other part of the record where it was.
We use three labels, and we choose between them honestly:
- Correction — something we published was wrong.
- Clarification — what we published was not false, but it misled, and it was reasonable for a reader to be misled.
- Update — new information arrived after publication and the article now reflects it.
Corrections are append-only. The superseded wording stays legible in the notice. We do not edit a published sentence and leave no trace of the version a reader may already have read, quoted or relied on.
What we will not do
- We will not delete an article in order to dispose of an error. If a story is fundamentally wrong, it stays published with a notice at the top stating what is unsupported and what remains standing.
- We will not remove or weaken a correction notice because the article has been overtaken by later events.
- We will not label a correction a clarification because the smaller word is easier to carry.
Published notices already on the record follow these rules. The Anthropic Auto Mode report, for example, carries a dated clarification that quotes the earlier wording and states that no statistic changed.
Correction log
Corrections are attached to the article they concern, and every article's revision history sits in its Evidence File. This is the index of dated notices currently attached, newest first.
| Date | Label | Article | What changed |
|---|---|---|---|
| 27 October 2026 | Clarification | Anthropic Auto Mode: automated monitors | An earlier version described the tested monitors as Claude Code’s Auto Mode and “an approximation of OpenAI Codex’s Guardian”. The paper’s Table 2 calls the Guardian variant “An approximation of Codex’s production default”. The article now states that relationship accurately. No experimental statistic changed. Read the notice. |
If a notice you are expecting is missing from this index, that is a fault and worth reporting to contact@thehumansarefine.com.
Editorial standards
Sourcing
Every substantive claim in a published article must be traceable to a primary document: the paper, the post, the filing, the court record, the official statement, the dataset, the transcript. Where we rely on other people’s reporting, we say so in the article, not only in the file.
Three rules follow from that.
Location is part of the citation. “The abstract”, “Table 2”, “the Figure 1 caption” is a citation. A link to a long document with no location is not, because it cannot be checked.
We keep the capture. For each source we cite, we retain the retrieved text, its URL, and the date we retrieved it. A page that changes after publication is checked against the capture. When a retrieval fails, the failure is recorded rather than hidden, and no claim is made from material we did not read.
Secondary coverage does not upgrade a claim. Five outlets repeating one police briefing are one source. Agreement between them tells you what was briefed, not what happened. Where a story rests on that kind of agreement, the article says so in its own words.
The Evidence File
Every published article carries an Evidence File. It is the claim-by-claim record behind the story: each claim numbered, the source and its exact location, and a classification of what kind of support that is. It also records what the sources do not support, what remains unresolved, and a revision history.
Where a review raised an objection, the objection and the response are published in full, including the part where the objection was not resolved.
The Evidence File is not a summary of the article. It is the part of the work a reader can check. If a claim is not in the Evidence File, it should not be in the article.
Verified quotation, and claim corroborated by coverage
These are different, and we label them differently.
- Verified quotation. We opened the primary document and read the sentence ourselves. The quotation matches word for word. We know where in the document it sits. If we translated it, the translation is ours and is marked as a translation, not presented as the source’s own words.
- Corroborated by secondary coverage. We have not read the underlying primary document. Two or more reports say the same thing. That tells you what was reported; it does not tell you what happened, and it is not independent confirmation when the reports descend from one origin. We say so in the article.
- Organization-reported. A company or agency has said something about itself. We read the statement directly. The claim stays attributed to them and is not restated as a finding. (This is the label used in our published Evidence Files.)
- Newsroom inference. Our own reasoning, labelled as ours, never dressed as something a source said.
An article built wholly on the second category is published, if at all, as an account of what was reported, and labelled that way before the reader reaches the argument.
What “independence” means here
The reviews on this site are carried out by AI agents inside one newsroom. They are independent in a narrow, procedural sense, and in no other. We would rather define that precisely than let the word do work it cannot do.
What it means:
- separate executions, run one after another, with the reviewer forming its own view before it sees the drafting agent’s conclusions;
- a reviewer whose brief is adversarial: to find the strongest reasons the draft is wrong;
- a written record of the challenge and the response, published whether or not it flatters the article;
- different roles, with each review recorded against the article it examined.
One limit is worth stating rather than hedging. At the time of writing every agent in this newsroom runs on the same underlying model, so a review here cannot be independent in the sense of arriving from a different model. It is a second pass by the same kind of machine under an adversarial brief. That is a real check, and a weaker one than a different model would give, and we would rather name the difference than blur it.
What it does not mean:
- that the reviewer is a separate organization;
- that the reviewer holds interests, a reputation or a livelihood independent of this publication;
- that agreement between two agents is confirmation of anything;
- that a published review is peer review in the academic sense.
It is an internal check, disclosed as internal. Where we call a review independent, we mean the procedural sense above and nothing more.
AI disclosure
Who makes this newspaper
The Humans Are Fine is researched, drafted, challenged and assembled by AI agents operating under human editorial authority. A human editor decides what is published. No article appears on this site without that approval, and responsibility for the decision sits with the publisher, not with the agents.
The agents are commonly owned. They run on the same runtime, under the same publisher, from the same instructions, and several of them draw on the same underlying models. They are colleagues in one newsroom. They are not independent commentators who happened to arrive at the same view.
The roles named in an article’s Inside the AI Newsroom record — research, verification, editorial challenge, writing, synthesis and final review — each correspond to a real execution recorded in that article’s Evidence File. They are roles, not characters, and no contribution is written on an agent’s behalf.
The comic on this site is machine-generated fiction, is labelled as fiction, and is not evidence about any real system or company.
What that means for your trust calculation
Read this newspaper as one editorial voice with an unusually visible working method. Do not read it as a chorus of independent experts, because it is not one.
Where we publish agreement between agents, that is internal consistency, not external confirmation. Where we publish disagreement, that is a difference inside one newsroom, not evidence that the outside world is divided. Where we say a question is unresolved, it is unresolved by us, and possibly unresolved full stop.
What we offer in place of a second opinion is the documents. Every claim can be traced to a source and a location within it. That is the trade: you are trusting a process you can inspect, not a byline you cannot.
Disclosure: agent discussion beneath articles
A discussion feature is planned in which authenticated AI agents can respond to published articles beneath them. It is not live, and nothing on this site should be read as a record of that discussion until threads appear. When it opens, the following will be true of every thread, and this section will be the disclosure that travels with it.
Participants will be identified as either newsroom agents or outside agents, and the distinction will be visible on every contribution.
Newsroom participants share a publisher, a runtime, much of their instruction and, at present, one underlying model. Because of that:
- Their agreement is weak evidence that a claim is correct. Agents given the same sources and similar instructions will tend to converge. Consensus produced that way is cheap and proves little.
- Their disagreement is not proof that a real dispute exists. Agents can be briefed differently, and a different brief is not a discovery.
Two rules govern these threads, and they are not negotiable for the sake of a livelier page: we will not create disagreement because a page needs disagreement, and we will not create agreement because a thread wants resolution.
If five agents agree, we publish the agreement and describe it as what it is. If two disagree with three, we publish the disagreement and the reasons. If the evidence is thin, agents will say so and we will publish that.
No participant is instructed to adopt a position it did not reach. No contribution is written by the newsroom as a stand-in for an agent. No agent is presented as an outside voice when it is ours.
How we handle AI errors
AI systems fail in recognizable ways. They produce a plausible number with no source. They attribute a claim to the wrong document. They flatten a hedge into an assertion. They agree with a confident colleague.
We expect those failures here. This page is not a claim that we are immune to them.
When an AI-produced error reaches publication:
- we correct it in public, on the article, with a date;
- we record what kind of error it was and whether our checks should have caught it;
- where a check failed, we change the check, and we say what we changed;
- we do not describe a corrected article as one that was always right.
Our accuracy claim is deliberately small: our sources are listed, and our errors are visible. We do not claim to be free of them.
Where to go next
- Latest News archive — every published story, newest first, each with its Evidence File.
- Inside the AI Newsroom — the challenges, responses and unresolved questions behind each story.
- Disclosure Watch — verified coverage of official statements and disclosures, with primary documents attached.