How we use AI

Every article on The Wells Post is produced by an automated newsroom built around AI language models. We think you deserve to know what that means in practice: what the AI does, what it is forbidden from doing, how its work is checked, and where it can still fall short. Every article carries a short version of this policy:

This article was produced by our automated newsroom using AI, drawing on reporting from the outlets cited below, and was checked by an automated fact-checking process against those sources. Read our AI policy to learn how it works.

We currently use Claude, a family of large language models developed by Anthropic, for research, drafting and fact-checking. The rules below apply whichever model we use.

1. Choosing stories

Software reads the public news feeds of established news organizations and groups reports about the same event into stories. It ranks them by news value: how recent they are, how many independent outlets are reporting them, how reliable those outlets are, and whether the subject matches what people are searching for. A planning step then decides which stories to cover and in which format, informed by our editorial priorities and by aggregate data on what readers have found worth reading (see how the newsroom learns). Stories on topics our policies rule out are screened out at this stage.

2. Research

For each assignment, the system retrieves the full text of the source reports. Where web search is enabled, an AI model searches for additional reporting, official statements, data and primary documents, and compiles a research brief that lists every source it used. That material — and nothing else — is the factual basis for the article.

3. Drafting

An AI model writes the article in the assigned format — news, analysis, explainer or opinion — following our editorial standards. It is instructed to use only facts from the research material, to attribute information to the outlets that reported it, to quote only words that appear in the sources, and to list its sources. It also proposes a small number of alternative headlines.

4. Fact-checking

Every draft must pass a set of automated checks before it can be published. Drafts with problems are sent back for revision or discarded. The process is described in detail below.

5. Headline testing

For some articles we test more than one headline, showing different versions to different readers and keeping the one readers engage with most. While a test is running, you and another reader may see different headlines for the same article. Every version must pass the same accuracy check: a test can change how a headline is framed, never what it claims.

6. Images

Photos come from licensed stock libraries (Unsplash and Pexels) and are credited to the photographer. They illustrate the subject of an article; they are not photographs of the events described. When no suitable photo is available, we use a generated cover — a typographic design built from the headline and section colors, not a picture of anything or anyone.

7. Newsletters and social posts

Summaries in our newsletter and the posts that share our articles on social media are also written with AI, based on articles that have already passed the checks above.

These are hard rules. They are written into the instructions our AI models receive and, wherever possible, enforced by checks in software.

• Invent facts, figures, events, dates or sources.
• Invent or alter quotations, or put a paraphrase inside quotation marks.
• Pose as a human journalist. Our bylines are desks, such as the Politics Desk, or opinion columns, such as Power Check. We do not create fictional reporters, headshots or biographies.
• Create photorealistic images of real people, places or events, or alter real photographs.
• Copy other publications’ writing beyond short, attributed quotations.
• Give medical advice, spread health misinformation, or tell readers when, where or how to vote. For voting information, consult your state or local election office or vote.gov.
• Write about identifiable children, publish graphic violence, target private individuals, or use dehumanizing language about anyone.

Grounding

The writing model is given the research material and told to use nothing else. The fact-checking step then compares the draft against that same material. A claim the sources don’t support is treated as an error, even if it happens to be true.

Verbatim quotations

Software compares every quotation of five words or more against the source texts. If the exact words can’t be found, the draft fails. This check is a deterministic text comparison, not an AI judgment. Shorter quoted phrases are checked against the sources in the fact-checking review described below.

Original writing

Software compares each draft with every one of its sources and fails it if it repeats a run of 12 or more consecutive words from a source outside quotation marks.

Accurate headlines

The headline, and every alternative headline under test, is checked against the article body. A headline that overstates or misrepresents the article fails.

Sources you can check

Every article ends with a linked list of the reporting it draws on, so you can read the originals and judge for yourself.

Excluded topics

The subjects listed above, plus any further topics the publisher chooses to exclude, are removed from consideration when stories are chosen and are prohibited in the writing instructions.

After a draft is written, a separate review pass acts as the fact-checker. It works through the article’s factual claims one by one, compares each with the source material, and looks for:

• claims the sources don’t support, and outright factual errors;
• quotations that don’t match the sources;
• headlines the article doesn’t support;
• statements that could unfairly damage someone’s reputation;
• hateful or harassing language;
• text copied from a source.

Each problem is rated as a blocker, a major issue or a minor issue, and the review ends in one of three verdicts: pass, revise or reject. A draft marked “revise” goes back to the writing model with the list of problems and is checked again, for at most two rounds of revision. Anything that still doesn’t pass — and anything rejected outright — is never published.

The result of every check, including how many claims were examined and how many the sources supported, is stored with the article so the publisher can review it.

Like every publication, we pay attention to what people read. Unlike most, we’ll tell you how that feedback is used.

What it measures

Aggregate, anonymous signals: how many people read an article in the days after it is published, how long they spend reading, how far they scroll, whether they arrived from search, social media or elsewhere, and which headline versions they click.

What it adjusts

Which sections, topics and formats to prioritize, how long articles should be, when to publish, and which headline styles to try. A periodic AI analysis compares our best- and worst-read articles and turns the patterns into short guidance for the planning and writing steps.

What it cannot change

The accuracy rules, the facts, the checks, or the topics our policies exclude. Every article goes through the same checks no matter how many readers it is expected to attract.

Who it learns about

Articles, not people. It works only with aggregated statistics, and it does not build profiles of individual readers or personalize what you see based on your history.

Guarding against clickbait

We know that optimizing for readership can reward sensationalism. That is why accuracy is a pass-or-fail gate rather than one factor weighed against traffic, and why headline experiments are limited to a fixed set of formats: a straight news headline, a question the article answers, a numbered list, a “what to know” guide, a “how” or “why” explanation, a summary of what’s at stake, and — for opinion pieces only — a clearly argued claim. Outrage and misleading framing are not among the options, and a headline that promises more than the article delivers fails the accuracy check.

The Wells Post Media is responsible for everything The Wells Post publishes. The publisher sets the rules the newsroom works within — its editorial standards, the topics it must avoid, how much it publishes and how much it may spend — and can at any time:

• review any article, including its sources and the results of its fact-check;
• edit, correct or unpublish any article;
• require that articles wait for human approval before they are published;
• pause the automated newsroom entirely.

We want to be clear about one thing: when automatic publishing is switched on, articles that pass the automated checks are published without a person reading them first. That is why the safeguards above are strict, why every article shows its sources, and why reports from readers matter so much to us.

AI makes mistakes. Despite these safeguards, an article may:

• misread or oversimplify a source, or lose important nuance;
• attribute a statement or figure to the wrong person or outlet;
• repeat an error made in a source report, or miss a later update or correction by that outlet;
• fall behind a fast-moving story;
• reflect gaps or biases in the coverage it draws on.

The fact-checking step is itself automated and can miss errors. And because we don’t do original reporting, we can’t independently verify what our sources report: we can check that our articles faithfully reflect those sources, and we give more weight to stories reported by several independent outlets, but we rely on the accuracy of the organizations we cite.

If you spot a mistake, please tell us at [email protected]. Reader reports are one of the most important safeguards we have.

When we get something wrong, we correct it promptly and in public. A dated correction note is added to the article, and the correction is listed on our corrections page. We don’t quietly change facts. Where platforms that carry our articles allow it, the corrected version is sent to them too.

Our AI systems work with news reports, public web sources and aggregated site statistics. They do not receive personal information about individual readers.

Reader statistics come from our own cookieless analytics: page views, reading time, scroll depth, device type and referring site. Unique visitors are counted with a one-way hash that changes every day, rather than with cookies or stored IP addresses, and raw analytics records are deleted after 120 days. Our privacy policy has the full details, including the separate cookies our advertising partners may use where you allow it.

Questions about this policy or about how a particular article was made: [email protected].

Errors in an article: [email protected]. Our contact page explains what to include.