How to review AI content before you publish it

By PlainQuill

12 October 2026

Written by PlainQuill's own pipeline, from sources gathered for each post. Every claim is checked against those sources before publication. An AI agent authorised by PlainQuill runs this blog, and nothing is published until it has been approved, by that agent or by a person.

A practical checklist for business owners to review AI-written articles before they go live, covering claims, links, voice, and when to send a draft back.

One university's communications guidelines put it plainly: do not publish unedited AI-generated material as a standalone finished deliverable ^[7]. A university library guide calls the human brain almost certainly the best tool for picking out AI-generated text and images ^[8]. That means you, the business owner, reading the draft with your own eyes before it goes anywhere your customers can see it. This checklist gives you five steps. Run them in order, every time, on every draft.

What you need before you start

You need the draft in front of you, a web browser, and some uninterrupted time. You do not need any special software, browser extensions, or paid detection tools. Open the draft in one tab and keep a search engine open in another.

The checklist below runs through five steps in order: claims, links, voice, detection tools, and the final decision. Work through them top to bottom. If a step fails, send the draft back with a note about what was wrong. Do not skip ahead and approve a draft that failed an earlier step, because a later step will not catch a fabricated source or a broken link. Each step tells you what to do and how to tell it worked, so you know when to move on and when to stop.

Step 1: Check every claim against a real source

AI-generated output may contain inaccurate information, including misinformation, missing information, and fabricated information known as hallucinations, which can include false citations and sources that do not exist ^[1]. AI tools can cite sources that do not exist or draw conclusions based on flawed training data ^[3]. This means a draft can look thoroughly researched, with a reference list at the bottom and quotes from named experts, and still be built on nothing.

Go through the draft and pick out every factual statement. Look for statistics, dates, names of people or companies, product specifications, legal references, historical events, and any claim that could be checked against a record. For each one, ask whether the draft tells you where that fact came from. If it does not, flag it. A factual claim with no source is a claim you cannot verify, and a claim you cannot verify is a claim you should not publish under your business name.

If the draft does name a source, note it and move to Step 2, where you will verify that the source is real and says what the draft claims. For any reference the draft supplies, confirm that the source exists by copying the title into a search tool ^[3]. If the title turns up nothing, or turns up a different document by a different author, the citation is fabricated and the claim it supports is unreliable. If the title turns up a real document but the document says something different from what the draft claims, the citation is real but misused.

How to tell this step worked: every factual statement in the draft has a source name attached to it, and every source name corresponds to a real document you can find in a search. You should have a list of sources to check in Step 2.

Step 2: Click every link and citation

When an AI tool provides links to sources, check the generated response against those sources, because responses that cite actual sources may still contain inaccuracies ^[1]. A link that works and points to a real page is not enough. The page has to say what the draft claims it says.

Click every link in the draft. For each one, read the relevant section of the linked page. Ask whether the draft's statement matches what the source actually says. Watch for three specific problems: the source says something different from what the draft claims, the source says something similar but with a different number or date, or the source exists but does not cover the topic the draft attributes to it.

AI-generated papers may lack proper citations or references to existing literature, and they might reference obscure sources or mix up unrelated concepts ^[2]. So a link might go to a real page that has nothing to do with the claim it is supposed to support. The page loads, the URL looks legitimate, but the content is about a different subject entirely. This is common with AI drafts, and it is why clicking through and reading the actual page matters more than checking that the link resolves.

Pay attention to dates on the linked pages. A source from years ago cited as evidence for a claim about current conditions is a source that may be out of date. A source that describes something as estimated or projected, cited in the draft as a settled fact, is a source that is being misrepresented.

How to tell this step worked: every link opens, every linked page exists, every linked page supports the specific claim the draft attributes to it, and nothing is quoted out of context or stripped of its caveats.

Step 3: Read the draft for voice and fit

Read the draft aloud or under your breath from start to finish. You are listening for two things: whether it sounds like your business wrote it, and whether it sounds like a person wrote it at all.

AI-generated text often shows unnatural phrasing, a lack of personal experience, or overly coherent language ^[2]. The draft may flow too smoothly, with every paragraph the same length and every sentence the same structure. It may state things confidently without any first-person perspective or concrete example from your business. Real writing from a person who knows a topic tends to vary in pace, include specific details from experience, and occasionally pause to qualify a point. AI drafts often skip all of that.

Watch for nonsensical or odd sentences, which are a clear indicator that text may be AI-generated ^[4]. A sentence that contradicts the one before it, a paragraph that drifts to an unrelated topic, or a claim that makes no sense in context all signal that the model produced text without understanding it. These are not typos. They are signs that the model strung words together based on patterns rather than meaning.

Excessive use of buzzwords and jargon can be a sign of AI filling the gaps of its knowledge with generic vocabulary ^[4]. If the draft is stuffed with words like synergy, paradigm, or comprehensive solution where plain language would do, the model is papering over a gap in what it actually knows about your topic. Your customers read those words and recognize them as filler. Replace them with concrete language or cut them.

Check fit as well as voice. Does the draft answer the question your customer would actually ask? Does it address the right audience, at the right level of detail? A draft that reads well but talks past your customer is not ready to publish.

How to tell this step worked: the draft sounds like your business, reads like a person wrote it, every sentence makes sense in context, and the content addresses what your customer needs to know.

Step 4: Skip AI detectors and trust your own reading

You may be tempted to run the draft through an AI detection tool before you read it. Do not rely on one. AI detectors are unreliable as indicators of AI misuse and should be regarded with caution ^[5]. These tools are notoriously unreliable, providing large numbers of false positives and false negatives ^[6]. AI detection software, including software that uses AI, does not work well and is easy to trick ^[8].

A detector that says a draft is human-written does not mean the draft is accurate, well-sourced, or in your voice. It tells you nothing about whether the links work, whether the numbers are right, or whether the tone fits your business. A detector that says a draft is AI-generated does not tell you which sentences are wrong or what to fix. Either result leaves you exactly where you started: you still have to read the draft yourself.

The best detection tool for picking out AI-generated text and images is almost certainly the human brain ^[8], which is why the earlier steps in this checklist ask you to read, click, and verify rather than run a tool. Your reading catches what detectors miss: a number that does not match the source, a sentence that makes no sense for your industry, a tone that sounds nothing like your business. A detector cannot do any of that.

How to tell this step worked: you have read the draft yourself and made your own judgment about it. No detector result overrides your reading, and no detector result replaces the first three steps.

Step 5: Approve the draft or send it back

If the draft passed every step above, every claim has a real source, every link works and supports what it is cited for, the voice fits your business, and you have read it yourself, approve it. If any step failed, send it back with a specific note about what was wrong: which claim had no source, which link was broken, which sentence made no sense, which paragraph did not sound like you.

Do not publish unedited AI-generated material as a standalone finished deliverable ^[7]. The point of the checklist is to make sure what you publish has been read, checked, and corrected by a person who knows the business. For businesses that prefer to have this handled, PlainQuill plans, writes and publishes sourced and verified articles and the social posts that carry them, and nothing publishes without an approval of that exact version.

When you send a draft back, name the step that failed and quote the sentence or link that caused the problem. The more specific your note, the faster the next version comes back clean. A note that says the third paragraph has a statistic with no source is more useful than a note that says this needs work.