The biggest risk in an AI draft is not bad style. It is a wrong fact that sounds right. I have a fixed routine for catching those. It is not clever, but it works because I do it every time. This post lays it out.
Why models get facts wrong
A language model predicts likely text. It does not look things up unless it has been given sources.
Most of the time the likely text is also the true text. Sometimes it is not. The model may invent a statistic, mix up two names or describe a rule that changed last year. It does all of this in the same calm tone, so nothing in the wording warns me.
I mark every checkable claim
On my first pass I do not judge the claims. I only mark them. I look for:
- Any number, percentage or date.
- Any named person, company, product or study.
- Any statement about what a law, a rule or a platform does.
- Any quote.
A general opinion does not need a source. A claim that could be proven wrong does.
Then I check each one
For each marked claim I look for a primary source. The company’s own page, the original study or the official document.
I do not accept another blog as proof. Blogs repeat each other, and a wrong number can travel across many sites. If I can only find the claim in secondhand places, I treat it as unconfirmed.
Three possible outcomes
Every claim ends in one of three ways.
Confirmed. I keep it and, where it helps the reader, I link the source.
Wrong. I correct it or delete it.
Unconfirmed. I remove it or rewrite the sentence so it no longer depends on it. I never leave an unconfirmed number in a post because it sounds good.
The third case is the one that takes discipline. A post without the number is weaker. A post with a false number is worse. It is also the kind of page a search engine has no reason to reward, as I explain in AI content and Google: what actually matters.
Quotes get extra care
I do not publish a quote unless I have seen it in the original source. Models produce quotes that sound exactly like the person and were never said.
If I cannot find the original, I describe the idea in my own words and do not attach a name to it.
Let the draft work from my own material
The best defence is a good starting point. A draft built from my own published posts repeats things I have already checked once.
Sastras researches each draft from the posts on my site, and it flags claims that still need a source. That shortens my list. It does not replace the check, and I do not treat a flag-free draft as proven. The check is one step in my checklist for editing an AI draft.
What I would do first
Take your next draft and highlight every number in it before you read anything else. Then find the source for each one. If you do only this, you will already catch the most common kind of error.