You can produce a thousand words with a language model in under two minutes. Most of those words are filler: hedged claims, restated questions, transitions that connect nothing to nothing. The model is not lying to you on purpose — it's optimizing for fluency, and fluency is not the same thing as being right or being useful.
Most articles on this site are written with model help. We say so in the byline and in the aiAssisted note at the top of every piece that qualifies. That disclosure only means something if the writing behind it is held to a standard. This page is that standard — the checklist a draft has to survive before it earns a publish date.
What gets a draft killed
There are a few failure modes that send a draft back regardless of how clean it reads. The first is the unverifiable number. A model will happily tell you a tool is "40% faster" or has "over 50,000 users" because that shape of sentence appears constantly in its training data. If we can't trace a figure to a primary source — a pricing page, a changelog, a status dashboard, a benchmark we ran ourselves — the figure comes out. A claim you can't check is worse than no claim, because it spends your trust without earning it.
The second is the fabricated specific. Invented quotes, made-up customer names, a "developer we spoke to" who does not exist. These are the most dangerous outputs a model produces because they read as credible and editorial. We don't write them, and a draft containing one is rejected outright rather than patched.
The third is the article that says nothing. If you can delete a paragraph and lose zero information, the paragraph was decoration. A 1,500-word piece that could have been 600 words wastes your time to inflate a word count, and that math never favors the reader.
The round number is the tell. "10x faster," "1,000+ teams," "saves hours every week" — these are the phrases a model reaches for when it has no actual measurement. When you read one anywhere online, treat it as unsourced until proven otherwise. On this site, those phrases are a rejection trigger, not a flourish.
The checks every draft has to pass
Before a draft is published, it goes through a fixed set of checks. None of them are about style. All of them are about whether the article holds up.
- Every factual claim traces to a source. Pricing, feature availability, version numbers, and limits are checked against the vendor's own pages on the day of writing. Things change; "free tier includes X" can be false by next quarter, so dated claims get a date.
- The recommendation survives a counter-argument. If we suggest a tool, the draft has to name who it is wrong for. A review that only lists upsides is marketing wearing an editorial jacket.
-
No placeholder links ship. Affiliate links route through our own
/go/redirects and resolve to real, approved programs. A broken or fake affiliate URL never reaches production — it's a hard rule in the publish pipeline, not a manual courtesy. - The structure earns its length. Headings describe real sections, not padding. If a section exists only to hit a word target, it gets cut.
- A human reads it end to end. The model drafts; a person decides. Nothing publishes on autopilot without that read, because the model cannot tell the difference between a sentence that is true and a sentence that merely sounds true.
That last point is the whole game. The model is a fast, tireless first-drafter and a confident, occasionally wrong researcher. Treating its output as a starting point rather than a finished product is the only posture that produces something worth your time.
Why we label every AI-assisted piece
Disclosure is not a legal box to tick — though it is also that. The U.S. Federal Trade Commission expects material connections and the nature of content to be clear to readers, and search engines increasingly reward content that demonstrates genuine experience and expertise over generated bulk. Both of those point the same direction: tell people how the work was made.
The aiAssisted flag on a post is set to true whenever a model wrote any part of the body, and the note renders automatically. We would rather over-disclose than have you discover the seams yourself and wonder what else we were quiet about. The standard is simple: if knowing how an article was made would change how much you trust it, you should be told before you read it, not after.
Keeping that standard consistent across dozens of drafts takes a written checklist that doesn't drift. We keep ours — the rejection triggers, the source-verification steps, the disclosure rules — in a shared workspace so every draft is measured against the same bar instead of whatever the writer remembered that day.
The point of all of this is not to make AI-assisted writing sound virtuous. It's to make the distinction that actually matters to you as a reader: not human-versus-machine, but checked-versus-unchecked. A human can write a lazy, unsourced, padded article just as easily as a model can. The bar is the same either way. The only question worth asking about any article is whether someone took responsibility for every claim in it before it reached you. Here, someone did.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)