A number came off a live page recently and the reasoning behind it was impeccable. The page gave a figure for the size of an operator's game catalogue. The operator had since put that catalogue behind a login, so nobody could confirm the number where it came from, and the rule said unconfirmed figures do not ship. Out it came.
The page went out safe, defensible and materially worse. Four independent write-ups carried that figure. Not four rough approximations of it. The same number, counted separately, sitting in public where anyone could find it inside a minute.
What I said when I saw what the rule had done was blunter than this article. Four sources agreeing is a fact. Deleting it because the operator stuck a login in front of their own catalogue is not caution. It's a filing error wearing the clothes of rigour.
Two Rules Wearing the Same Coat
Two separate rules had collapsed into one. The first says never launder somebody else's work into your own voice: don't take their rating, their phrasing or their research and pass it off as your finding. That one is about attribution, it's basic honesty, and it does not bend.
The second says verify before you claim. Just as firm, and about something completely different. That one is about knowledge.
A policy with one dial can only be turned up. Turned far enough, it starts deleting things that are true.
Fold those two together and you get something that reads like good practice and behaves like amnesia. It tells a writer to pretend not to know something four independent parties have published, because the one organisation with the authority to confirm it has put a login page in the way. The rule fires. The check passes. The reader gets less.
What the Ban Was Actually Protecting
The ban on other people's content is there to stop one specific and genuinely serious thing. A writer reads three reviews, absorbs the conclusions, and serves them back up as independent judgement. That's passing off, it's what makes most published comparison worthless, and a model will do it at speed unless something in the process stops it.
None of which has anything to do with a countable number. A catalogue holds however many titles it holds. Where several people have counted separately and agree, the count isn't anybody's property. It's the world, and putting it in your own words isn't citation. It's knowing something.
The Order of Preference
So the rule in my standard is ordered now instead of binary, and that turns out to be the whole difference. Four steps. Stop at the first one that works.
- The primary source, counted or quoted directly. Always try this first, and put real effort in before you give up on it.
- Several independent write-ups that agree. Three or four in agreement is a fact. State it plainly as your own finding, and never name, link or quote them.
- An honest range, or the conservative end of one, where the sources genuinely disagree.
- Leave it out, and only when nothing corroborates it at all.
The version that failed had step one and step four and nothing in between, so every figure behind a login fell straight through to deletion.
The Line Is Invention
This loosens nothing, and the article would be dishonest if it pretended otherwise. The boundary hasn't moved. It just turned out to sit somewhere other than where the rule had drawn it.
Invention is the failure that matters. A figure no source carries, produced because the sentence wanted a number in it, is the most damaging thing that can go out under my name, and models are remarkably good at exactly that: statistics that sound precise with nothing at all underneath them, confident historical detail with no date on it, a plausible round number standing where an awkward real one should be.
Uncertainty Is Not Invention
Comparative claims still fail safe. Any largest, first or fastest that has not been checked against named alternatives in the same working session gets dropped or swapped for a plain figure, and that rule has never been relaxed by a degree. It's the one with an actual failure behind it.
But not knowing something for certain and making something up are different acts, and a policy that can't tell them apart will keep choosing the empty page. Corroboration is not citation either. Those two distinctions carry the whole weight of the thing.
The Cost of the Safe Version
There's a commercial argument underneath the editorial one and it's short. A page with no figure loses to a page that has one.
Assistants and search engines reward the answer that comes in specifics, and readers do the same without being asked. A page carrying exactly the facts every competing page carries has given nobody a reason to rank it and nobody a reason to quote it. No reader is ever going to hear that the source was behind a login, and it wouldn't move them if they did.
Which gives you the question to put to an AI content policy, and it isn't how much the thing forbids. Ask what it has deleted. Every risk-reducing rule in my own documents now carries a second question beside it, asking what it costs on the days it fires correctly, and any rule with no answer comes back out of the file. It's the same overshoot that flattens the prose, one layer down. A policy whose cost nobody has ever measured isn't a policy. It's a reflex with a document attached.