The strongest objection to this site, and what we can show against it
· 4 min read
When we read what people say they want from news, the sharpest thing said
against a product like this one came from inside the audience most likely to
use it. A professional writer of fifty-five years put it roughly this way: a
reporter has to deliver who, what, where, when, why and how, and it is the last
two that cause the trouble, because both require judgment. So when people ask
for "just the facts", are they asking for the why and the how to be left out?
That is the best argument against this site, and it is not answered by
insisting we would never do such a thing. It is answered by saying exactly what
comes out, and then by counting.
What actually comes out
Over the last 600 articles we hold in full, the engine changed 438 of them —
73 per cent. Across all 600, 2.4 per cent of the words were removed. That
is the size of the intervention: about one word in forty, and in a quarter of
articles, none at all.
The words that carry cause and sequence are not among them. Counting *because,
since, after, following, due to, as a result, although, despite, however,
therefore, according to, said, told, reported, under, amid* and their
neighbours across the same 600 articles: 3,208 before, 3,250 after. Not one was
lost.
The count goes up rather than staying level, and the reason is worth saying
because it is the same argument in miniature. The engine replaces an
attribution verb that grades what was said with one that does not: insisted,
maintained, claimed and suggested all become said. A reader is told who
said a thing and that they said it, and is not told, in the reporter's voice,
whether to believe them. The why survives the edit; a verdict about the why
does not.
Two more counts from the same run. **Every number survived: 1,952 in, 1,952
out. Every quotation mark survived: 1,970 in, 1,970 out** — nothing inside
quotation marks is ever edited, because editing inside a quotation puts words
in somebody's mouth.
Where the objection lands anyway
It still lands in two places, and we would rather write them down than wait to
be told.
A verdict can be the news. When a judge calls conduct disgraceful, the word
is the finding, not the decoration; when a coroner calls a system's failure
avoidable, that is the report. The engine keeps a word that is attributed and
keeps every word inside a quotation, which covers most of these — but the
general rule "a word that grades an event goes" is a rule with genuine
exceptions, and we hold the version we removed from as well, and show both.
Colour can carry information. "The street was silent for an hour
afterwards" is a description doing real work. We take out the words that tell
you how to feel about a fact; we do not always get the line right between that
and a word that is telling you what the scene was. Where we get it wrong there
is a report button on every article, and the reports go in a public
ledger unedited, including the ones we cannot fix.
Why we do not simply write our own version
The obvious escape from all of this is to have a machine rewrite the article.
We will not, and the reason is the same objection from the other side: a
generated sentence can be fluent, neutral and untrue. Everything on this site
is the outlet's own words with some of them removed, or the outlet's own
sentences shortened. No hosted language model touches an article here, and
the summariser is extraction rather than generation: every word in a summary is
a word from the article.
That is a harder engineering problem and a weaker-looking product. It is also
the only version of this we can defend, because a fact cannot be invented by a
process that can only delete.
The part we cannot answer
If the objection is that reading a de-sensationalised report leaves you feeling
informed at a moment when you are not, then no measurement above settles it.
Two per cent of the words and all of the numbers is what we can count; how it
feels to read is not.
What we can do is refuse to hide the alternative. Every article carries the
outlet's headline as filed, their wording beside ours, the count of what we
changed and a link to the original. If you think we have taken something out
that was carrying the story, the original is one tap away and the ledger is
open. A product that is confident about this should be willing to be checked,
and this one is built to be.
Written by Mundane Read. If we have something wrong, tell us and it goes in the corrections ledger.