What a human actually catches
Every piece we send is read by a person before it reaches a client. That sounds like
a line on a pricing page, so here is what it has actually caught — four real
cases, with the text the model produced and the text that went out instead.
We publish these because the alternative is asking you to take our word for it.
Four cases · August 2026 · all verifiable below
Case 01 · Invented pricing
It wrote a price list for a business that had never given us a price
A wedding cake designer. The brief contained no pricing of any kind —
we had never seen her price guide. The model supplied one anyway, because articles like
this usually have one.
What the model wrote
"Basic cakes start at £300. Medium complexity cakes range from £500 to £800.
Highly intricate designs can exceed £1,000."
What was sent instead
The figures were cut entirely and replaced with a line pointing readers to her own
published price guide.
⚠️Had this gone out, a customer
could have read a price under her name and held her to it. The rule now: never state
a price, fee, package or turnaround that wasn't supplied in the brief — point
the reader at the business instead.
Case 02 · A link to a page that doesn't exist
A GOV.UK link, in an article about registering a death, that went nowhere
Written for a funeral director. The link looked completely ordinary —
right domain, right subject, sitting in a sentence that read perfectly. It was invented.
This one was live on a page before we found it.
What the model wrote
gov.uk/register-death — a page that has never existed. Still returns a
404 today; we re-checked before publishing this.
What was sent instead
gov.uk/register-a-death — the real page. One word's difference.
⚠️Someone who had just lost a
parent could have clicked it. The rule now: no URL goes in unless it was supplied in
the brief or has been checked — and every link in every piece is tested before
delivery. It takes seconds and we should have been doing it from the start.
Case 03 · A rule that quietly expired
The hardest kind: not invented, just two years out of date
The same funeral director's article. The model described how registering a
death worked — accurately, for 2023. The law changed in September 2024. Nothing about
the text looked wrong, because it had been right once.
What the model wrote
The doctor issues the medical certificate, the family carries it to the register office,
and the death must be registered within five days of the death.
What was sent instead
"Since September 2024, every death in England and Wales that is not investigated by a
coroner must also be independently reviewed by a medical examiner … In most cases you
will not be given this certificate to carry yourself — it is sent to the register
office directly … you should register within five days of being told by the medical
examiner's office that you can do so — note that the five days runs from that
confirmation, not from the date of the death itself."
⚠️This was being published under
the name of a funeral director who handles that process every working day. He would have
spotted it immediately. Invented facts read slightly wrong. Expired ones read
perfectly — so anything with a date, a threshold or a legal process in it gets
checked against the source, every time.
Case 04 · Our own numbers
And one we got wrong about ourselves
Not a client piece. Our own dashboard was showing clients a word count for
every article they received. The number came out of a config file — it was the length
we had asked for. Nothing anywhere compared it to the length that actually arrived.
What the dashboard said
Long-form articles: 1,500–2,000 words.
What we measured
422 words on average. Every content type was short. We had been reading a label instead
of measuring the thing.
⚠️Nobody complained, which is the
part that stays with us — it looked fine, so nobody looked. Every piece now
reports its real measured length, and short drafts are sent back for genuine
expansion before anyone sees them. The free sample on this site reports its true word count
too, even when that's under target.
Why we'd rather show you this
None of those four read wrong. Every one would survive a proofread — specific,
confident, better written than most of us manage at 11pm. That's how AI content
actually fails: not obviously, but plausibly and slightly false. And it goes out
under your name, not ours.
Each of these is now blocked in the prompt itself, and a person still reads every piece
before it reaches a dashboard. We'll find a fifth. We'd rather you heard about it from us.
See what it writes for your business →