12 things my own AI tried to slip into my copy

AI & Agents · 2026-10-08

During the redesign of jstov.uk in September 2026, ScriptGrain rewrote 17 of my project descriptions in my voice. Then every line got fact-checked against the original fact sheets, and 12 of them had to be rejected or repaired.

ScriptGrain is mine, which is the awkward part. It measures how a person writes and then writes in that voice, and I'm a solo AI developer in London, so when it slips things into my copy there's nobody else to blame. Just me, marking my own homework and finding it full of fibs.

Small ones, mostly. But small fibs about yourself on your own website are still fibs.

Privacy that got a promotion

The first kind was privacy wording. The source said one thing, plainly, and the rewrite came back with something stronger, along the lines of "that's all they're used for". Which sounds lovely. It sounds like a person who cares. It's also a promise the fact sheet never made.

Here's why I think it happens. When people talk about data, they reassure. They say things like "don't worry, that's all it's for". A voice model has absorbed how confident, kindly people sound on the subject, so it produces that tone whether or not the source earned it.

Not what I want on a page with my name on it.

Things I apparently did personally

This group was the most annoying, because it felt personal.

One line implied I personally read every answer, when the tool does that. Another announced "I built every model myself", which appeared nowhere in the source. Another suggested I'd typed data in by hand, with a phrase like "which I fed in every week". And one just said "I built it" where the original never did.

None of these are wild inventions. That's the problem. They're exactly what a person says about their own work: I built it, I fed it, I read them all. The model was guessing what somebody in my position would naturally claim, and then claiming it on my behalf with total confidence.

Benefits nobody listed

Then came the promises. Ease-of-use lines like "drop it into a scene without fiddling". Invented benefits like "so you can see where a rival is growing". Neither was in the source material.

These are the hardest kind to catch, I reckon. They read well. They read like the sort of thing you say when you've used your own product a lot and you're pleased with it. A benefit that sounds like lived experience is very hard to spot as invented, because lived experience is exactly what it's imitating.

Just the ordinary human thing of selling a bit.

The number that grew up

This one's my favourite, in a wincing sort of way.

There's a public architecture reference for Jarvis, and in it there's a real measurement: a median of 3.0 seconds over 847 turns. A genuine number, measured once. Somewhere in the rewrite, that turned into a standing claim about how the thing performs, as if 3.0 seconds were a guarantee rather than something that was measured.

You can see how it goes. A measurement is a fact about the past; a promise covers every future use. The model took the first and quietly upgraded it to the second, because confident present-tense claims sound more like a person than "it was 3.0 seconds when we measured it".

That line got repaired back to what the number actually is.

Pure noise

Some of the 12 weren't claims at all. They were filler. "Right, that's the lot." "One feed, basically." "Anyway, that's the idea." And a stray "you know" dropped in mid-sentence.

The annoying thing is that some of that sounds like me. That's rather the point of a voice tool. But sprinkled through short project descriptions, those phrases were doing nothing. No information, no clarity. They just sounded like somebody wrapping up a chat, which is what a voice model reaches for when it wants a paragraph to feel finished.

Out they went.

What actually fixed it

I'd love to tell you the answer was a cleverer prompt. It wasn't. A better prompt would only make the model more fluent, and fluency is the thing that was hiding the problem in the first place.

What worked was dull. A fact sheet for each project, listing only what was true. Then a line-by-line diff of the rewrite against it, flagging every line that said something the sheet didn't. That's how the 12 turned up.

The diff doesn't care how good a sentence sounds. A sentence that reads beautifully and says something untrue gets caught exactly as fast as a clumsy one, and I can't talk myself out of it because the wording flattered me.

If you want the longer version of what ScriptGrain measures and how it works, there's a write-up on ScriptGrain.

What I take from it

A voice model adds confident first-person claims and filler because that's what people sound like. People reassure, take credit, round a measurement up into a promise, and tidy an ending so it feels closed. Copy a person well enough and you get all of that too.

Which is what I was afraid of and, weirdly, what I was pleased to find. ScriptGrain caught itself out, more or less, because it had to show its work against sources it couldn't charm. I'd rather find 12 of these in my own drafts than have a stranger find one on the live site.