Humanised AI content

Humanised AI content is machine-drafted text revised so that its measurable habits (sentence rhythm, contractions, phrasing, punctuation) fall inside the range of human writing, as distinct from text rewritten only to fool a detector. The revision can be measured: before-and-after counts of the tells, and a voice-match score against a real writer.

The numbers

What humanised AI content is

The definition is narrow on purpose. It's not "AI content that sounds nice" or "AI content a person tidied up a bit". It's a specific process: scan for the known tells, count them, revise, rescan. The full list of what counts as a tell (30 phrases across four groups, plus ten detectable sentence shapes) lives at /tools/ai-cliche-checker and there's no point repeating it here. What matters for this glossary entry is the shape of the process, not the inventory.

Humanising also isn't the same as writing to a brand voice from scratch. It's closer to editing: take a machine draft, strip the giveaways, and if you've got a measured voice profile, pull the prose towards that instead of towards a generic idea of "human". See /glossary/voice-match for what "pulled towards a voice" actually means in numbers.

What changes when you measure it

Here's the thing about humanising: you can't tell if it worked by eye. You have to count.

The counts that move are the tells per 1,000 words, the em dash count, the short-sentence share, and the contraction rate. In SGR-001, running a banking brief through a voice profile took average sentence length from about 17 words to about 11, doubled the share of sentences under eight words (24% to 48%), lifted contractions by roughly 70%, and took ten em dashes down to none. One run, unedited, so treat it as directional.

There's also a voice-match delta: how close the output sits to a target voice, measured, not guessed. In SGR-002, ScriptGrain's drafts scored 0.83 against a study profile and 0.91 against the held-out post itself. Bare GPT-5 scored 0.74 and 0.80. That gap is what "humanised" is supposed to close.

Humanising versus evading a detector

Here's the distinction, plainly: humanising fixes the writing. Evading fixes the score.

Those aren't the same target, and chasing the second one is a trap, because detector scores don't reliably track voice quality. In SGR-002, the arm using GPT-5 with a one-page measured prompt scored highest on a detector's human score (0.78) while missing the target contraction rate by a mile: 0.5 per 1,000 words against a request for roughly 14, with judged formality at 7.2 against a target of 3.7. Meanwhile ScriptGrain's own drafts, which matched voice far better, scored lower on the detector (0.67).

So beating a detector and sounding like a person are different problems, and detectors are unreliable in both directions: some human writing scores as machine-made, some machine writing scores as human. ScriptGrain reports detector scores because they're a data point. It doesn't sell evasion, and it won't ship a bypass feature. The humaniser revises towards natural variation, sentence-length spread, short sentences, contractions, removal of listed tells, and rescoring up to three passes towards a default of 0.85, reporting a miss honestly rather than rounding it up.

What it cannot tell you

Measuring tells and voice-match doesn't tell you whether the content is accurate, useful, or worth publishing. It tells you whether it reads like a person wrote it and, if you've supplied one, whether it reads like you. That's a narrower claim than it sounds. See /humanize-ai-writing for the mechanics, and /reference/ai-isms-in-marketing-copy for what these tells look like once they've escaped into actual marketing copy.

Worked example

a 300-word AI-written announcement paragraph, the usual suspects present: three "not X but Y" constructions doing the heavy lifting, tells running at roughly 6 per 1,000 words, not one sentence under eight words anywhere in the block, formal throughout with no contractions.

after revision, the three "not X but Y" constructions are gone, replaced with plain statements. Tell density drops from 6 per 1,000 to under 1. Four sentences now sit under eight words, breaking up what was a wall of uniform 25-word sentences. Contractions appear where they'd naturally fall in speech. These numbers are illustrative, meant to show the kind of change humanising produces, not a guarantee for any given paragraph.

the words carrying the actual announcement barely changed. What changed was the scaffolding around them, the rhythm, the hedges, the padding that told a reader (or a detector) that no one was actually talking to them. Fix that, and the content reads like someone wrote it because, by the end, someone did.

Sources

More terms