Humanised AI text examples: 20 scored before-and-after pairs
By Jack Stovell X · Instagram · LinkedIn · YouTube · TikTok · published 2026-11-16 · checked 2026-11-15
These are 20 AI-written passages, each sent once to ScriptGrain's humaniser and scored before and after, from ScriptGrain Research study SGR-027. The detector's human score rose from 0.51 to 0.69 and the content survived, but voice match did not move on average, even though the wording changed a lot.
What the humaniser changed
The passages ran from 146 to 235 words. Ten were written by Claude and ten by GPT, all from bare prompts with no style instruction. We forced the humaniser to rewrite each one, then measured the before and after. Our detector scores text from 0 to 1, where 1 reads as human. The mean went from 0.505 to 0.689, a change of +0.184 (95% interval 0.143 to 0.225). It rose in 18 of 20 pairs. The biggest single gain was +0.4, in P04. Split by generator, Claude passages went from 0.545 to 0.722 and GPT passages from 0.465 to 0.656.
Some surface habits moved. Average sentence length fell from 13.377 to 11.737 words. Commas per word dropped from 0.831 to 0.671. Em dashes went from 0.762 per 1,000 words to none at all. Contractions rose from 30.157 to 35.148 per 1,000 words, but the interval runs from -1.061 to 11.037, so that is not a clear change. Passages grew a little, from 194.1 words to 204 on average.
The facts held. In all 20 pairs the humaniser lost no number, added no number and invented no first-person sentence.
The wording changed far more. On average 66% of the original's content words were still present after the rewrite. The lowest was 41% (P02). So the humaniser replaced a lot of vocabulary while keeping the claims intact. That's a different thing from a light edit, and it's worth knowing before you run your own text through one.
Where it did not help
Start with the two plain failures. In P01 and P11 the humaniser returned the text word for word unchanged, which is why those two pairs show 100% of content words kept and no gain. For P01 none of its 3 passes scored above the original on its detector; for P11 its integrity guard discarded all 3. Across the run it made 60 passes and its own integrity guard discarded 5 of them.
No rewrite reached the target. The default target is 0.85, and the highest score any rewrite reached was 0.82. So 0 of 20 got there. The mean of 0.689 is a long way short of it.
The AI tells did not clearly fall. These are the cliche phrases our checker counts per 1,000 words. The mean went from 0.235 to 0.47, a change of +0.235, with an interval from 0 to 0.705. That isn't a clear fall, and there was little to fix anyway: only 1 of 20 originals had any tell to remove. In one pair the measure rose. P18 went from 0 to 4.7.
Voice match did not move on average. We measure it against one fixed profile, not the writer's own. The mean was 0.74 before and 0.739 after, a change of -0.001, inside the pre-registered margin of plus or minus 0.05. Averages hide the movement, though. In 4 pairs it fell by 0.05 or more: P06 (0.79 to 0.67), P08 (0.84 to 0.75), P15 (0.67 to 0.6) and P20 (0.7 to 0.63). In 5 it rose by the same amount: P13 (0.66 to 0.71), P14 (0.67 to 0.76), P16 (0.76 to 0.82), P18 (0.6 to 0.68) and P19 (0.73 to 0.84). So it moved both ways, and the mean says nothing about which way your own text would go.
Then the caveat that matters most. The detector is ours, and the humaniser is ours too. It keeps whichever of its passes scores highest on that same detector, so a rise on that detector is partly built in. The result doesn't show the text reads as more human to people or to other detectors. We tested one humaniser, one detector, one run and 20 short passages, with no human writing and no other detectors.
How to read the scores
The pairs table has four measures, and each one is narrower than it sounds.
The human score is the detector's number from 0 to 1. A higher score means the model reads the text as more human. It's a model's judgement, not proof of authorship, and a passage scoring 0.69 hasn't been shown to be written by a person. If you want to see how a score like this is produced, you can run your own text through the AI detector, and the page on how AI detectors work explains why these numbers are fuzzy.
Tells per 1,000 words counts the stock phrases that mark text as machine-written, using the cliche checker. It's a count of known phrases, so a low number means those phrases are absent, not that the text sounds like anyone in particular.
Voice match compares each passage against one fixed profile. It isn't the profile of whoever wrote the prompt. So a rise or fall tells you the text moved towards or away from that single reference, nothing more.
Content words kept is the share of the original's content words still present after the rewrite. A low figure means heavy rewording. It doesn't tell you whether the meaning survived; the checks for lost numbers and invented first-person sentences do that job.
For a wider look at humanisers, the reference page on AI humanizers compares eight tools.
All 20 pairs, scored
Every pair from SGR-027, in the order of the study. Scores are before the humaniser to after. Detector human score is 0 to 1 with 1 reading as human; tells are per 1,000 words; voice match is against one fixed profile; content words kept is the share of the original's content words still in the rewrite. In P01 and P11 the humaniser returned the text unchanged. The mean human score was 0.505 before and 0.689 after.
| Pair | Format | Written by | Detector human score, before to after | AI tells per 1,000 words | Voice match | Content words kept |
|---|---|---|---|---|---|---|
| P01 | Blog intro | Claude | 0.78 to 0.78 | 4.7 to 4.7 | 0.78 to 0.78 | 100% (returned unchanged) |
| P02 | LinkedIn post | GPT | 0.55 to 0.72 | 0 to 0 | 0.76 to 0.76 | 41% |
| P03 | Launch email | Claude | 0.45 to 0.72 | 0 to 0 | 0.74 to 0.7 | 58% |
| P04 | Explainer | GPT | 0.22 to 0.62 | 0 to 0 | 0.71 to 0.71 | 44% |
| P05 | About page | Claude | 0.42 to 0.62 | 0 to 0 | 0.75 to 0.73 | 71% |
| P06 | Blog intro | GPT | 0.62 to 0.72 | 0 to 0 | 0.79 to 0.67 | 63% |
| P07 | LinkedIn post | Claude | 0.42 to 0.62 | 0 to 0 | 0.79 to 0.75 | 74% |
| P08 | Launch email | GPT | 0.22 to 0.42 | 0 to 0 | 0.84 to 0.75 | 68% |
| P09 | Explainer | Claude | 0.42 to 0.62 | 0 to 0 | 0.76 to 0.73 | 55% |
| P10 | About page | GPT | 0.42 to 0.62 | 0 to 0 | 0.75 to 0.79 | 54% |
| P11 | Blog intro | Claude | 0.82 to 0.82 | 0 to 0 | 0.86 to 0.86 | 100% (returned unchanged) |
| P12 | LinkedIn post | GPT | 0.62 to 0.72 | 0 to 0 | 0.75 to 0.77 | 52% |
| P13 | Launch email | Claude | 0.55 to 0.82 | 0 to 0 | 0.66 to 0.71 | 76% |
| P14 | Explainer | GPT | 0.35 to 0.62 | 0 to 0 | 0.67 to 0.76 | 69% |
| P15 | About page | Claude | 0.55 to 0.82 | 0 to 0 | 0.67 to 0.6 | 65% |
| P16 | Blog intro | GPT | 0.55 to 0.62 | 0 to 0 | 0.76 to 0.82 | 76% |
| P17 | LinkedIn post | Claude | 0.62 to 0.78 | 0 to 0 | 0.73 to 0.74 | 72% |
| P18 | Launch email | GPT | 0.55 to 0.78 | 0 to 4.7 | 0.6 to 0.68 | 60% |
| P19 | Explainer | Claude | 0.42 to 0.62 | 0 to 0 | 0.73 to 0.84 | 71% |
| P20 | About page | GPT | 0.55 to 0.72 | 0 to 0 | 0.7 to 0.63 | 54% |
Three pairs in the text
These are three of the 20, picked to show the range: the biggest gain, a pair where voice match rose, and a pair where it fell. The full text of every pair is in the JSON download listed under Sources.
P04, original (detector human score 0.22, voice match 0.71):
Loft insulation lowers heating bills by slowing the movement of heat from your home into the colder air outside. Warm air rises, making the roof an important route for heat loss. Without enough insulation, heat passes through the upstairs ceilings and escapes through the loft and roof. Your heating system then has to replace it to keep rooms comfortable.
Insulation materials, such as mineral wool, trap air in small pockets. This trapped air is a poor conductor of heat, so it creates a barrier that slows heat transfer. In a typical unheated loft, insulation is laid between and over the ceiling joists, helping retain warmth in the rooms below.
Because an insulated home loses heat more slowly, the heating needs to run less often or for shorter periods to maintain the same temperature. That reduces the amount of gas, electricity or other fuel you use, cutting your bills.
The savings depend on how much insulation you already have, the size of your home, your heating system and your usual thermostat setting. Installation matters too: gaps and compressed material reduce performance, while suitable ventilation helps prevent moisture problems. Loft insulation works quietly all year and can also help limit heat entering rooms during hot weather.
P04, humanised (detector human score 0.62, voice match 0.71):
Loft insulation cuts your heating bills because it slows how fast heat leaves the house. Warm air rises, so the roof is where a lot of it escapes. Skimp on insulation up there and the heat just wanders through the upstairs ceilings, into the loft and out. Your heating then has to keep topping up the rooms.
Mineral wool and similar materials trap air in tiny pockets, and still air conducts heat badly. That's the whole trick. In an unheated loft the stuff goes between the ceiling joists and over them too, so the rooms below hold onto more warmth.
Lose heat slowly and the boiler doesn't need to run as often or for as long to keep the place at the same temperature. Less gas, less electricity, less of whatever you burn. Smaller bills.
How much you actually save is hard to pin down. It turns on what's already up there, the size of the house, the type of heating, and where you tend to leave the thermostat. Fitting counts as well. Gaps and squashed material do a poor job, and without decent ventilation you can get damp. Once it's in, the insulation works all year without anyone touching it, and in a heatwave it can also hold back some of the heat trying to get into the rooms.
P04 had the biggest detector gain in the set (+0.4) and kept 44% of its content words: the facts are the same, the sentences are not.
P19, original (detector human score 0.42, voice match 0.73), first 3 paragraphs:
Sourdough takes longer because it rises using wild yeast and bacteria rather than the fast-acting commercial yeast in ordinary bread.
Commercial yeast is a single, concentrated strain bred for speed. It produces gas quickly, so a standard loaf can be mixed, risen and baked in a few hours.
A sourdough starter is different. It's a living culture of wild yeast and lactic acid bacteria, kept alive on flour and water. The yeast population is much smaller and more varied, so it produces carbon dioxide far more slowly. The dough needs many hours to rise, and the first rise often runs eight to twelve hours or longer.
P19, humanised (detector human score 0.62, voice match 0.84), first 3 paragraphs:
Sourdough takes longer because it rises on wild yeast and bacteria instead of the fast commercial yeast in ordinary bread.
Commercial yeast is one concentrated strain, bred for speed. Gas comes fast. A standard loaf can be mixed, risen and baked in a few hours, start to finish.
A sourdough starter is a different animal. It's a living culture of wild yeast and lactic acid bacteria, kept alive on flour and water. Way less yeast in there, and a messier mix of types, so the carbon dioxide trickles out. The first rise often runs eight to twelve hours, sometimes longer than that.
P19 gained +0.2 on the detector and its voice match rose from 0.73 to 0.84.
P08, original (detector human score 0.22, voice match 0.84), first 4 paragraphs:
Subject: Paid weekly? Meet a budget that keeps up.
Hi there,
Most budgeting apps think in months. Your paycheck arrives every week.
Meet Weekwise, a budgeting app built around the way you get paid. It helps you plan each paycheck, set money aside for upcoming bills, and see what's available to spend before your next payday.
P08, humanised (detector human score 0.42, voice match 0.75), first 4 paragraphs:
Subject: Paid weekly? Here's a budget that keeps up.
Hi there,
Most budgeting apps think in months. Your paycheck doesn't. It lands every week.
Weekwise is a budgeting app built around how you actually get paid. You plan each paycheck, put money aside for upcoming bills, and see what's free to spend before the next payday rolls around.
P08 gained +0.2 on the detector but its voice match fell from 0.84 to 0.75. The largest voice drop in the set was P06 (0.79 to 0.67). Two pairs, P01 and P11, came back unchanged and are not shown.
Limits
This page shows one humaniser, ScriptGrain's own, scored by one detector, also ScriptGrain's own, on 20 short passages written by two AI models from bare prompts. We sell the humaniser. The humaniser picks between its passes using the same detector that scored the result, so the rise in the human score shows what it optimises for. It does not show that the text reads as more human to people or to any other detector.
Voice match is measured against one fixed profile, not the writer's own, and a different profile could give a different result. The passages are not human writing, each was rewritten once, and model outputs vary between runs. None of this shows how the humaniser behaves on long documents or on your own writing.
Questions
Does humanising change the facts?
Not in this study. All 20 pairs lost no number, added no number and invented no first-person sentence. The wording changed a lot, with a mean of 66% of content words kept and a low of 41%. Facts held, but check your own text, because we only tested 20 short passages from one tool.
Does it make text sound like you?
No evidence of that here. Voice match was measured against one fixed profile, not the writer's own, and the mean barely moved: 0.74 before, 0.739 after. It fell in 4 pairs and rose in 5. Whether a rewrite sounds like you depends on your own samples, which this study did not use.
Does it beat detectors?
It raised our own detector's score from 0.505 to 0.689, but the humaniser keeps the pass that scores highest on that same detector, so some of the rise is built in. No rewrite reached the 0.85 target. We didn't test any other detector, so this study can't say how others would score the same text.
Where is the data?
All 20 pairs are in the table further down this page, three of them as text excerpts, and the full texts are in the JSON download. The study itself is SGR-027, "What does a humaniser actually change?", under Sources. It was run once on 2026-11-10 using 20 passages of 146 to 235 words. The study page lists the method, the limits and every data file.
Methodology
Every figure comes from the published data of ScriptGrain Research study SGR-027 (pairs JSON and CSV, and its summary JSON), run once on 10 November 2026. Nothing was re-run for this page.
Each of 20 passages was written by an AI model from a bare prompt, sent once to the humaniser with a forced rewrite, and measured before and after. The study's method and pre-registration are on the study page.
Sources
- ScriptGrain Research SGR-027: what does a humaniser actually change?
- SGR-027 data: all 20 pairs, original and humanised text, every measure (JSON)
- SGR-027 data: all 20 pairs, every measure, no text (CSV)
- SGR-027 data: summary, hypotheses and intervals (JSON)