Research · Facts and data sheets
Peer-reviewed numbers on how machine text differs from human text, how detectors read it, and where they get it wrong. Every figure carries its source.
§ 01
Key findings
Published results on how AI text differs from human text, and on how well detectors tell them apart. Each card names its source; the full list is at the end of the page.
0.0macro-F1
Structure alone, with no word choice at all, separates AI from human B2B blog posts. After the AI rewords its own posts, it still scores 96.1.
Macro-F1, 0–100. Rewording barely moves it.
[1] SlopShape v3, arXiv 2609.15369, Sept 2026
0.0%
Readers who use AI heavily caught AI articles 92.7% of the time, with 4.0% false positives. Everyone else was at chance: 56.7% caught, 51.7% false positives.
Heavy AI users
Everyone else
[3] Russell, Karpinska & Iyyer, ACL 2025
0%
of essays by non-native English writers were flagged as AI by perplexity-based detectors.
[4] Liang et al., 2023
0.0%0.0%
DetectGPT's detection rate at 1% false positives, before and after one paraphrase pass.
[5] Krishna et al., NeurIPS 2023
0.0%0.0%
Detection of a text watermark (SynthID) before and after one paraphrase pass.
[6] DAMAGE authors' test
43–0%
of human-written texts were flagged by leading detectors after “extremely minor” AI polishing.
[7] Saha & Feizi, APT-Eval 2025
0%0%
A leading detector's flag rate, before and after per-author fine-tuning.
[11] Chakrabarty et al., 2025
0.0×
GPT-4o uses trailing “-ing” analysis clauses (“…, highlighting the need for…”) at 5.3 times the human rate.
[8] Reinhart et al., PNAS 2025
0×
“Delves” appeared at 28 times its expected rate in 2024 PubMed abstracts.
[9] Kobak et al., Science Advances 2025
~0%
fewer views for posts flagged with LinkedIn's “seems like AI slop” report, according to LinkedIn. The report launched 30 Jul 2026.
[10] LinkedIn, via TechCrunch
2 Aug 2026
EU AI Act Article 50 transparency duties for AI-generated content start to apply.
[12] EU AI Act, Article 50
“14 days” → “14 weeks”
A humanizer we tested quietly changed a duration in its output. Rewording is cheap; changing a fact is expensive. That's why every rewrite here is checked against its source.
Source
Returns are accepted within 14 days of delivery.
Humanizer output
We accept returns within 14 weeks of delivery.
Fact changed · a fact check would block this
[13] Our lab test, Oct 2026
§ 02
How AI detectors work
Every detector on the market leans on one or more of these families. Each measures something real, and each has a documented way to fail.
Fig. 2.1 · Statistical
Schematic, not to scale
What it measures
How predictable each word is to a language model. When nearly every next word is the obvious one, the text reads as machine-made.
We treat predictability as one weak signal, never as a verdict on its own.
§ 03
AI vs human
The clearest gap isn't vocabulary. It's how the argument is put together, and whether a reader is left with anything to do.
Table 1 · Structure
No way for the reader to act
2.6×
0.0×
Thesis announced up front
1.8×
0.0×
Summary section
3.3×
0.0×
Close restates the thesis
6.4×
0.0×
Old-versus-new framing
2.9×
0.0×
[2] Sitefire / SlopShape
Table 2a · Grammar
0.0×
GPT-4o, relative to the human rate.
Revenue rose 12% in Q3, highlighting the strength of the new strategy.Example sentence, for illustration
[8] Reinhart et al., PNAS 2025
Table 2b · Vocabulary
0×
2024 PubMed abstracts, relative to the expected rate.
This review delves into the mechanisms underlying…Example sentence, for illustration
[9] Kobak et al., Science Advances 2025
Table 3 · Rhythm, range and specifics
Sentence-length spread
Default model prose keeps sentences in a narrow band. People swing from four words to forty.
Emotional range
Default model prose stays level and upbeat. People get annoyed, wry, unsure, then certain.
Specifics
Default model prose generalises. People name the client, the date, the number, the street.
AI
Human
§ 04
How we benchmark
A detector that is right 95% of the time can still be wrong about most of the people it flags. Our method is built around false positives.
Writing we never train on: native and non-native authors, lightly edited drafts, technical and casual registers. It's where false positives hide.
Score new text against the same writer's known samples, not against an imagined average human. A plain writer is allowed to stay plain.
Every rewrite is diffed against its source for numbers, units, names and links. Any change fails the call instead of shipping.
False-positive rates per group at fixed thresholds, published next to detection rates. Never a single accuracy number.
Fig. 4 · Detect benchmark
1Corpus make-up
Which human and AI texts are in the held-out set, and how they were kept out of training.
2False positives by group
Native and non-native writers, lightly edited drafts and technical writing, each at fixed thresholds.
3Generic vs calibrated
The same texts scored with and without the writer's own baseline, side by side.
4One paraphrase pass
How much detection drops after a single rewording, the test word-level detectors fail.
5Fidelity error rate
How often a rewrite changes a number, unit, name or link, and how often the check catches it.
We'll publish the numbers here, good or bad, with the corpus description and thresholds. Until then, the figures on this page are from the cited studies.
§ 05
Data sheets
Short PDFs that summarise each topic with full citations. In production now.
DS-01 · PDF
Statistical, classifier, watermark and structure detectors: what each measures and the cited failure rates.
DS-02 · PDF
The five patterns that separate AI and human posts, with rates and a reviewer's checklist.
DS-03 · PDF
Why numbers, units, names and links must survive a rewrite, and how a fact check fails closed.
DS-04 · PDF
Held-out corpus, writer calibration and per-group false-positive reporting. Draft.
§ 06
Responsible use
We sell to publishers, content teams and people writing in their own voice. Some uses are off the table, whatever plan you're on.
What IntactVoice is for
What it is not for
Detect returns likelihoods and the reasons behind them, sentence by sentence. A person makes the call.
A voice profile is built from a writer's own samples, with their agreement. Profiles can be deleted at any time.
Rewrite will not change a number, unit, name or link. If it would, the call fails and shows you the diff.
Use outside these lines breaks our terms and ends access. Questions about a use case? Write to support@intactvoice.com.
§ 07
Figures are quoted as published. Where a source reports a range or a platform's own claim, we say so.
SlopShape v3, arXiv 2609.15369, Sept 2026
Structure-only classifier on B2B blog posts; macro-F1 before and after the model rewords its own posts.
Sitefire / SlopShape
Share of AI and human posts showing each structural pattern.
Russell, Karpinska & Iyyer, ACL 2025
Human readers as detectors; heavy AI users versus everyone else.
Liang et al., 2023
Perplexity-based detectors on essays by non-native English writers.
Krishna et al., NeurIPS 2023
Effect of one paraphrase pass on DetectGPT at a 1% false-positive rate.
DAMAGE authors' test
Effect of one paraphrase pass on a text watermark (SynthID).
Saha & Feizi, APT-Eval 2025
Human texts with “extremely minor” AI polishing, scored by leading detectors.
Reinhart et al., PNAS 2025
Grammatical and rhetorical features of model text versus human text.
Kobak et al., Science Advances 2025
Excess vocabulary in 2024 PubMed abstracts.
LinkedIn, via TechCrunch
“Seems like AI slop” report, launched 30 Jul 2026; reach figure as stated by LinkedIn.
Chakrabarty et al., 2025
Per-author fine-tuning and a leading detector's flag rate.
EU AI Act, Article 50
Transparency obligations for AI-generated content apply from 2 Aug 2026.
Our lab test, Oct 2026
A third-party humanizer changed a duration in its output. Internal test, not peer reviewed.
Detect API
Run Detect on your own writing and get sentence-level signals with the pattern behind each one.