This measures style, not authorship. Humans write slop and models can avoid it. Do not use it to make decisions about people (grading, hiring, misconduct).

← Back to the detector

The rule set

91 features: the 87 in the scored model and 4 artefact-variant features that are defined but not scored. Each feature is code written by hand (regexes, word lists, counts); a logistic regression learned on the v2 training split weights them. A positive weight means a higher value pushes the score toward AI style. Features are sorted by |weight| within each family. Model slop_lr, export sha256 c32b8dc271b5, built from commit ac9dd98.

The direction observed on our training data often disagrees with the folk tell: those rows say "reversed" in their rationale.

constructions (4)

'It isn't X. It's Y.' constructions.not_x_its_y_xsent

Antithesis across two sentences (ports of Paech's slop-score patterns). The single-sentence 'not X but Y' is null on v2 train; this form survives the antitell prompt's ban: AUC .57 on antitell (13% of AI vs 4% of human texts overall), null on natural.

Source
PAECH-SS; WP; SLHCK; GRAPHITE
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.159

'rather than' constructions.rather_than

'rather than' per 1000 words, a contrastive framing models favour. AI more: AUC .56 overall (fires in 16% of AI vs 4% of human texts), .59 natural.

Source
GRAPHITE; WP
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.142

Formula phrases ('when it comes to', 'plays a role') constructions.formula_phrases

Russell et al.'s formula phrases plus 'plays a (X) role' and 'a testament to'. AI more on natural (AUC .58), null overall.

Source
Russell2025experts; Kobak2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.052

Rule of three ('fast, cheap, and reliable') constructions.rule_of_three

Lists of exactly three short items ('X, Y, and Z'), not part of a longer list. AI reaches for triads: AUC .58 overall, .65 natural, null on antitell.

Source
PANGRAM-9; WP; A16Z
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.041

dramatic beats (3)

One-line paragraphs dramatic_beats.one_line_para

Share of paragraphs (after the first; headings and list items left out) that are a single sentence of <= 12 words. Predicted as an AI emphasis move; reversed on v2 train: humans write more of them, AUC .63 overall (human > AI), on every slice; .64 without the flagged Wikipedia slice.

Source
SPEC; A16Z
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
share of paragraphs
Highlights
units: the one-line paragraphs
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.213

Ending restates the opening dramatic_beats.restatement_ending

Overlap (Jaccard of content-word stems) between the first and the last paragraph. The essay habit of closing on the opening's words: AI higher, AUC .63 overall, .57 natural, .68 antitell. Texts with >= 3 paragraphs.

Source
PANGRAM-G; A16Z
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
Jaccard
Highlights
hits: shared words in the first and last paragraph
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.175

Short sentence after a long one dramatic_beats.short_after_long

Sentences of <= 6 words right after one of >= 25 words, per 100 transitions. Predicted AI; reversed on the natural slices (AUC .60, human > AI), null overall.

Source
SPEC; A16Z; MunozOrtiz2024 (contra)
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 100 transitions
Highlights
hits: the short sentences
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.037

global statistics (5)

Adjacent sentences share words global_statistics.sentence_overlap

Mean Jaccard overlap of content words between neighbouring sentences. Predicted higher in AI; reversed on v2 train: humans carry words from one sentence to the next more (topic chaining), AUC .73 overall (human > AI), .75 natural, .73 antitell.

Source
FrohlingZubiaga2021
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
Jaccard
Highlights
hits: content words repeated from the previous sentence
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.351

Punctuation variety global_statistics.punct_diversity

Entropy of the punctuation profile over 13 marks (comma, period, colon, semicolon, ! ? dashes, hyphen, parentheses, quotes, ellipsis). Humans use a wider mix of marks; AI leans on commas and periods. On v2 train AI is lower: AUC .65 overall, .75 antitell, null on natural.

Source
WP; ECON; GRAPHITE; Przystalski2025
Evidence type
community
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
bits
Highlights
units: every punctuation mark counted
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.309

Repeated 4-word sequences global_statistics.ngram_repetition

Share of word 4-grams that occur twice or more in the text. Predicted higher in AI (it loops); reversed on v2 train: humans repeat phrases more, AUC .71 overall (human > AI), .73 natural, .69 antitell.

Source
PAECH-SS; PAECH-SF; PANGRAM-G; ShaibDiversity2025; FrohlingZubiaga2021
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
share of 4-grams
Highlights
hits: repeated 4-grams
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.191

Top-10 content words' share global_statistics.topk_word_share

Share of content-word tokens taken by the text's 10 most frequent content words. Predicted higher in AI; reversed on v2 train: humans concentrate on fewer words, AUC .66 overall (human > AI), .69 natural, .65 antitell.

Source
SPEC; FrohlingZubiaga2021
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
share of content words
Highlights
hits: occurrences of the top-10 content words
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.088

Repeated function-word frames global_statistics.template_rate

Content words replaced by X; share of 5-grams of the skeleton ('the X of the X') that occur twice or more. AI reuses syntactic templates even while it varies its words: AI higher on v2 train, AUC .63 overall, .71 natural, .59 antitell.

Source
ShaibTemplates2024
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
share of 5-grams
Highlights
hits: word runs whose skeleton repeats
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.076

grammar (18)

Prepositions grammar.prepositions

Biber's preposition list per 1000 words. Predicted higher in AI; reversed on v2 train: humans use more, AUC .60 overall (human > AI), .69 antitell, null on natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.456

Clause-joining 'and' (', and I ...') grammar.clausal_coordination

'and' starting a new clause with a pronoun subject. Sources disagree; on v2 train humans chain clauses more on the natural slices (AUC .64, human > AI), null elsewhere.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.314

'because' grammar.because

Predicted human; reversed on v2 train: AI uses 'because' more, AUC .56 overall (AI > human), .60 antitell (the ELI5 explanations).

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.256

Agentless passive grammar.agentless_passive

be/get + participle without 'by'. Contrary to the folk 'AI is passive', humans use more: AUC .57 overall (AI fewer), .59 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.185

Post-nominal -ing ('the man standing there') grammar.present_participle_whiz

Noun-like word followed by an -ing word. AI more: AUC .58 overall, .65 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.172

Participial clauses grammar.present_participle

Clause-initial '-ing' words and ', V-ing the/a/to ...' tails ('..., revealing the cracks'). AI uses more: AUC .57 overall, .68 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.102

Second-person pronouns grammar.second_person

you, your ... per 1000 words. Sources disagree; on v2 train humans address the reader more: AUC .58 overall (human > AI), .63 natural.

Source
GRAPHITE; Reinhart2025
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
-0.091

'not' and n't grammar.negation_analytic

Analytic negation per 1000 words. Humans more on the natural slices (AUC .59), null overall.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.088

Phrasal 'and' (N and N) grammar.phrasal_coordination

'and' joining words of the same kind ('innovation and collaboration') and 'a, b, and c' lists. AI more: AUC .56 overall, .62 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.084

Post-nominal -ed ('the data collected in 2020') grammar.past_participle_whiz

Noun-like word + -ed + preposition. AI more: AUC .55 overall, .63 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.082

Existential 'there is/are' grammar.existential_there

'there is', 'there are', 'there's' ... Humans more: AUC .57 overall, natural and antitell.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.072

Private verbs (think, feel, know ...) grammar.private_verbs

Verbs of thinking and feeling. Humans more on the natural slices (AUC .55), null overall.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.052

Demonstrative pronouns ('this is', 'that was') grammar.demonstrative_pronoun

this/that/these/those standing alone before a verb or punctuation. Humans point back more: AUC .58 overall (AI fewer), .59 natural and antitell.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.043

First-person pronouns grammar.first_person

I, me, my, we, us, our ... per 1000 words. Human writing is personal; AI keeps an impersonal voice even when asked to tell a story: AUC .62 overall (AI fewer), .67 natural, .58 antitell (rate-matched by the antitell prompt, still lower). Kept next to mean_word_length although the two correlate (rho -.75): stance, not vocabulary.

Source
MILD-REPO; GRAPHITE; Reinhart2025; JiangHyland2024
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.035

Contractions grammar.contractions

n't, 're, 'll, 've, 'd, 'm and it's/that's/there's ... per 1000 words. AI writes out the full forms more: AUC .58 overall (AI fewer), .63 natural.

Source
PANGRAM-G; GRAPHITE; WAPO (contra); Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.033

'be' as main verb grammar.be_main_verb

Forms of 'be' not followed by a participle or 'to' ('He is tall', 'it's fine'). Plain copular clauses are a human, spoken feature: AUC .59 overall (AI fewer), .68 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.021

'do' as pro-verb ('so did I') grammar.pro_verb_do

do/does/did standing in for a verb phrase. Humans more: AUC .55 overall, .58 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.013

'no', 'neither', 'nor' grammar.negation_synthetic

Synthetic negation. Humans more on antitell (AUC .55), null elsewhere.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.002

hedging filler (8)

'will', 'would', 'shall' hedging_filler.modals_predictive

Predictive modals per 1000 words. Humans speculate and plan in will/would; AI uses fewer: AUC .65 overall, .64 natural, .66 antitell.

Source
Herbold2023; Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.325

Amplifiers (very, completely, extremely) hedging_filler.amplifiers

Biber's amplifiers. Humans intensify more: AUC .60 overall (AI fewer), .65 antitell, null on natural.

Source
Reinhart2025; JiangHyland2025meta
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.280

Sentence-initial 'Well', 'So', 'Now' hedging_filler.discourse_particles

Sentences opening with a spoken particle (well, now, anyway, oh, so). Humans more: AUC .59 overall, natural and antitell.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.186

Hedges (maybe, sort of, kind of, almost) hedging_filler.biber_hedges

Biber's hedges. Null overall and on natural; on antitell AI uses more (AUC .56, reversed), probably because the antitell prompt asked for a casual human voice.

Source
Reinhart2025; JiangHyland2025meta
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.151

Plain wordy phrases ('in order to', 'the fact that') hedging_filler.wp_human_wordiness

Wordy constructions and simple superlatives Wikipedia lists as signs of human writing. Humans more: AUC .57 overall, .59 antitell, null natural.

Source
WP
Evidence type
community
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.125

First-person hedges ('I think', 'pretty sure', 'IIRC') hedging_filler.human_hedges

Epistemic hedges in the first person. Humans more: AUC .56 overall, .57 natural, .55 antitell (rate-matched by the antitell prompt).

Source
MILD-REPO; GRAPHITE; Herbold2023; JiangHyland2024
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.120

'can', 'could', 'may', 'might' hedging_filler.modals_possibility

Possibility modals. Sources disagree; humans slightly more on natural (AUC .55), null elsewhere.

Source
Herbold2023; Reinhart2025
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.086

Emphatics (really, just, a lot, so + adjective) hedging_filler.emphatics

Conversational emphasis. Humans more: AUC .56 overall, .61 natural.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.016

lexical (19)

Fiction slop words (gaze, whispered ...) lexical.slopscore_fiction_words

Hits (words and phrases) of the list per 1000 words. Words over-represented in model fiction (antislop top 200). One of three lists that survive the antitell prompt: AUC .61 overall, .68 natural (WritingPrompts .88), .60 antitell.

Source
PAECH-AS; PAECH-SF; PAECH-SS
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.373

Conjuncts (however, therefore, thus ...) lexical.connectives_biber

Biber's conjunct list. Sources disagree; on v2 train humans use more: AUC .61 overall (human > AI), .70 antitell, null natural. The antitell prompt banned formal transitions, so this is partly a prompt effect.

Source
Reinhart2025; Herbold2023
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.290

Mid-2026 AI vocabulary (challenge, foster, quietly ...) lexical.wp_era_2026

Hits (words and phrases) of the list per 1000 words. Wikipedia's weakly sourced list for 2026 models (challenge, clearer, dependable, foster, leverage, matters, practical, quietly, steady ...). Survives the antitell prompt: AUC .61 overall, .64 natural, .59 antitell.

Source
WP; GRAPHITE
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.225

Claude-style words (deliberate, measured, steady) lexical.graphite_claude5

Phrases Graphite attributes to Claude Opus 4.x/5, plus vague-concrete numbers ('a handful', 'two or three'). Survives the antitell prompt: AUC .58 overall, .60 antitell.

Source
GRAPHITE
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.224

Noun-preposition bundles ('the importance of') lexical.np_pp_bundles

'the X-tion of', 'in the X of' ... Predicted higher in AI; reversed on antitell (AUC .61, human > AI), null on natural.

Source
JiangHyland2024
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.212

Swearing and internet slang lexical.profanity_slang

damn, hell, crap, lol, tbh, gonna, kinda ... Humans more: AUC .58 overall (AI fewer), .61 antitell. Only a positive count is evidence.

Source
Reinhart2025; Guo2023HC3; MILD-REPO
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.174

Banned words of the antitell prompt lexical.mildrgb_banned

delve, crucial, vital, pivotal, robust, leverage, navigate, tapestry ... (the list the antitell generation prompt banned). AI more on natural (AUC .57), null overall.

Source
MILD-REPO
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.109

Excess words (Kobak et al., content/style) lexical.kobak_mixed

Hits (words and phrases) of the list per 1000 words. AI more on natural (AUC .56), null overall.

Source
Kobak2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
-0.108

2023-24 (GPT-4) vocabulary lexical.wp_era_gpt4

Hits (words and phrases) of the list per 1000 words. Wikipedia's GPT-4-era list. AI more on natural (AUC .58).

Source
WP
Evidence type
community
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.091

Evaluative adjectives (Liang et al.) lexical.liang_review_adj

Hits (words and phrases) of the list per 1000 words. Adjectives over-used in AI-written peer reviews (commendable, innovative ...). AI more: AUC .56 overall, .63 natural.

Source
Liang2024rev
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.076

Essay slop words (slop-score) lexical.slopscore_essay_words

Hits (words and phrases) of the list per 1000 words. Words over-represented in model essays (Paech's slop-score). AI more: AUC .61 overall, .73 natural, null on antitell.

Source
PAECH-SS
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.066

'core', 'modern' lexical.wapo_core_modern

GPT-4o favourites per the Washington Post analysis. AI more on natural (AUC .55), null overall.

Source
WAPO
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.064

Instruct-model words (camaraderie, palpable ...) lexical.reinhart_overrep

Words instruction-tuned models over-use (Reinhart et al.). AI more on natural (AUC .60), null overall.

Source
Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.063

Adverbs (Liang et al.) lexical.liang_review_adv

Hits (words and phrases) of the list per 1000 words. AI more on natural (AUC .58), null overall.

Source
Liang2024rev
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
-0.062

slop-forensics words lexical.slopforensics_words

Hits (words and phrases) of the list per 1000 words. Top words of model output over varied prompts. AI more on natural (AUC .64), null overall.

Source
PAECH-SF
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
-0.026

Human-voice markers (pretty, I guess, years ago) lexical.human_informal

Informal markers from Graphite and mild-rgb. Humans more: AUC .55 overall, .56 natural and antitell.

Source
GRAPHITE; MILD-REPO
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.025

2024-25 (GPT-4o) vocabulary lexical.wp_era_gpt4o

Hits (words and phrases) of the list per 1000 words. Wikipedia's GPT-4o-era list. AI more on natural (AUC .55).

Source
WP
Evidence type
community
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
+0.015

Wikipedia's AI-vocabulary box lexical.wp_aivocab

additionally, bolster, crucial, delve, enhance, foster, intricate, pivotal, showcase, testament ... per 1000 words. AI more on natural (AUC .59), null overall.

Source
WP; Kobak2025; JuzekWard2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.011

AI vocabulary (Pangram's list) lexical.pangram_guide_list

Hits (words and phrases) of the list per 1000 words. A broad list of words and phrases that AI over-uses, compiled by the detector vendor Pangram; it also holds generic words (health, life, key). AI more on v2 train: AUC .59 overall, .75 natural, null on antitell (banned words).

Source
PANGRAM-G
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
lexicon
Status
main
LR weight (standardised; + = AI-ward)
-0.009

mined (12)

Plain quantifiers (all, any, much, many, lots; some of, most of, one of) mined.human_quantifiers

People quantify plainly: all, any, much, many, lots, and partitives (some of, most of, one of, both of ...). On v2 train humans use more (AI<H): AUC .71 overall, .66 natural, .74 antitell; human-ward on every slice (weakest on Wikipedia, .56).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.617

Explainer vocabulary (specific, rules, gap, shift, decades, ordinary, genuinely ...) mined.explainer_vocab

The 2026 explainer register: abstract framing nouns and adjectives (specific, rule(s), gap(s), shift(s/ed), decades, ordinary, genuinely, final, remained) that none of the community or literature lists has. On v2 train AI uses more: AUC .69 overall, .69 natural, .69 antitell; AI-ward on every slice (.59-.77). Six mined words were dropped at integration as topical or carried by one slice (system(s), central, centuries, ancient; specifically had no signal).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.453

Plain multiword phrases (able to, as well, due to, for example, have to) mined.plain_multiword

People use plain multiword phrasing (able to, as well, due to, for example, have/has/had to, want to, a few, a little). On v2 train humans use more (AI<H): AUC .69 overall, .68 natural, .71 antitell; human-ward on every slice (weakest on Wikipedia, .56). Fairness note: learner and student writing uses these too, which pushes it toward 'human' (the safe direction).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.323

Alternatives with 'or' mined.or_alternatives

AI offers alternatives ('X or Y', 'whether ... or ...'); 'or' was the largest AI coefficient of bag-of-words LR on half A. On v2 train AI uses more: AUC .60 overall, .59 natural, .62 antitell; flat on AITA (.49) and WritingPrompts (.47).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.292

Complement 'that' + pronoun ('that they', 'that we') mined.that_pronoun

People follow complement 'that' with a pronoun subject ('that they', 'that we', 'that there'); AI more often writes a noun subject. On v2 train humans use more (AI<H): AUC .64 overall, .65 natural, .64 antitell.

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.283

Connectives through, toward(s), within, across, whether, itself mined.connective_prepositions

AI links ideas with a small set of spatial-logical connectives (through, toward(s), within, across, whether, itself). Mined subset of Biber's prepositions that goes the other way from the full list (grammar.prepositions is AI<H). On v2 train AI uses more: AUC .67 overall, .70 natural, .66 antitell; AI-ward on every slice (.62-.80).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.245

Lower-case hyphenated compounds mined.hyphen_compounds

AI writes more lower-case hyphenated compounds ('long-term', 'well-known', 'trill-filled'); URLs are skipped. On v2 train AI uses more: AUC .61 overall, .65 natural, .59 antitell; AI-ward on every slice. Replaces punctuation.hyphen (rho +.94, weaker in every group).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.184

Sentences open with an article ('The system ...', 'A small change ...') mined.article_openers

AI opens sentences with a noun phrase ('The system ...'); people more often open with a pronoun, a conjunction or a discourse word. Share of sentences whose first word (after markup and opening quotes) is the/a/an. On v2 train AI is higher: AUC .66 overall, .59 natural, .70 antitell; weakest on Wikipedia (.56), where encyclopedic human prose opens with 'The' too.

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
share of sentences
Highlights
hits: the opening article
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.171

Sentences end in a bare full stop mined.plain_period_endings

AI ends nearly every sentence with a bare full stop; people also end on ?, !, a closing bracket or quote, '...' or a colon. Counted on real sentences only (>= 4 words, not a heading or list line, ending in sentence punctuation), so unpunctuated headings, list items and Reddit lines do not count. On v2 train AI is higher: AUC .69 overall, .66 natural, .73 antitell (the mining definition on all sentences gave .75, partly from headings and list lines on Wikipedia and Stack Overflow).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
share of sentences
Highlights
hits: the full stop of each real sentence that ends in one
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.164

Run-on sentences (more than 40 words) mined.run_on_sentences

People write run-on sentences of more than 40 words; AI keeps sentences in a narrow band. Share of sentences with > 40 words. On v2 train humans are higher (AI<H): AUC .66 overall, .59 natural, .72 antitell. Fairness note: learners and young writers write run-ons too, which pushes them toward 'human' (the safe direction).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
share of sentences
Highlights
hits: the sentences over 40 words
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.112

Plain verbs and evaluations (get, go, want, good, great, ok, sorry) mined.plain_verbs

People use core verbs and plain evaluations (get, go, want, good, great, ok, sorry), where AI picks more specific verbs and adjectives. On v2 train humans use more (AI<H): AUC .61 overall, .69 natural, .56 antitell; strongest on Stack Overflow (.79) and WritingPrompts (.76).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.081

Mid-sentence colon continued in lower case ('the rule is simple: you wait') mined.colon_lowercase

AI uses a mid-sentence colon continued in lower case ('the rule is simple: you wait'); people more often write capitalised label colons ('Note: The ...'). Share of colons followed by a lower-case word, of all colons followed by a letter; undefined without such colons. On v2 train AI is higher: AUC .68 overall, .68 natural, .67 antitell (null on AITA, .47).

Source
first-hand: mined on v2 train (half A), replicated on half B
Evidence type
first-hand
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
share of colons
Highlights
hits: colons followed by a lower-case word
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.075

punctuation (13)

Parentheses punctuation.parentheses

Opening parentheses per 1000 words. Humans add asides in brackets; 2026 models almost never do (fires in 11% of AI vs 65% of human train texts). AUC .78 overall (AI fewer), .66 natural, .87 antitell.

Source
MILD-REPO; ECON; GRAPHITE; PANGRAM-G
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.597

Em dashes punctuation.em_dash

Em dashes (—) per 1000 words, the best-known AI tell. On v2 train it holds on the natural slices (AUC .66; WritingPrompts .69, Wikipedia .65) and is uninformative on antitell, where mild-rgb injected em dashes into the AI texts at the human rate. A real tell, but typographic: the --ascii-norm evaluation turns it into a hyphen.

Source
WP; WAPO; PANGRAM-9; GRAPHITE; ECON; ARS; Freeburg2026; Czuma2026a
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.370

Serial (Oxford) comma punctuation.oxford_comma

Among 'X, Y(,) and Z' lists, the share written with the serial comma. AI uses it consistently, people mostly leave it out: AUC .68 overall, .70 natural, .68 antitell. Undefined in texts without such a list.

Source
PANGRAM-G
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
share of triads
Highlights
units: each triad
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.286

Commas punctuation.comma_rate

Commas per 1000 words. Sources disagree; on v2 train AI uses more: AUC .59 overall, .68 natural, null on antitell.

Source
ECON; WP
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.122

Digits and symbols punctuation.numbers_symbols

Numbers and the symbols % $ € £ @ & / + = < > per 1000 words, outside URLs. Predicted higher in AI; reversed on v2 train: humans give more figures, dates and symbols, AUC .68 overall (human > AI), .63 natural, .72 antitell, human-ward on every slice. The probe was artefact-flagged on Stack Overflow because of URLs (40% of human answers); URLs are now skipped.

Source
MunozOrtiz2024
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.078

Typed dashes (' - ', '--') punctuation.dash_surrogates

A spaced hyphen or double hyphen used as a dash. People type these; models emit a real dash character. AI fewer on v2 train: AUC .56 overall, .63 antitell, null on natural.

Source
WP; Freeburg2026
Evidence type
community
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.076

Emotional marks and ALL CAPS punctuation.irregular_marks

'?!', '!!', '??' runs and ALL-CAPS words that are not acronyms, per 1000 words. Humans shout and exclaim; AI writes evenly. AUC .61 overall (AI fewer), .57 natural, .64 antitell.

Source
Guo2023HC3
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.050

Exclamation marks punctuation.exclamation

'!' per 1000 words. Humans exclaim more: AUC .56 overall, .58 antitell (rate-matched by the antitell prompt, still lower in AI), null natural.

Source
GRAPHITE; MILD-REPO; Guo2023HC3
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.021

Punctuation overall punctuation.punct_rate

All punctuation marks per 1000 words. Slightly lower in AI on antitell (AUC .57), null elsewhere. The probe counted marks per word; here per 1000 words (same ranking).

Source
MunozOrtiz2024; Guo2023HC3
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
+0.003

Curly quotes punctuation.curly_quotes

Any curly quote or apostrophe. AI > human on the natural slices (AUC .70) but that is the platform convention of the human side (SO .06 vs .72, Wikipedia .07 vs .77 of texts), not the writer.

Not in the scored model: platform artefact on v2: Stack Overflow and Wikipedia human text is straight-quoted, rasbt's AI text curly; mild-rgb injected typography into the antitell AI texts at human rates (docs/DATA.md)

Source
WP-S; MILD-AITA; MILD-ELI5
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
AI-ward (AI higher)
Unit
presence (0/1)
Highlights
hits
Kind
presence
Status
variant
LR weight (standardised; + = AI-ward)
–

Curly and straight quotes mixed punctuation.mixed_quote_styles

Curly and straight quotes both in one text. AUC .62 on the only clean slice (WritingPrompts), artefact elsewhere.

Not in the scored model: platform artefact on v2: Stack Overflow and Wikipedia human text is straight-quoted, rasbt's AI text curly; mild-rgb injected typography into the antitell AI texts at human rates (docs/DATA.md)

Source
MILD-REPO; WP-S
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
presence (0/1)
Highlights
units: all quotes
Kind
code
Status
variant
LR weight (standardised; + = AI-ward)
–

Double spaces punctuation.double_space

Two spaces inside a line. Human > AI (AUC .57), all of it from flagged slices.

Not in the scored model: old Reddit typing on WritingPrompts (no AI text has one); injected into antitell AI texts at human rates; null without the flagged slices (.53)

Source
MILD-REPO
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
presence (0/1)
Highlights
hits
Kind
presence
Status
variant
LR weight (standardised; + = AI-ward)
–

Ellipses punctuation.ellipsis

'…' and '...' per 1000 words. Human > AI on antitell (AUC .59).

Not in the scored model: signal only on antitell, where mild-rgb injected ellipses into AI texts; null without the flagged slices (.46)

Source
MILD-REPO; SLHCK; Guo2023HC3
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
variant
LR weight (standardised; + = AI-ward)
–

statistics (7)

Vocabulary variety (MATTR) statistics.mattr

Moving-average type-token ratio over windows of 100 words: how few words repeat. The folk and GPT-2-era claim is that AI repeats itself; on v2 train it is the opposite: 2026 models (and Opus/GPT-4o on WritingPrompts) vary their vocabulary more than humans: AUC .80 overall, .90 natural, .74 antitell. This is also the direction that should not penalise English learners, whose variety is lower.

Source
GRAPHITE; WP; PAECH-SS; Reinhart2025; Herbold2023; MunozOrtiz2024; Suvanto2026
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
AI-ward (AI higher)
Unit
type-token ratio
Highlights
units: words that repeat a word used within the previous 100 words (they lower the value)
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.556

Concrete specifics (numbers, names, months) statistics.specificity

Digits, capitalised words inside a sentence (names) and month names per 1000 words. Sources disagree; on v2 train humans give more specifics: AUC .63 overall (human > AI), .59 natural, .69 antitell. AI text stays generic.

Source
PANGRAM-G; WP; GRAPHITE
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.361

Mean word length statistics.mean_word_length

Mean letters per word: the one feature kept for the formal-register cluster (Latinate words, nominalisations, attributive adjectives; rho .73-.77 with it). AI prose leans on longer, more formal words. On v2 train AI words are longer: AUC .67 overall, .79 natural, .60 antitell.

Source
GRAPHITE; ECON; Reinhart2025
Evidence type
empirical
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
letters per word
Highlights
units: words of 10+ letters (the value uses every word)
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.202

Sentences of 8-20 words statistics.sentence_length_band

Share of sentences of 8 to 20 words, the band 2026 GPT models keep to. AI keeps more sentences in the band: AUC .60 overall, .65 antitell, null on natural.

Source
WP
Evidence type
expert
Direction in the sources
AI-ward (AI higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
share of sentences
Highlights
units: sentences in the band
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.113

Sentence-length variation (burstiness) statistics.burstiness

Coefficient of variation of sentence lengths. Human writers alternate long and short sentences more. On v2 train AI texts are more even (lower CV): AUC .67 overall, .70 natural, .64 antitell, although the antitell prompt asked the models for uneven rhythm.

Source
GRAPHITE; PANGRAM-G; GPTZERO; MILD-REPO; MunozOrtiz2024; Guo2023HC3
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
statistic (CV)
Highlights
units: every sentence (texts with >= 5 sentences)
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.104

'is' / 'are' statistics.is_are_rate

Copula 'is/are' per 1000 words; AI is said to avoid the plain copula ('serves as' instead of 'is'). Holds on v2 train: AI uses fewer, AUC .57 overall, .62 natural, .56 antitell.

Source
WP; GengTrotta (via WP)
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
per 1k words
Highlights
hits
Kind
rate
Status
main
LR weight (standardised; + = AI-ward)
-0.041

-ly adverbs statistics.ly_adverbs

Adverbs ending in -ly (stoplist of non-adverbs like 'family', 'early'). Predicted rarer in AI; reversed on v2 train: AI uses more, AUC .57 overall (AI > human), .70 natural, null on antitell.

Source
Reinhart2025; Liang2024rev
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
AI-ward (AI higher)
Unit
per 1k words
Highlights
hits
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.029

structure (2)

Uniform paragraph lengths structure.uniform_para_length

Coefficient of variation of paragraph word counts. AI paragraphs are cut to a similar size; people write a two-line paragraph next to a long one. The strongest single tell on v2 train: AUC .81 overall (AI lower), .76 natural, .84 antitell; .80 without the flagged Wikipedia slice. Texts with >= 3 paragraphs.

Source
PANGRAM-G; SPEC; MILD-ELI5; MunozOrtiz2024
Evidence type
empirical
Direction in the sources
human-ward (humans higher)
Direction observed on v2 train
human-ward (humans higher)
Unit
statistic (CV)
Highlights
units: every paragraph
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
-0.631

Mean paragraph length structure.para_length

Mean words per paragraph. Sources disagree; on v2 train AI paragraphs are longer: AUC .63 overall, .68 natural, .62 antitell; .62 without Wikipedia.

Source
PAECH-SS
Evidence type
empirical
Direction in the sources
mixed / contested
Direction observed on v2 train
AI-ward (AI higher)
Unit
words per paragraph
Highlights
units: every paragraph
Kind
code
Status
main
LR weight (standardised; + = AI-ward)
+0.136

Sources

Keys not listed here are "expert judgement", "first-hand" (found in our data) or the project spec.