scanforai

Methodology — open, inspectable, imperfect

ScanForAI runs the open-source aiscan engine; everything below is its actual scoring logic.

Every formula below ships in the source. If you can't inspect a detector's method, you can't contest its output; here you can do both.

Tier 1 — stylometric heuristics

Each signal maps a measurement to 0..1 (1 = machine-leaning) and carries a weight (its reliability). The overall score is the weighted average, ×100. Signals abstain (weight 0) when the text is too short to judge.

• burstiness (weight 3.0) — coefficient of variation of sentence lengths. Humans mix short and long sentences (CV ≈ 0.5–0.9); models write uniformly (CV ≈ 0.2–0.4). The strongest pure-text signal.
• banned_phrases (2.5) — density of curated AI-ese phrases ("it's worth noting", "delve", "testament to"…), per 1,000 words.
• contractions (2.0) — contraction rate vs. contractible material; unedited model prose under-contracts.
• antithesis (2.0) — the "not just X, it's Y" cliché structure.
• conjunction_starts (1.5) — humans open sentences with But/And/So; models rarely do.
• typography (1.5) — em-dash flood, curly quotes, ellipsis glyphs.
• ai_words, vocab_diversity, opening_repetition, uniformity (1.0 each) — lexicon hits, rolling type-token ratio, repeated sentence openers, near-equal adjacent sentence lengths.
• clause_train (2.0) — longest run of consecutive 25-word-plus, multi-comma sentences. Complementary to burstiness: a heavy long tail raises the CV (reads "human" there) while the prose still marches in same-shaped comma-trains.
• simile_tail (1.5) — density of the ", the way X ..." comparison frame; model narrative repeats one pet construction.
• long_tail (1.5) — share of sentences over 40 words plus any over 55; edited prose keeps the tail under ~5–8%.
• echo_grams (1.5) — distinct content five-word phrases repeated 3+ times in one text; models re-use their own scaffolding.
• aphorism_density (1.5, genre-gated) — share of short definitional/reversal epigrams ("An empty line is X. An initialed blank is Y.") plus a bonus for 3+ consecutive stacks. Corpus calibration (July 2026) found this separates expository AI (opinion/persuasive prose, ~17% definitional lines) from humans, but not fiction: current AI fiction scores ~0% while aphoristic human literary prose scores 8–12%. So the signal abstains on narrative text (detected by pronoun/past-tense/dialogue density) rather than punish strong literary voice — the classic false positive. It scores only expository prose, where the epigram-stack is a genuine tell.

The four shape signals were derived in July 2026 from a 349k-word fiction corpus that passed every signal above and still tripped perplexity-class detectors — the surviving tells were all sentence-shape tells, not word choice.

Tier 2 — local model perplexity (weight 6.0)

A small open model runs on our own hardware and measures per-token surprisal. Machine text is unusually predictable — a low, flat surprisal stream — which is the GLTR/GPTZero family of evidence. Predictability carries the signal (a light uniformity term rides along). Input is capped at 1,024 tokens. No third-party AI APIs are called, ever.

Genre gating

Most tells are genre-dependent. Corpus calibration (July 2026) showed several signals mean opposite things in different registers: aphorism density flags persuasive AI but is normal literary voice; simile density is higher in human fiction than in current AI fiction; contraction rate is tuned for fiction and mis-reads formal exposition. The scanner therefore classifies each input (fiction / expository / promotional / mixed, from pronoun, tense, dialogue, second-person and imperative density) and lets genre-sensitive signals abstain or re-baseline where they are not diagnostic — rather than punish a strong human voice. The detected genre is returned in every result. Lexicon and typography signals are genre-robust and are never gated.

Design rule: a genre gate may only make a signal quieter where the corpus shows a false positive; it never raises an accusation. Per-genre calibration tables are a work in progress as the harvest corpus grows.

Honest limitations

• Calibration constants (the human/AI breakpoints above) are first-pass guesses, not yet tuned on a labeled corpus. Treat absolute numbers accordingly.
• The phrase lexicon decays: "delve"-era tells fade as models change.
• A 2019 model approximates newer models' predictability imperfectly.
• Editing defeats surface tells — by design this tool tells you how.
• Tells are genre-dependent: aphorism density flags persuasive AI but is normal in literary fiction; the scanner genre-gates where the corpus demands it, and absolute numbers stay first-pass until tuned on a larger labeled set.

Independent evaluations are welcome — the scoring engine is a small, readable Python package with zero dependencies in its heuristic tier.