Glossary

Every metric the tool produces, with a plain-language definition, an example or two, and side-by-side comparisons showing what high and low values actually look like in prose. The eight AI-writing signs are at the bottom, each with a side-by-side showing AI-style phrasing versus the same idea written without the marker.

The field — what stylometry is

Stylometry

The quantitative study of writing style: the measurable patterns in how a writer chooses words, structures sentences, and organizes paragraphs.

The field draws on computational linguistics (turning language into numbers), corpus linguistics (studying language via large bodies of text), forensic linguistics (applying linguistic evidence to real-world attribution problems), and the digital humanities (using computational tools to read literary and cultural texts).

Its methods range from descriptive profiling — measuring features and reporting patterns, which is what this tool does — to formal authorship attribution using cross-validated classifiers on baseline corpora of known-author writing, which this tool deliberately does not do. Descriptive stylometry answers what does this text do linguistically? Authorship attribution answers who wrote this? Very different questions with very different evidentiary demands.

The framing this tool adopts is unambiguously descriptive: count features, surface patterns, and let the analyst interpret. See docs/limitations.md for what the tool is not.

Preprocessing

Quoted material

Direct quotations from sources. The tool removes them before analysis because they're the source author's choices, not the writer's. Double-quoted spans (straight and curly) and block quotes (indented or >-prefixed lines) are stripped.

Before stripping:

As Smith argues, "the rhetorical situation is always contingent on audience uptake," and this insight reshapes how we read student work.

After stripping (what the analyzer sees):

As Smith argues, and this insight reshapes how we read student work.

The report still notes how many quoted spans were removed and how many words they contained.

Sample size

How many words remain after quoted material is removed. Below 250 words the tool warns that results may be unreliable. The sweet spot is 400–1000 words.

Short samples make type-token ratio mechanically high (every word is unique) and per-500-word rates noisy. Long samples make TTR drop because common words start repeating.

§1 Lexical preferences

1.1 Lexical diversity (TTR / MATTR / MTLD)

How varied a writer's vocabulary is across the sample. The tool reports three related measures because each has different strengths.

High TTR (≈0.82)

The cartographer surveyed shifting dunes, inscribing their contours onto vellum. Each undulation yielded a fresh hieroglyph; no ridgeline rehearsed an earlier one.

26 tokens / 21 types. Almost no repetition.

Low TTR (≈0.42)

The thing about the thing is that the thing keeps doing the thing, and so the thing we should do is think about the thing the thing is doing to us.

31 tokens / 13 types. Heavy repetition.

TTR — raw type-token ratio. Distinct word forms divided by total word forms. Easy to read but length-sensitive: a 200-word sample will mechanically show higher TTR than a 2000-word sample by the same author, because common words inevitably repeat as text grows. For comparison across texts of different lengths the next two metrics are more defensible.

MATTR — Moving-Average Type-Token Ratio. Slides a 100-token window across the text and averages the TTR computed within each window. Because every window is the same size, MATTR is length-independent: two texts of very different total lengths can be compared on equal footing. Requires at least 200 tokens of material to compute.

MTLD — Measure of Textual Lexical Diversity. Walks the text counting how many tokens are needed before the running TTR drops below 0.72; each such span is a "factor". MTLD is total tokens divided by factor count. Bidirectional (forward and backward) and length-robust; widely used in published stylometric work. Requires at least 100 tokens to compute.

Comparison rule: the tool compares texts on MATTR when both samples are long enough (≥ 200 words each). MATTR diff ≤ 0.03 → Strong Match; 0.03–0.06 → Partial Match; > 0.06 → No Match. When MATTR is unavailable for either text, the tool falls back to raw TTR with looser thresholds (0.05 / 0.10) and the older 2× length-ratio guard. MTLD is reported in the per-text profile but does not participate in the comparison rating.

Small-sample note: under 100 words MATTR and MTLD are suppressed and only raw TTR is shown — and at those sizes even TTR is mechanically inflated. Under 200 words MATTR is suppressed. The tool surfaces an inline warning whenever a metric is omitted for length reasons.

1.2 Latinate / Germanic lean

English draws on two etymological streams. Latinate vocabulary (Latin/French roots) tends to be formal, abstract, polysyllabic. Germanic vocabulary (Old English/Norse roots) tends to be concrete, shorter, everyday. The same idea can be expressed in either stream, and a writer's habitual preference is a stable stylistic marker.

Latinate version

We endeavored to ascertain whether the methodology would facilitate substantive engagement with the material. Ultimately, the implementation necessitates careful articulation of underlying assumptions.

Germanic version

We tried to find out whether the way we taught it would help students get into the stuff. In the end, doing this means we have to spell out what we already think is true.

Ratio = Latinate hits / (Latinate + Germanic). ≥ 0.60 leans Latinate; ≤ 0.40 leans Germanic; in between is mixed. Both pure Latinate / pure Germanic texts share a lean → Strong Match. Opposite leans → No Match.

1.3 Pet Words & Habitual Phrases

Content words a writer reaches for repeatedly that aren't forced by the topic. With a topic hint, the tool separates these from words that are topical (required by subject matter).

Habitual word (filler):

I basically think there's basically two ways to handle this. The first one, basically, is to just talk to the student. The second is basically to wait and see.

"basically" appears 4× as a filler — not topic-required. The tool would flag it as habitual.

Habitual phrase (multi-word):

At the end of the day, what matters is how the writing sounds. And at the end of the day, that's not something you can fake.

"at the end of the day" appears 2× — flagged as a habitual phrase even though the individual content words don't recur.

Comparison: 2+ shared habitual markers → Strong; 1 shared → Partial; none shared → No. Both texts empty → Indeterminate.

1.4 Hedges, Fillers, Intensifiers

Three register-marking categories.

Informal hedges

It's kind of a problem. Basically, students are just trying to figure it out. I mean, honestly, who isn't?

basically, kind of, sort of, just, really, actually, honestly, like, I mean, you know

Intensifiers

The shift was absolutely dramatic — completely reshaping how we teach. The results were literally transformative.

very, extremely, absolutely, totally, completely, literally, definitely, clearly, obviously

Formal hedges

Perhaps the most useful framing is that the situation is arguably still in flux. It could be argued that consensus has not yet emerged.

perhaps, arguably, potentially, somewhat, it could be argued, one might suggest, it appears that

Both texts share dominant pattern (both informal, both formal, both low) → Strong. One informal-dominant + other formal-dominant → No Match.

1.5 Reading level

How difficult a text is to read, calculated from word length (syllables per word) and sentence length. Reported as two numbers: a US grade level (Flesch-Kincaid) and a 0–100 ease score (Flesch Reading Ease). Reading level often correlates with the Latinate/Germanic ratio (§1.2) — Latinate-heavy prose is usually harder to read because Latinate words are longer.

Grade 21 · Reading Ease −29 (very difficult)

The methodological challenges of artificial intelligence in higher education necessitate careful examination of underlying assumptions and their pedagogical implications. Universities must articulate clear policies that delineate appropriate use across multiple institutional contexts.

Long sentences, polysyllabic Latinate vocabulary, embedded subordination — reads as graduate / specialist.

Grade 4 · Reading Ease 91 (very easy)

I think AI in school is kind of a mixed bag. Everyone is using it, but nobody wants to talk about it. Teachers are stuck in this weird spot where they know what is going on, but they do not know how to deal with it.

Short sentences, everyday Germanic vocabulary, coordination over subordination — reads as elementary.

Flesch-Kincaid Grade Level. US school-grade equivalent for reading difficulty. Grade 8 = an 8th-grade reader can follow it; grade 12 = high school senior; grade 16 = college senior; above 16 = graduate / specialist. Computed from average sentence length + average syllables per word.

Flesch Reading Ease. A 0–100 scale where higher = easier to read. 90+ = elementary; 60–70 = standard / plain-English (roughly the level newspapers aim for); 30–50 = college-level; below 30 = very difficult (or, as with academic prose, "specialist"). Uses the same inputs as Flesch-Kincaid Grade Level but inverts the scale.

Comparison rule: the tool compares texts on Flesch-Kincaid Grade Level. Diff ≤ 2 grade levels → Strong Match; 2–4 → Partial Match; > 4 → No Match. Indeterminate if either sample is under 50 words or fewer than 3 sentences.

§2 Syntactic patterns

2.1 Sentence Length Distribution

Word counts per sentence. Mean tells you average length; standard deviation tells you whether the writer alternates short and long sentences or stays in a narrow band. SD is often more diagnostic than mean.

Low SD — uniform rhythm

The room was quiet. The students worked steadily. No one spoke aloud. The clock ticked through the period. Pages turned softly under hands.

5 sentences, mean ≈ 5.8, SD ≈ 1.3.

High SD — varied rhythm

The room was quiet. Students worked steadily, heads bent over their notebooks in the half-light of an October morning that had not quite decided whether to commit to autumn or hold one more week of summer. Pages turned. The clock ticked.

4 sentences, mean ≈ 11.5, SD ≈ 13.5.

Means within 3 words AND SDs within 3 → Strong; one of those → Partial; neither → No Match.

2.2 Sentence Openers

The first grammatical element of each sentence, in one of six categories. Writers tend to have a default opening pattern.

Six categories with examples:

Pronoun subject: "I think the policy is unclear." / "This means we have to revise."
Noun subject: "Universities are built around assumptions about authorship."
Transitional connector: "However, the situation has shifted." / "On the other hand, the evidence is mixed."
Adverbial / prepositional: "In this essay, I argue..." / "As a result, faculty must adapt."
Participial / gerund: "Struggling to articulate the point, she paused." / "Given this situation, we have options."
Coordinating conjunction: "But the real issue is different." / "And that means we need a new approach."

Same top two categories within 15 percentage points of each other → Strong; top matches but proportions diverge → Partial; top differs → No Match.

2.3 Coordination vs. Subordination

Coordinating structures (and, but, or, so) join clauses additively — one thing, then another. Subordinating structures (because, although, while, that, which) embed one clause inside another, producing complex, syntactically dense prose.

Coordination-heavy

The policy was unclear, and students were confused, and faculty were stuck, and nobody wanted to be the first to say anything. So we waited. And the longer we waited, the worse it got.

Strings of and / so joining independent clauses. Paratactic, additive rhythm.

Subordination-heavy

Because the policy remained ambiguous when students first encountered the new tools, which they had already begun integrating into their drafts before any formal guidance arrived, faculty who attempted to enforce consistent standards found themselves negotiating a terrain that the institution had not yet mapped.

Single sentence with multiple embedded clauses (because, which, that, who). Hypotactic, hierarchical.

Subordination/coordination ratio: ≥ 1.5 = high subordination; 0.75–1.5 = balanced; < 0.75 = high coordination.

2.4 Punctuation Rates

Frequency per 500 words of semicolons, colons, em dashes, parentheses, exclamation points, question marks, ellipses, and commas. Em-dash and semicolon use are especially strong personal-style markers.

Em-dash-heavy

The result — and this matters — was a kind of stylistic homogenization. Students started turning in work that sounded like everyone else's work — polished, uniform, oddly empty.

3 em dashes in ~35 words = ~43 per 500w.

Semicolon-heavy

The result was a kind of stylistic homogenization; students turned in work that sounded like everyone else's work; the polish was uniform; the prose was oddly empty.

3 semicolons in ~30 words = ~50 per 500w.

5+ punctuation types matching in rate → Strong; 3-4 → Partial; fewer → No Match. "Matching" means both within 50% of each other or both at zero.

§3 Discourse organization

3.1 Paragraph Structure & Topic Sentence Position

Where the paragraph's main claim sits relative to its sentences.

Topic first

Voice is something writers build, not something they find. When a student says they're "trying to find their voice," what they usually mean is that they're trying to figure out which choices feel theirs. Those choices accumulate across drafts.

Topic last

Students often describe voice as something external — out there, waiting. They talk about "finding it" the way you'd find a lost wallet. But voice doesn't work like that. Voice is something writers build, not something they find.

Other positions: embedded (middle sentence) and distributed (no single sentence dominates). The detector is heuristic — it scores sentences by claim-marker presence and topic-term density. Recommend human review.

3.2 Transition Strategy

How paragraphs connect to each other. Three approaches.

Explicit transitional

"However, this approach has limits..."

"Furthermore, the data suggest..."

"On the other hand, some argue..."

Overt connectors at paragraph starts.

Metadiscursive

"Now I want to discuss the implications..."

"Let me turn to the second problem..."

"In this section I will examine..."

Writer narrates the argument's structure.

Implicit

"Universities have responded slowly. Faculty meetings convened..."

"The technology was already widely adopted. Students reported using it..."

Reader infers the connection from content alone.

3.3 Evidence-to-Claim Sequencing

Four common patterns for relating claims to support.

Claim → Evidence → Interpretation: "AI tools homogenize student writing. In Smith (2024), only 12% of AI-assisted drafts retained the student's pre-tool sentence-length signature. This suggests that surface-level grammar correction is inseparable from voice flattening."
Claim → Evidence (no interpretation): "AI tools homogenize student writing. In Smith (2024), only 12% of drafts retained their original signature." (No "this suggests..." — the evidence speaks for itself.)
Evidence → Claim: "In Smith (2024), 88% of AI-assisted drafts converged toward a shared sentence-length profile. Across multiple institutions, similar patterns appear. The pattern is now widespread."
Claim → Elaboration: "AI tools homogenize student writing. Imagine a class of twenty, each writer with their own rhythms, and then imagine the same class after three weeks of running everything through the same tool: the voices converge." (No external evidence — reasoning, hypotheticals, personal experience.)

3.4 Metadiscourse

Moments where the writer talks about the text or the reader rather than the subject matter. Two subcategories.

Textual metadiscourse

In this essay I will argue that voice is built, not found. As I mentioned above, the apprenticeship matters. The following section takes up the implications. To summarize: smoothing is loss.

References to the text's own structure.

Interpersonal metadiscourse

You might think this is overstated. I believe it isn't. One could argue that students will find their voice eventually. It is important to note that "eventually" is doing a lot of work in that sentence.

Direct address to the reader, writer stance.

§4 Register and stance

4.1 Register Classification

The text's overall formality level. Four labels.

Formal

The proliferation of generative tools necessitates a reconceptualization of authorship in higher education. Institutions must articulate policies that delineate appropriate use, lest the prevailing ambiguity undermine the integrity of scholarly assessment.

No first/second person, no contractions, Latinate vocabulary, complex subordination.

Semi-formal

I argue that the proliferation of generative tools requires us to rethink authorship in higher education. Institutions need to articulate policies that specify appropriate use, since the current ambiguity is undermining assessment.

First-person allowed ("I argue"), avoids second person, mostly Latinate but lighter.

Informal

I think AI tools are basically forcing us to rethink what counts as someone's own work. Schools have to get their act together and just say what's okay and what's not, because right now everyone is making it up as they go.

First & second person, contractions, Germanic vocabulary, informal hedges, shorter sentences.

Mixed

The proliferation of generative tools necessitates a reconceptualization of authorship. But honestly, schools haven't figured out what to do, and students are basically left to figure it out on their own.

First sentence formal; the rest informal. Features from two levels in different parts of the text.

4.2 Register Consistency

Whether the register holds throughout the text or shifts between paragraphs. The tool reports each shift with paragraph index and direction.

Shifting register example:

The contemporary moment presents an unprecedented challenge for pedagogical practice in the writing classroom.

...

So basically, what I'm saying is, none of us have any idea what we're doing right now, and I think we should just be honest about that.

Paragraph 1 is formal; paragraph 2 is informal. The tool would flag this as a register shift at paragraph 2.

Register inconsistency can indicate multiple authorial processes (different sections drafted differently) or normal variation in a developing writer. The tool reports the shift; it doesn't interpret the cause.

4.3 Pronoun Profile

The mix of first-singular, first-plural, second-person, and third-person pronouns. One of the most reliable register markers.

First-singular dominant

I noticed something when I started reading my students' drafts. My sense is that voice is what I'm really trying to preserve. I kept asking myself why I cared.

Third-person dominant

Students develop voice through accumulated stylistic decisions. They figure out which patterns feel like theirs; they refine their preferences across drafts. Their voice emerges from their choices.

Second-person dominant

You might think voice is something you just have. But the more you write, the more you realize you're making choices. Your voice is whatever pattern your choices fall into.

Common in instructional / advice prose.

Impersonal / no pronouns

Voice in writing develops through accumulated choices about syntax, vocabulary, and paragraph structure. The development is gradual. The result is recognizable without being announced.

Common in formal academic prose.

§5 Comparison Ratings

Each of the 15 comparison features gets one of four ratings, based on thresholds defined in the methodology doc.

Strong Match

Both texts behave the same way on this feature, within the spec's thresholds.

Partial Match

The texts agree on the dominant pattern but diverge on secondary measurements.

No Match

The texts behave differently on this feature.

Indeterminate

Not enough signal — sample too short, no relevant tokens, length difference too large, or fewer than two paragraphs.

§6 AI-Writing Signs

Eight stylistic markers commonly found in large-language-model output, drawn from Wikipedia's Signs of AI writing catalog. The tool reports per-500w rates and example hits; it does not rate these as match/no-match. They are descriptive markers, not diagnostic tests. Humans use every one of these patterns.

For each sign below, the left column shows AI-style writing that triggers the marker; the right column rewrites the same idea without it.

AI Vocabulary

Words documented as statistically frequent in LLM output post-2022: delve, tapestry, underscore, pivotal, vibrant, meticulous, showcasing, intricate, fostering, highlighting, enhance, and others. Three time-stratified lists (2023–mid-24, mid-24–mid-25, mid-25+) — newer lists are more diagnostic because older words have leaked into general writing.

With AI vocabulary

Education is undergoing a vibrant transformation, with technology playing a pivotal role in shaping the evolving landscape. This intricate interplay underscores the crucial need for meticulous implementation, fostering deeper engagement.

Without

Education is changing, and technology is now central to how it works. The mix of policy and practice means we need to roll these tools out carefully so students actually learn from them.

Promotional Phrasing

Travel-guide / press-release style: boasts a, nestled in, in the heart of, diverse array, natural beauty, commitment to excellence, state-of-the-art, world-class, renowned for.

With promotional phrasing

Nestled in the heart of the university's curriculum, the writing program boasts a diverse array of courses, renowned for its commitment to excellence.

Without

The writing program offers a range of courses, including first-year composition, advanced rhetoric, and the senior thesis seminar.

Significance Emphasis

Generic statements that connect the subject to broader importance without earning it: stands as a testament to, pivotal role, underscores its importance, evolving landscape, indelible mark, setting the stage for.

With significance markers

This breakthrough stands as a testament to human ingenuity, underscoring its pivotal role in shaping the future of education. The development marks a turning point with lasting impact.

Without

This is an important advance. It will change how schools work.

Vague Attribution

References to unnamed authorities: experts argue, industry reports, several sources, leading scholars suggest, observers have noted, research indicates.

Vague

Industry reports and leading experts have noted the technology's impact, with several sources suggesting that adoption rates continue to climb. Research indicates measurable benefits across institutions.

Specific

A 2024 EDUCAUSE survey of 422 institutions found that 73% had drafted at least preliminary AI policies, up from 31% in 2023 (EDUCAUSE 2024, p. 14).

Negative Parallelisms

Contrastive constructions that deny one characterization before asserting another: not just X but Y, not only X but also Y, it is not X, it is Y.

With negative parallelism

AI is not just a tool, but a transformative force. It is not only reshaping education but also redefining what we mean by authorship. It is not a passing trend; it is a fundamental shift.

Without

AI is reshaping education and changing what authorship means. It looks like a lasting shift, not a passing trend.

Participial Pseudo-Analysis

Sentences ending in a comma + -ing clause that makes an unattributed analytical claim. Common starters: highlighting, underscoring, emphasizing, ensuring, reflecting, fostering.

With participial tails

Schools are adopting AI tools, fostering meaningful engagement across communities. Faculty have updated assessment practices, ensuring sustainable academic integrity. Students benefit, highlighting the importance of inclusive implementation.

Without

Schools are adopting AI tools because they find them useful. Faculty have updated assessment practices in response. Students benefit when the rollout includes accommodations for varied needs.

Rule of Three

Three parallel items in a list or series: X, Y, and Z. Common in human writing too, but overused in LLM output for rhythmic effect. The tool flags triplets where all three items share the same part of speech (all adjectives, all nouns, or all proper nouns).

With rule-of-three (overused)

The new framework engages institutions, disciplines, and communities. It addresses policy, pedagogy, and practice. It serves students, faculty, and administrators. The result is thoughtful, inclusive, and sustainable.

Four triplets in three sentences.

Without

The new framework brings together everyone who has a stake in how the institution handles AI tools. It covers what we should teach, how we should assess, and who should be at the table when policies are set.

One triplet, used purposefully.

Conclusion Formulas

Stock closing patterns: despite its challenges, looking ahead, the future outlook remains, in conclusion, moving forward, navigating these challenges.

With conclusion formulas

Despite its many challenges, the future outlook for AI in education remains promising. Looking ahead, institutions will continue navigating these challenges as the field moves forward.

Without

There are real problems with how AI is being adopted in schools, and we don't know yet which ones will be solved by better policy and which ones will just stay with us. The honest answer to "what happens next" is that we'll find out.