Skip to content

Detector tell · 4 patterns · weight 2

The connectors that convict formal human writing

Four rows, weight 2 each, matching the standard connectives of academic and administrative English. They are the only phrase family with a verdict floor of its own, shared with the structural markers.

Our detector calls this category “Formal connector”. The patterns below are read out of functions/api/tools/detect.js when this page is built, and the examples are produced by running the shipped scorer over the 48 licensed samples in our benchmark. If a claim here stops matching the engine, the build fails.

What our own benchmark saw

Appeared in 2 of the human samples and 3 of the AI samples, 5 occurrences in total.

Human-written academic writing, authors at non-anglophone institutions document scored 24 — Reads mostly natural

Matched: “Consequently”

  • “Consequently, the students solved the problem correctly.”

    Our scorer gave this sentence 52 out of 100.

Source:
Katarina Tri Utaminingtyas, Rachmadina Eka Herdianti, Inti Hayatul Fitria & Anton Prayitno, "Small Groups: Student Productive Interactions in Learning Cooperative (Case Study of Mathematics Learning at Junior High School in Pakis, Malang)", Educational Process: International Journal 6(2), 2017
Licence:
CC BY (journal-level license recorded in DOAJ for Educational Process: International Journal, ISSN 2147-0901)
How we know the date:
Published 15 June 2017 — verified against Crossref (api.crossref.org/works/10.22521/edupij.2017.62.3 returns published date-parts [2017,6,15]) — over five years before any general-purpose LLM was publicly available, so human authorship is a date fact. Non-native authorship is a strong inference rather than a certainty: all four author names are Indonesian, and the study was run in a junior high school in Pakis, Malang, Indonesia (stated in the article title). The prose carries textbook L1-Indonesian transfer features (subject-verb disagreement, missing articles) that no copy-editor removed.
Human-written government notice document scored 27 — Reads mostly natural

Matched: “Additionally”

  • “Additionally, this rulemaking broadens the opportunities for military instructor pilots or pilot examiners to obtain civilian ratings based on military experience, expands opportunities for logging pilot time, and removes a burden from sport pilot instructors by permitting them to serve as safety pilots.”

    Our scorer gave this sentence 52 out of 100.

Source:
Federal Aviation Administration, "Regulatory Relief: Aviation Training Devices; Pilot Certification, Training, and Pilot Schools; and Other Provisions", final rule, Federal Register document 2018-12800
Licence:
Public domain — work of the US federal government (17 U.S.C. §105)
How we know the date:
Publication date 2018-06-27, returned by the Federal Register API for document_number 2018-12800; the document number itself encodes the year and is a permanent citable identifier. Four years before public LLMs, and produced by a federal agency under notice-and-comment rulemaking.
AI-written AI asked plainly document scored 42 — Mixed — some AI-sounding passages

Matched: “Furthermore”

  • “Furthermore, advances in collaboration technology have largely eliminated the practical barriers that once made remote arrangements difficult to sustain.”

    Our scorer gave this sentence 63 out of 100.

Source:
claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.
Licence:
generated for this benchmark
How we know the date:
No human wrote or edited any part of this text — it was produced token-by-token by the model in the session that wrote this file, and the session transcript is the primary record. The prompt was a bare content request with no style instruction, so this is the model's untouched default register. Residual risk (stated, not dismissed): the model could in principle have reproduced memorized training text; that was not independently checked against a plagiarism index.
AI-written AI asked plainly document scored 30 — Reads mostly natural

Matched: “Additionally”

  • “Additionally, language learning provides direct access to another culture, allowing learners to engage with literature, film, and everyday conversation in their original form rather than through translation.”

    Our scorer gave this sentence 52 out of 100.

Source:
claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.
Licence:
generated for this benchmark
How we know the date:
No human wrote or edited any part of this text — it was produced token-by-token by the model in the session that wrote this file, and the session transcript is the primary record. Bare content request, no style instruction; the em-dash is the model's own default punctuation habit, not an instructed tell. Residual risk (stated, not dismissed): the model could in principle have reproduced memorized training text; that was not independently checked against a plagiarism index.
AI-written AI asked plainly document scored 42 — Mixed — some AI-sounding passages

Matched: “Moreover”

  • “Moreover, the reputational damage often outlasts the technical incident itself, as customers hesitate to share payment information with a company that has already lost it once.”

    Our scorer gave this sentence 52 out of 100.

Source:
claude-opus-5 (Anthropic), generated 2026-08-26. The model was run as a Claude Code agent tasked with building this corpus; it issued the recorded prompt to itself and wrote the completion in the same session. Output copied verbatim into this file with no human drafting, editing, or trimming.
Licence:
generated for this benchmark
How we know the date:
No human wrote or edited any part of this text — it was produced token-by-token by the model in the session that wrote this file, and the session transcript is the primary record. Bare summarization request with no style instruction. Residual risk (stated, not dismissed): the model could in principle have reproduced memorized training text; that was not independently checked against a plagiarism index.

2 of those samples were written by a person, years before any general-purpose model existed, with a licence and a publication date on record. 3 were machine-written. That is the whole reason these pages exist: a hit here is a fact about a phrase, not about an author.

About the non-anglophone sample above

READ THIS BEFORE PUBLISHING ANY ESL NUMBER. The five "esl-nonnative" samples are NOT learner-corpus essays. No learner corpus with a license compatible with commercial use could be sourced: PELIC is CC BY-NC-SA (NC excludes us), the Cambridge Learner Corpus FCE set, ICLE, ICNALE, NUCLE and the BEA-2019 W&I+LOCNESS data are all behind restrictive or registration-gated licenses, and the ETS TOEFL11 corpus is a paid LDC product. WHO and FAO publications were also rejected as CC BY-NC-SA. Rather than substitute something unlicensed, these five are CC BY / CC BY-SA scholarly prose written by authors whose institutional affiliations are all in non-anglophone countries (Indonesia, Turkey, Iran, China), selected because visible L2-transfer features survived peer review uncorrected. That is evidence of non-native authorship, not proof of any individual author's first language, and the register is academic rather than the undergraduate essay that produces our worst false positives. Treat any FPR computed on this slice as a floor, not as the ESL-essay false-positive rate.

In short: do not read this page as evidence about how detectors treat non-native English writers. It is one licensed, dated document that happens to contain the phrase.

A constructed example — written by us, not evidence

This sentence was written for this check. It is not from the benchmark, nobody wrote it as real prose, and it says nothing about how anyone writes. Its only job is to prove the rule still fires, and the generator fails the build if it stops.

“Moreover, the survey was repeated in March. Furthermore, response rates rose.”

  • /\bmoreover\b/gi → “Moreover”
  • /\bfurthermore\b/gi → “Furthermore”

What it means when a person writes this way

These four words are what people were taught to write. They are the connective tissue of a legal notice, a grant application, a lab report and a translated abstract, and a writer using them is usually doing exactly what their register asks of them. The human samples above that trigger this label are documents with a licence and a publication date on record years before any general-purpose model existed. There is no question about who wrote them.

The patterns, as the engine holds them

4 of the 70 rows in our phrase table carry this label. They are printed here as they are written, because a paraphrase of a regular expression is a different regular expression.

/\bmoreover\b/gi weight 2

Matches “moreover” , in any capitalisation .

/\bfurthermore\b/gi weight 2

Matches “furthermore” , in any capitalisation .

/\badditionally\b/gi weight 2

Matches “additionally” , in any capitalisation .

/\bconsequently\b/gi weight 2

Matches “consequently” , in any capitalisation .

What a hit does to your score

The phrase model is one of four in our ensemble, and its formula is patternScore = min(98, max(5, 10 + 7 x total weight)). One occurrence of the heaviest row in this category takes that model from 10 to 24 out of 100 — and the pattern model carries 40% of the ensemble when the em-dash signal fires and 55% when it does not.

This is the only phrase family in the engine with a verdict floor of its own. Hits of this label and of "Structural marker" are counted together: three of them force the score to at least 68, four to at least 78, regardless of what every other model in the ensemble says. If you write formal English and a detector has convicted you, this is the most likely mechanism.

Check your own text

Our detector is free and shows you which sentences it flagged and why, including this category. It also gets things wrong, and we publish how often.

Run the detector Our measured error rates If you have been accused Which registers score highest

Other things the detector looks for