AI Writing Detection

The 2026 state-of-the-art in detecting AI-generated prose — matured from "amateur stylometry" to industrial-grade classification with sub-1% false-positive claims. The Economist's Is AI Writing Taking Over Westminster (Economist) promotes detection from academic curiosity to a governance mechanism for AI-authored policy documents.

The vendors

Pangram — market leader; claims 0.01% false-positive rate. A University of Chicago study called its error rate "near-zero" on longer texts.
Multiple competitors emerging; Pangram is the reference case cited in the piece.

The "smells" that leak from LLM prose

Love of contrasts. "not X but Y." "The technology has not changed. The ownership has."
Tricolons. Three-item lists ("careful sequencing, clear communication, and sustained commitment"). Humans have always liked groups of three (Cicero), but LLM enthusiasm is superhuman.
Em-dashes everywhere.
Certain rhetorical structures (parallelism, syntactic mirroring) at unusually high densities.

Modern detectors don't rely on these tells — they use similar machine-learning techniques to the models they're detecting. The stylistic markers are what a human reader notices before running the detector.

The Westminster case (2026 sample)

Pangram flagged as "mostly written by AI":

A policy paper by Louise Haigh MP (close adviser to Andy Burnham). Haigh's defence: AI was used for "version control."
"The Productive State" — a ~70-page manifesto from Mainstream, a Labour group linked to Burnham; one co-author used Claude to "expose line-level differences and unify the voice of both authors' independently drafted sections through copy editing."

Why detection has a governance role

The Economist's structural argument: policy proposers are effectively outsourcing their own working-out, in a role where the working-out is the deliverable. Any writer knows that putting words on a page tightens one's arguments. "AI certainly produces sloppy prose, but it also papers over sloppy thinking." Detection is how a reader gets some signal that this outsourcing has happened. But detection can only surface whether AI wrote the final product — not how authors used it. So the reader can flag the risk but not settle the debate.

The trust problem: when AI's involvement is disclosed only after the press come knocking, trust takes some earning back. Voluntary disclosure norms would help; a professional standard (as with academic-integrity policies) may follow.

Where this compounds

Individual-user parallel: Cognitive Offloading tracks the same "outsource your reasoning trace" phenomenon at consumer scale. The Westminster case is the professional-authorship variant.
Curriculum / education parallel: Is AI Making Us Dumber (Charlie Gedeon, TEDxSherbrooke) and Learning Software Engineering During the Era of AI (Raymond Fu, TEDxCSTU) track the education-side detection problem — the Pangram-class detector is what makes enforceable policy possible.
Regulated-profession parallel: Vibe Lawyering tracks the parallel move in law. Different profession, different verifier (courts / judges do adversarial checking), same underlying pattern.

Connects to

Is AI Writing Taking Over Westminster (Economist) — primary source.
Cognitive Offloading — the general phenomenon this concept detects.
Higher-Order Thinking — the thing being offloaded when AI writes for you.
Vibe Lawyering — regulated-profession parallel.
FOBO (Fear of Becoming Obsolete) — the anxiety this detection surfaces.
Andy Burnham — the political figure whose think-tank orbit produced two of the flagged documents.
Anthropic — Claude is named as one of the tools used.

2026-08-01 — the Economist's own LLM stylometry investigation (from the paired Leader "AI is getting better at writing")

The 2026-08-01 Economist Leader "AI is getting better at writing. Humans must get better." (brief in the 2026-08-01 edition MOC — not promoted to a dedicated source page in this ingest) reports the Economist's own methodology datapoint:

55,940 sentences across 14 model variants (ChatGPT / Claude / Gemini / Grok families) compared against the Economist's human corpus.
Fading tells: em-dashes, "leveraging", some Latinate bloat markers.
Persistent tells: sycophancy, structural parallelism, tricolons, contrast constructions.
Editorial upshot: the smells the Westminster piece first named are actively being trained away — human editing is now the bottleneck, and detection has to lean more heavily on ML-classifier tooling (Pangram-class) than on human stylistic pattern-matching.

Worth carrying because it's a rare Economist-internal LLM eval, not vendor-published; corroborates the vault's read that the detection arms-race is real and the human-visible smells are a shrinking signal.

2026-09-26 — the AI-writing arc as an Economist cover-Leader concern (chamber shares + LinkedIn + elite-publication tells)

The cover Leader (Dont Let AI Kill the Author) promotes the concern from "is AI writing at Westminster?" (2026-07-11 Westminster feature) to "AI-drafted text is a first-order civic problem across parliamentary institutions." Three-piece arc adds five specific first-in-vault anchors:

UK House of Commons: 1 in 10 words drafted by AI (~10%). Cover-Leader anchor.
US House of Representatives: 1 in 7 words drafted by AI (~14%). Cover-Leader anchor — first-in-vault US-House share.
LinkedIn: 41% of longer posts detected as AI-drafted (Pangram study; per AI Often Makes Writing Worse (Economist)). Vault's first specific professional-content channel-share.
Steve Yemm (UK MP): June 2026 parliamentary intervention on technology-shaping-society was flagged by Pangram as entirely AI-written. First-in-vault named MP + specific speech (per AI-Written Speeches Are Taking Over Politics (Economist)).
Elite-publication instances: Wall Street Journal ran an AI-drafted op-ed; French publisher Grasset defending its author against AI-drafting charge in a "prize-winning novel"; Gmail default "My Dearest [Husband's Name]…" (per AI Often Makes Writing Worse (Economist)).

The editorial arc-shape shift: Detection has moved from governance-mechanism-for-policy-documents (2026-07-11) to civic-infrastructure-layer-for-democratic-speech (2026-09-26). The Economist cover Leader's "loyalties in conflict: good prose vs creative destruction" framing is the vault's cleanest signal that the newspaper's own editorial-line on AI-writing has hardened — the "AI text is winning literary awards, occupying op-ed pages, and filling the mouths of politicians" verbatim is the load-bearing sentence.

The prose-quality thesis (per AI Often Makes Writing Worse (Economist)): AI makes the median writer better and the elite writer worse. Median rises, right tail collapses, standard-setting layer erodes. The 41% LinkedIn share is the professional-content signal; the elite-instances (WSJ + Grasset + Gmail default) are the standard-setter signal. Distinguishes this from the pure "is AI writing?" frame: yes, and the distributional effect is asymmetric.

2026-09-27 — The executive-etiquette response: disclose and call it out rather than detect (How AI Is Changing Communication (HBR IdeaCast))

HBR IdeaCast's AI-series episode adds the inside-the-firm angle. Wade Foster (Zapier CEO) calls out "Claudisms" and "AI slop" in colleagues' writing directly ("what do you actually think?"). Host Alison Beard defends the em dash, now treated as an AI tell. HBR's Scott Berinato calls LLMs "an averaging machine". The response proposed isn't detection tooling but norms: declare whether AI was author, co-author or assistant, and hold that the author should put in more effort than the reader (AI Communication Etiquette). Foster also reports that newer models seem worse at prose while getting better at code, which is one operator's impression. It fits the Economist's "median up, elite down" thesis above. See Authenticity Crisis (AI Communication).

2026-11-04 — The practitioner's verdict on detectors and humanizers (Gaspar, AI Daily Brief)

How to Deslopify Your AI (AI Daily Brief) adds the user-side read. Detectors measure structure — column and bullet density, same-length sentences, a few more features — and turning one on in LinkedIn lights up nearly everything as fully or mostly AI-generated. Nufar Gaspar's only sanctioned use is on your own work, as a way to move out of the average; she gives them little weight because they do not detect quality or originality, which is what a reader cares about. Humanizers (skills or processes that make text sound human) are "slop trying not to look like slop": humanized text is still flagged about three-quarters of the time, and the reader's problem was always the missing point of view, not the style. The Druckenmiller-op-ed debate reappears: stripping em-dashes and tells is a proxy for pretending you invested time; the better norm is to label — Figma's published rules for AI writing require AI-written content to say so at the top (AI Communication Etiquette). Whittemore's own tell of the month: he can no longer say "load-bearing".