wikipedia-not-ai — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited wikipedia-not-ai (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Derived from Wikipedia's empirically documented field guide on LLM writing patterns. Applicable to any text — not just Wikipedia.
Scan for signals across the categories below. Weight clusters more heavily than isolated occurrences — one "pivotal" is coincidence; five AI-vocab words in a paragraph is a strong signal.
Output format: Quote specific examples from the text, categorize each signal, then give an overall verdict: Likely AI / Possibly AI / Likely Human with brief reasoning.
Important: Newer models (GPT-5.1+, Claude 4+) actively suppress known tells like em dashes and classic AI vocabulary. Absence of classic markers is NOT proof of human origin. Look for subtler patterns: overly cautious hedging, Latinate word preference, "quietly" narratives, structural uniformity, the four deeper tells in Section 12 (abstraction trap, sensing without sensing, treadmill effect, subtext vacuum), and the in-disguise patterns in Sections 14–16 (decorative triplets, drumroll-in-disguise, false idioms/calques) — these are the most reliable signals on modern models.
LLMs statistically overuse specific words. Check for density/clusters.
See [word-lists.md](references/word-lists.md) for the complete list by model era.
Quick scan: pivotal, underscore, tapestry, delve, meticulous, vibrant, intricate, testament, bolstered, garner, fostering, showcasing, align with, enhance, highlighting
One word = coincidence. Five+ in a short passage = strong AI signal.
Adds statements about how mundane facts "represent a shift," "mark a pivotal moment," or "contribute to the broader landscape" — even for population data or etymology.
Trigger phrases: stands as, serves as, marks a pivotal, underscores its importance, reflects broader, symbolizing its enduring, setting the stage for, indelible mark, deeply rooted, evolving landscape
Rigid structure, usually at the end: "Despite its [positive adjective], [subject] faces challenges including... Despite these challenges, [vague optimism about the future]."
Tacks present participle phrases onto sentences as pseudo-analysis: "...highlighting its importance," "...underscoring the significance," "...reflecting the broader trend."
Attributes claims to no one specific: "Experts argue," "Observers have noted," "Industry reports suggest," "Several sources indicate," "Some critics argue."
Warm, advertisement-like prose even on neutral topics: nestled, vibrant, rich cultural heritage, breathtaking, diverse array, boasts, showcasing, groundbreaking, renowned, in the heart of.
LLMs take information from 1–2 sources but present it as widely held or as an incomplete list. Watch for: "including..." or "such as..." when the enumeration actually covers everything the sources say; opinions from a single source framed as consensus; "widely regarded," "generally considered" without evidence of breadth.
LLMs often describe actions or developments as happening "quietly" — a pseudo-narrative detail that adds false drama or modesty. Rare in factual human writing, common across multiple LLM families.
Replaces simple "is/are/has" with elaborate constructions:
Systematically chooses Latin/Greek-derived words over simpler Germanic equivalents, even in plain contexts:
This is distinct from copula avoidance — it's about lexical register, not syntax. As Orwell noted, bad writers are "haunted by the notion that Latin or Greek words are grander than Saxon ones." LLMs embody this tendency statistically.
Pseudo-balance: "Not only X, but also Y", "It's not just X, it's Y", "Not X — it's Y." Sounds thoughtful but is formulaic. See Section 14 for the more specific decorative triplets extension of this pattern.
Compulsive tripling: "adjective, adjective, and adjective" or "phrase, phrase, and phrase" even when two or four would be more natural. The decorative subtype is now treated separately — see Section 14.
Avoids repeating a word by substituting awkward synonyms throughout: subject → "the eponymous figure" → "the key player" → "this individual."
• **Bold Term**: description text pattern**bold**, # Header, - list item with dashes appearing in documents, emails, wiki markup, or any context where markdown is not the native format. Wikipedia added an edit filter specifically for this. Strong signal when combined with other tells.Text meant for the AI chat interface that got pasted into the document:
[INSERT NAME HERE], [Describe the specific section]Technical residue from specific AI tools:
turn0search0, turn0search1 — ChatGPT search citation artifactscontentReference[oaicite:0]{index=0} — ChatGPT reference rendering bugoai_citation, +1 inline — ChatGPT markup bugs{"attribution":{"attributableIndex":"X-Y"}} — ChatGPT JSON attributionutm_source=chatgpt.com or utm_source=openai in URLs[attached_file:1], [web:1] — Perplexity artifacts<grok_card ...> — Grok artifactsLLMs select illustrations by keyword-matching file names or titles, not by actual visual relevance:
When LLMs generate persuasive, argumentative, or discussion text (not just articles):
AI companies actively suppress known tells. Be aware:
Less common in newer models, useful for auditing older content:
Certain phrases sound human in isolation but become AI tells when used in the wrong context. Flag these when the context doesn't justify the emotional weight.
Legitimate only in: life-threatening situations, unexpected personal turning points, major life discoveries. ❌ Avoid in: software reviews, product descriptions, workflow descriptions, business writing. Example of misuse: "That combination changed how I work — and I didn't see it coming." (software review) Example of correct use: "The diagnosis came back positive. I didn't see it coming."
Same rule as above. Reserve for genuine surprise in high-stakes personal or life contexts. ❌ Avoid in: tool reviews, feature descriptions, onboarding narratives, professional writing. When in doubt: replace with a plain statement of the fact itself. "I now manage most of it through Claude" is stronger than "what I didn't expect: I now manage most of it through Claude."
Legitimate only in: references to authentic brands, original authorship, pure emotional states, sincere personal convictions. ❌ Avoid as a general intensifier in: product reviews, performance claims, workflow descriptions. Example of misuse: "genuinely changed how I work" (software review — use the plain verb instead) Example of correct use: "This is genuinely his work, not a ghostwriter's." / "I genuinely don't know what to do."
General rule: If a phrase adds emotional weight that the context hasn't earned, it's a tell. Software does not merit the same language as a near-death experience.
These four patterns operate below the level of vocabulary and syntax. They survive when a model is prompted to "avoid AI-isms" because they reflect the model's underlying nature, not its surface lexicon. Especially useful for auditing newer models (GPT-5+, Claude 4+) that have learned to suppress classic tells.
LLMs reach for abstract conceptual words instead of concrete, tangible ones — because they have read everything but experienced nothing. The text is unimaginable in the literal sense: you cannot form a mental image from it.
Markers:
Test: read a paragraph, close your eyes, and ask "what do I see?" If the answer is "nothing — just concepts," it is likely AI.
Russian/Ukrainian markers: комплексный, фундаментальный, многоаспектный, экосистема, парадигма, измерение, динамика процесса, ландшафт, поле, пространство (в переносном значении). Same effect: the abstract noun replaces the concrete image.
LLMs string together sensory descriptions that are statistically plausible but physically wrong — because they know about touch, smell, and sound without ever having a body in a room.
Classic example: AI describes spider silk as "smooth" because "silk" co-occurs with "smooth" in training data (processed textiles). Anyone who walked into a real web knows it is sticky and elastic.
Markers in health/nutrition/science writing:
Test: would someone who has actually done the thing recognize the description? If the language tracks word-associations rather than embodied reality, it is AI.
The text moves a lot but goes nowhere. Each paragraph rephrases the previous one without advancing the argument. The rate of new information per paragraph is near zero.
This is structural, not lexical — the words are different in each paragraph, but the idea-content is the same. It happens because LLMs are autoregressive: they know the next word from the latest word, but they do not know the last word. There is no destination conditioning the path.
Markers:
Test: try to write a one-sentence thesis the text is arguing toward. If you can't, the text doesn't have one.
AI explicitly explains every implication, joke, and emotional beat. It does not leave gaps for the reader to close, because it lacks a working theory of mind for "the reader" — ambiguity reads to the model as a hallucination risk to avoid.
Markers:
Test: cross out every sentence that explains what the previous sentence already showed. If the text gets significantly shorter and stronger, it was AI.
In Russian/Ukrainian popular science this is one of the strongest tells — a confident author trusts the reader to draw the conclusion. «Учёные обнаружили X. Это означает, что нам стоит делать Y» — the second sentence is often the AI tell. Good writing leaves Y as the reader's discovery.
LLMs distrust that a fact is interesting on its own, so they constantly announce importance and stage the reading experience. Where a human writer uses these devices sparingly (one drumroll per essay), AI uses them as a default rhythm — every 2–3 paragraphs. Closely related to Subtext vacuum (Section 12): the model is talking at a reader it can't actually model, so it manually choreographs the emotional reaction.
EN: Here's the kicker / Wait, it gets better / Plot twist: / Buckle up / But that's not all / Here's where it gets interesting / And here's the magic
RU: И тут на сцену выходит… / А теперь – внимание! – самое интересное / И вот здесь начинается магия / Но это ещё не всё / Готовьтесь, сейчас будет поворот / Спойлер:
UA: І тут на сцену виходить… / А тепер – увага! – найцікавіше / І ось тут починається магія / Але це ще не все / Спойлер:
For the masked variant of this pattern — drumroll dressed up as a neutral connective — see Section 15.
EN: Let's pause here for a moment / Before we move on / Take a breath / Step back for a second
RU: И вот здесь нужно остановиться / Давайте на секунду задержимся / Прежде чем двигаться дальше / Сделаем паузу
UA: І ось тут потрібно зупинитися / Давайте на секунду затримаємося / Перш ніж рухатися далі / Зробимо паузу
EN: What do you think? / Interesting, right? / You know what's funny? / Can you guess where this is going?
RU: А ты как думаешь? / Интересно, правда? / Знаете, что самое забавное? / Догадываетесь, к чему я клоню?
UA: А ти як думаєш? / Цікаво, правда? / Знаєте, що найкумедніше? / Здогадуєтеся, до чого я веду?
A human writer chooses to use a drumroll — and uses it once, where it earns its place. An LLM defaults to it because the model doesn't trust that the underlying fact will land. If the same text contains 3+ drumrolls, 2+ fake pauses, or pseudo-dialogue every few paragraphs, that pattern is structural, not stylistic.
Test: delete every drumroll, fake pause, and rhetorical question in the text. If the remaining content is still coherent and the meaning is unchanged, the staging was filler — strong AI signal.
Stand-up comedy, podcast transcripts, motivational speaking, and children's writing legitimately use these devices often. Weight them by genre — in straight prose (essay, news, analysis, marketing copy, popsci), even one or two of these per page is unusual; in scripted oral genres, the bar is much higher.
Rule of three is a human technique with a long history — when used iteratively (weak → stronger → strongest), or to draw a real contrast, or to enumerate three actually distinct things, it earns its place. Decorative triplets are different. The third member is synonymic with the first two; it adds rhythm, not information. LLMs default to decorative triplets because three feels balanced statistically, even when the third slot has nothing new to put in it.
A real triplet has either an escalation, a contrast, or three genuinely distinct elements:
A decorative triplet has three near-synonyms doing the same work:
EN decorative triplets (typical AI output):
RU декоративные триплеты:
UK декоративні триплети:
Drop the third element. If the meaning is unchanged — it was decorative. If the meaning narrows or loses force — it was a real triplet.
One decorative triplet in a long article is noise. Three or more in the same piece — especially in headline positions (titles, opening sentences, summary blocks) — is a structural AI tell. Pair this with significance inflation (Section 2) and you have one of the cleanest signatures in current AI marketing copy.
The Section 13 drumrolls are theatrical — they declare themselves («А тепер увага», «Wait, it gets better»). Drumroll-in-disguise is the same function wearing a neutral suit. The model still wants to flag importance, but it uses connectives that read as ordinary academic prose, so the staging is invisible at first glance. The give-away is function over content: the phrase announces that something noteworthy is about to follow — but the thing that follows is not particularly noteworthy. The staging fires by reflex, regardless of payload.
EN:
RU:
UK:
A human writer reaches for «варто зазначити» once or twice in a long piece, in front of something that genuinely needed to be flagged because the reader wouldn't otherwise see its importance. The AI reaches for it as a metronome, often three to five times per page, in front of facts that don't need flagging at all.
Test: for each instance, ask — would the surrounding paragraph be weaker if the connective were deleted? If no, the connective was a tick, not a tool.
The phrase is grammatically respectable. It appears in academic and journalistic prose. So readers' AI-detection antennae stay down. But the function is the same as the theatrical drumroll: choreograph the reader's reaction, manufacture importance the text hasn't earned. When you train yourself to notice the rhythm — every second or third paragraph leading with «Стоит отметить» — the disguise drops.
Closely related: Importantly, ... / Crucially, ... / Critically, ... / Significantly, ... as paragraph openers. RU: Важно, что... / Принципиально, что... / Ключевым моментом является то, что... These are drumroll-in-disguise with extra emphasis built into the adverb itself.
LLMs are trained on multilingual corpora where the English distribution dominates. When the model generates Russian or Ukrainian, it sometimes maps an English idiom directly onto the target language, producing constructions that are grammatical but not idiomatic — they sound plausible, but a native speaker would not say them. The same happens in reverse for less-resourced languages: the model imports English phrasing into RU/UK because its statistical pull is stronger than the actual native usage.
The canonical example for this skill: «не дома в теме» — calqued from English «at home in [a subject]» (= comfortable with, well-versed in). In Russian, «дома в теме» is not an idiom. A native speaker says «разбирается в теме», «свободно ориентируется». The calque survives because each word is correct in isolation; only the construction is wrong.
Borrowings are conscious imports that have entered the language and become normative (бэкап, дедлайн, фидбэк). Calques are unconscious — the model has not consulted whether the target language uses this expression, it has only translated word-by-word. Borrowings appear in dictionaries; calques do not.
RU calques from English (typical AI output):
UA calques from English (often via Russian AI training data):
EN calques from formulaic training data (rarer, but exist):
Three reasons. First, the model has seen the English version many times in its training, and the word-by-word substitution into RU/UK produces something that scans grammatically. Second, the model has no embodied sense of what people actually say — only of what could be said. Third, post-training filters look for vocabulary tells more than for syntactic-idiomatic ones, so calques pass through where delve would have been caught.
For any suspect phrase, ask: does a native speaker actually say this, or am I just understanding it? Comprehensibility is not idiomacity. If you can imagine reading the phrase but can't imagine saying it without quote marks, it's likely a calque.
When in doubt, search the phrase in the target language on a major search engine restricted to native-language sources. If the only hits are translated content, machine-generated content, or marketing pages — the construction has not entered native usage.
One calque can happen in human writing too — bilinguals leak. But three or more in a single piece, especially when paired with abstraction-trap vocabulary (Section 12) and decorative triplets (Section 14), is a strong AI signature. Calques are particularly diagnostic because they survive even aggressive prompt instructions to "sound natural" — the model doesn't know the construction is non-native.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.