--- title: "Free AI Writing Checker: Content Shape Checker" url: "https://aivisibilitycensus.com/content-shape-checker/" description: "Free AI writing checker: four tested checks for AI patterns, every problem sentence quoted with a fix, and a checklist of claims to verify. No signup." publisher: "AI Visibility Census" date_published: "2026-10-29" date_modified: "2026-10-29" --- # Free AI Writing Checker: Content Shape Checker Free toolNo signupRuns in your browser # Free content shape checker A free AI writing checker that reads structure first, plus the stock phrasing that held up in testing. Paste a draft and get every problem sentence quoted with a fix, a tested reading of how AI-written it looks, and a checklist of the claims to verify before you publish. [Check your content](#checker) [See an example report](#example) ![Content Shape Checker report for an AI-written blog post about employee directories: fails 3 of 4 checks, AI-style signals High, 2 claims to verify with 1 unsourced, and the strongest habit named as even rhythm, with sentence lengths in a narrow band.](https://aivisibilitycensus.com/wp-content/plugins/human-shape-checker/assets/img/report-scorecard.webp) 4 structural checks 18 of 26 AI-written test texts read High 0 bytes of your text sent to a server Every claim listed to verify, unsourced stats flagged The checker ## Paste a draft or drop a file Paste from Google Docs or Word to keep your headings. In plain text, mark headings with ##. Content to check Upload file Your text is analyzed in your browser. It is not uploaded, stored or logged. Why it reads structure ## Swapping words doesn't fix it Most AI writing checkers count tells: em dashes, "delve", "in today's fast-paced world". Writers learned to delete those in an afternoon, and current models had already dropped most of them. The draft still reads generated, because most of the giveaway is structure. In April 2026, researchers at the University of Maryland and Google DeepMind took 61,608 stories, written from the same 10,272 prompts by human authors and five language models. They stripped out every stylistic feature and kept only structure. Structure alone told human from AI at 93.2% macro-F1. The five models also wrote stories that looked like each other. The human stories were spread all over the map. That study was on fiction, so we tested every rule on 1,952 business, news and essay documents and kept only what told human from AI writing. Most of what survived is structure: rhythm, named specifics, how a piece narrates its own meaning. One check is about words, because current models have swapped "delve" and "crucial" for plain words like "quietly" and "practical", and the swap is countable. [Read the paper](https://arxiv.org/abs/2604.03136) or [see how we tested it](#method). | | Typical word-level AI checker | Content Shape Checker | | --- | --- | --- | | Looks at | Individual words and phrases | How the piece is built (rhythm, sections, named specifics), plus the stock phrasing that held up in testing | | You get | A percentage | Each flagged sentence, the reason and the fix | | Tested against human writing | Rarely published | Yes. On held-out business writing, two or more checks failed on 31% of AI texts and 2% of human ones | | AI detection | A percentage, usually with no reasons | A Low, Medium or High reading of AI-style signals, with the evidence for each. Never a verdict on who wrote it | | Facts and figures | Varies | Lists every claim to verify and flags statistics with no source. It does not judge whether they are true | | Where your text goes | Varies. Check the privacy policy | Nowhere. It stays in your browser | What it checks ## Four shape checks, plus claims and AI-style signals 1. 01 Over-explanation Generated text narrates what things mean. The checker flags trailing clauses like ", highlighting the importance of...", claims that something "plays a crucial role", the sentence that interprets a number just given, and sections that close on "Ultimately". BeforeNew starters found their manager in under a minute, highlighting the crucial role a directory plays in onboarding. AfterNew starters found their manager in under a minute. Last year it took them most of a morning. 2. 02 Real things, named People with jobs name customers, tools, numbers and places. The checker counts those anchors per section and flags category language like "organizations of all sizes". BeforeOrganizations of all sizes are looking for ways to connect their teams. AfterA 340-person engineering firm in Cork timed it: new starters needed 11 minutes to find who owned a system. 3. 03 Rhythm and variety Models write to a narrow band. The checker measures the spread of sentence length, how many sentences run long, how much paragraphs and sections vary, how often you use an aside, and how often you reach for a list of three, each against human business writing. BeforeEvery sentence 12 to 16 words. Five sections of about 120 words each. No parentheses, no questions, four lists of three. AfterA six-word sentence, then a 35-word one with the caveat in parentheses. A 60-word section next to a 300-word one. 4. 04 Stock AI phrasing The words and constructions models lean on. In our tests, 2023 models leaned on "landscape", "crucial" and "seamless", and 2026 Claude models on "quietly", "steady" and "practical". Plus "not just X, but Y", "Y, not X" and "Additionally". One word never fails a piece: it takes at least three, at a high rate. BeforeIn today's landscape, a directory isn't just a list of names. It quietly keeps teams steady, not scattered. AfterThe directory answers one question: who do I ask? 5. + Claims to verify *Not scored* Language models write figures and studies that sound right and do not exist. The checker cannot tell you which ones are true, so it does the part it can do honestly: it lists every statistic, study, superlative, quote and date in your draft, and flags the numbers with no source nearby. You get a checklist to work through before you publish, in the report and in the PDF. FlaggedStudies show that employees spend up to 20% of their time looking for information. Checklist itemStatistic, research claim: no source nearby. Which study? Link it, name it, or cut the number. 6. + AI-style signals *Not scored* A free, private reading from a small model that reads structure: sentence length, how many names and numbers you use, parenthetical asides, how much section and paragraph lengths vary. You get Low, Medium or High and what drove it. In our test, High caught 18 of 26 AI-written texts and 2 of 48 human ones. A prompt to look closer, never proof of who wrote it. What drove itNames and numbers: less than typical human writing (2.1 vs 5.4 per 100 words). ReadingAI-style signals: High. The structure looks a lot like AI-written blog posts in our test set. AI detection, tested ## A free AI detector that shows its working Most free AI detectors give you a percentage and no reasons. This one gives a Low, Medium or High reading, shows what drove it, and publishes how often it gets it wrong. 18 of 26 AI-written test texts read High 2 of 48 human-written test texts read High 300 words minimum for a reading 0 bytes of your text sent anywhere [Check a draft](#checker) ![AI-style signals reading of High for an AI-written test post, listing what pushed it toward AI-style: shorter average sentences, fewer names and numbers, fewer long sentences, no parenthetical asides and less variety in section length, each compared with typical human writing.](https://aivisibilitycensus.com/wp-content/plugins/human-shape-checker/assets/img/report-signals.webp) *A real reading of an AI-written test post. Every signal shows this draft's number next to typical human writing.* ### Every test result, not the best ones Each mark is one text the model never saw in training. Human-written texts mostly land in Low. AI-written ones mostly land in High, and 3 of 26 slipped through as Low. Swipe sideways to see the full chart. ### How we tested it 1. **Collected real human writing.** Company blog posts, handbooks and style guides published before 2022, before ChatGPT existed. 2. **Wrote AI versions of the same posts.** Three AI models wrote posts on the same titles at similar lengths, so topic and length cannot give the answer away. 3. **Scored texts it never saw.** We trained on posts from two AI models and tested on posts from a third it had never seen, plus human texts it had never seen. Every number on this page comes from that held-out part. ### Where it gets it wrong - Medium is weak evidence: 9 of 48 human texts landed at Medium or High. - Our AI test texts came from Claude models. ChatGPT and Gemini drafts may score differently. - Short, punchy copy can read as AI-style even when a person wrote it. - Carefully edited AI text can read Low. - It never proves who wrote something. Do not use it to accuse anyone. [Full method and limits](#method) Example report ## One draft, run twice We generated a typical blog post, "Why Employee Directories Matter in Today's Workplace", and ran it through the checker. Then we rewrote it around one specific situation and ran it again. These are the tool's real outputs. Generated draft Fails 4 of 4 - **5** sentences narrating what things mean - **0.4** numbers or names per 100 words - **2.8** words of sentence-length spread (human writing: 8.7) - **9** stock AI phrases, including 3 "not just X" constructions - **1** statistic with no source ("studies show" 20%) Rewrite Passes 3 of 4 - **0** sentences narrating meaning - **14.6** numbers or names per 100 words - **Notes** on rhythm: no sentence over 30 words, sections still similar in length - **0** stock AI phrases - **3** claims listed to verify, 2 with no source The rewrite's company and figures are invented to show the shape, and the checker rightly lists them as claims to verify. It still gets a note on rhythm: every sentence is short, which reads clipped next to human writing. Passing the shape checks says nothing about whether your numbers are true. Both drafts are under 300 words, where the tool's results are rougher. ![The generated draft marked up by the checker: sentences that narrate meaning, category phrases and stock AI phrasing highlighted in different colors.](https://aivisibilitycensus.com/wp-content/plugins/human-shape-checker/assets/img/report-markup.webp) *The marked-up text. Each color is one check. Hover a mark on the live tool to see the reason.* ![Rhythm and variety panel for the generated draft: sentence-length spread of 2.8 words against 8.7 for human writing, no long sentences, uniform paragraphs and sections, no asides, and a flat sentence-length chart.](https://aivisibilitycensus.com/wp-content/plugins/human-shape-checker/assets/img/report-rhythm.webp) *Sentence rhythm. Generated text tends to run flat; human prose looks jagged.* ![Claims to verify checklist from a test draft: 7 claims, 3 with no source nearby, each tagged as statistic, research claim, price or superlative with its source status.](https://aivisibilitycensus.com/wp-content/plugins/human-shape-checker/assets/img/report-claims.webp) *The claims checklist from a test draft. Every figure, study and superlative is listed, and the ones with no source nearby are flagged. It never says whether a claim is true.* Example export ### "Copy report" gives you Markdown you can paste into a doc, a ticket or a freelancer brief ``` # Content Shape Checker report 256 words, 6 sections. 1 claim to verify, 1 with no source nearby. | Dimension | Result | Strikes | |---|---|---| | 1. Over-explanation | FAIL | 5 | | 2. Real things, named | FAIL | 10 | | 3. Rhythm and variety | FAIL | 5 | | 4. Stock AI phrasing | FAIL | 9 | ## Over-explanation - "This means that a directory can save a significant amount of time." - "Ultimately, better communication drives better outcomes for the entire organization." - "In short, collaboration is the key to unlocking your organization's potential." ## Rhythm and variety - Sentence lengths sit in a narrow band (spread 2.8 words; human business writing we tested averaged 8.7). - Section lengths are uniform (variation 0.1; human average 0.64). - Almost no asides: 0 parentheses, semicolons and questions per 1,000 words (human average 8.6). ## Stock AI phrasing - "An employee directory is not just a list of names, it's a powerful tool for collaboration." - "With a comprehensive directory, teams can communicate seamlessly across departments." - "A robust directory empowers employees to connect with colleagues across various teams." ## Claims to verify (not scored) 1 of 1 has no source nearby. The checker does not judge whether a claim is true. - [ ] "Studies show that employees spend up to 20% of their time looking for information." (statistic, research claim; no source nearby) ``` Excerpt. The full export lists every flagged sentence. How to use it ## Three steps, about two minutes 1. Paste or upload Use the full draft with its headings. The section checks need them. 2. Read the strongest habit first The report names the one habit doing the most damage. Fixing it usually clears several flags at once. 3. Fix, then run it again Cut the sentences that narrate meaning, swap category language for a name or number, let one sentence run long and one section run short. Work through the claims checklist, then re-check. Who it's for ## Anyone who has heard "this sounds like ChatGPT" ### Content leads Review a freelancer's draft in two minutes and send back the exact sentences to fix, plus every figure that needs a source. ### SEO and GEO teams Check a page before it ships, including whether the opening sentence answers the question. ### Agencies Paste the Markdown export into a client review so feedback points at sentences, not vibes. ### Founders writing their own copy Find out which habits you picked up from your AI drafts, and drop them. Shape and AI citation ## Human shape and AI citation pull in different directions The [AI Visibility Census](https://aivisibilitycensus.com/) measures which firms AI answers name and which sources they cite. Pages that get cited tend to open with the answer, state figures with their window and keep claims self-contained. Some of that looks tidy, which is the opposite of what human-shape advice says. So the checker scores the opening sentence separately. Keep an answer-first opening and self-contained FAQ answers. Get your human shape from named specifics and varied rhythm, not from leaving things messy. FAQ ## Questions people ask Does it detect AI-written content? Partly, and we publish how well. The AI-style signals section gives a Low, Medium or High reading from a small model that reads structure, not word choice, trained on human-written company blog posts from before ChatGPT and AI-written posts on the same titles. On texts it had never seen, High caught 18 of 26 AI-written texts and wrongly flagged 2 of 48 human ones. Medium also caught 9 of 48 human texts, so treat Medium as weak. The AI test texts came from Claude models, so drafts from other tools may score differently, and punchy copy with lots of short sentences can read as AI-style even when a person wrote it. It gives no percentage and never proves who wrote a piece. Does it fact-check my content? No, and it does not pretend to. It lists every claim a reader could check (statistics, studies, superlatives, quotes and dates) and flags statistics with no source nearby. A person still has to check each one. Language models can produce figures and citations that sound right and do not exist, so the list is a starting point for verification, not a verdict. Is the Content Shape Checker free? Yes. There is no signup, no paid tier and no limit on how many drafts you check. Each check runs on up to 3,000 words. Do you store or read my content? No. The analysis runs in your browser. Text you paste and files you upload are never sent to our server or anyone else's, so there is nothing for us to store. What can I upload? Word (.docx), plain text (.txt), Markdown (.md) and HTML files, or paste directly. Pasting from Google Docs or Word keeps your headings and lists, which the checker needs to judge section structure. Why can't I upload a PDF? Text pulled out of a PDF loses its headings, lists and paragraph breaks, which is exactly what the checker measures. Copy the text out of the PDF and paste it, then add headings back with ## if you need to. What research is it based on? It started from StoryScope (Russell, Rajendhran, Pham, Iyyer and Wieting, University of Maryland and Google DeepMind, April 2026), which separated human from AI fiction at 93.2% macro-F1 using structure alone. Because that study was on fiction, we tested every rule on 1,952 business, news and essay documents, kept what separated human from AI text, and dropped what did not. The stock-phrasing lists come from Wikipedia's Signs of AI writing guide, and the rhythm measures match findings by Desaire and others (2023). On held-out business writing, two or more checks failed on 31% of AI texts and 2% of human ones. If my content passes, will AI search cite it? Not on that basis alone. Pages that get cited in AI answers also tend to open with the answer, state figures with their source and time window, and make claims that can be lifted out cleanly. The checker flags whether your opening sentence is answer-first, but it does not predict rankings or citations. Does it work on fiction? It is built for business and marketing writing: blog posts, landing pages, case studies, newsletters, reports. Fiction has its own structural tells (how endings resolve, how many subplots stay open) that this version does not check. How much text do I need? At least 80 words to run, and up to 3,000. It works best on a single article or page of 300 to 1,500 words. Past 1,500, results average across the whole piece, so weak sections can hide behind strong ones; check long documents a section at a time. The section checks need at least four or five headed sections to mean much. ## How the checker works The checker started from [StoryScope](https://arxiv.org/abs/2604.03136) (Russell, Rajendhran, Pham, Iyyer and Wieting; University of Maryland and Google DeepMind, preprint April 2026), which separated human from AI fiction at 93.2% macro-F1 using structure alone across 61,608 stories. That study was on fiction, so in version 1.5 we tested every rule on business writing and news, kept what separated human from AI text, and dropped what did not. ### How we tested the checks 1,952 documents. Our own set: 48 company blog posts and 64 handbook and guide pages written by people before 2022, 48 posts on the same titles from Claude Haiku, Sonnet and Opus, and 10 generic Claude Sonnet posts. Two public sets from the [Ghostbuster benchmark](https://github.com/vivek3141/ghostbuster-data) (Verma and others, 2023): Reuters news articles and student essays, each paired with AI versions that GPT-3.5 and Claude wrote from prompts based on the human originals. We set the thresholds on 986 of them and report results on the other 966, which the thresholds never saw. On the held-out business writing, the four checks together failed AI texts far more often than human ones: an average of 0.9 failed checks per AI text against 0.1 per human text. Two or more fails happened on 31% of AI texts and 2% of human ones. The previous version scored 0.6 against 0.5, barely better than chance. - **Over-explanation.** Trailing clauses that announce meaning (", highlighting the importance of..."), significance claims ("plays a crucial role"), sentences that interpret a number just given, and sections that close on "Ultimately" or "In short". These are strongest against 2023-era GPT: 47% of its held-out essays failed, against 6% of human essays, and it flagged 17 to 19% of AI news articles against 3% of human ones. Current Claude models rarely do them, so a pass is weak evidence. Phrases such as "in other words" and "put simply" are no longer scored: human writers used them as often. - **Real things, named.** Numbers, names, tools, dates and places per section, against category language such as "organizations of all sizes". Unchanged. AI texts failed it 27% of the time, human texts 4%. - **Rhythm and variety.** Spread of sentence length, share of sentences of 30 words or more, variety in paragraph and section length, asides (parentheses, semicolons, questions) and lists of three, each compared with the human business writing in our set. AI texts failed it 50% of the time, human texts 6%. Repeated section openers and heavy bullet use are no longer scored: human writers did both as often as AI. Less variation in sentence length and fewer parentheses, semicolons and questions in AI text were also reported by [Desaire and others (2023)](https://doi.org/10.1016/j.xcrp.2023.101426). - **Stock AI phrasing.** Replaces "profundity inflation". Stock vocabulary in two lists, taken from the AI-vocabulary lists in Wikipedia's [Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) guide before we tested them. In our tests the 2023 GPT and Claude texts leaned on the first list ("landscape", "crucial", "leverage", "seamless"), and the 2026 Claude texts had dropped most of it and leaned on the second, plain words like "quietly", "steady", "practical" and "matters". Also negative parallelism ("not just X, but Y", "Y, not X", "rather than") and sentence starters such as "Additionally" and "Moreover". AI texts failed it 15% of the time, human texts 0%. Vague intensity words such as "powerful" and "incredibly" no longer have a check of their own: the human writers in our set used them more often than the AI did. A few that are also on the stock-vocabulary list, such as "robust" and "seamless", still count there. One or two hits never fail a piece: a fail needs at least three. A fifth section, **claims to verify**, is not part of the shape score and does not come from StoryScope. Language models can produce statistics and citations that sound plausible and do not exist, so the checker lists every statistic, price, study, superlative, quote and date, and flags statistics and research claims with no source in or next to the sentence. It looks for links, "according to", named studies and publication years. It cannot tell whether a claim is true, or whether a named source says what the sentence claims. The limits of the shape checks: our AI business texts all came from Claude, and the GPT texts in the public sets are from 2023. Student essays failed the rhythm and phrasing checks more often than business writing did (28% and 14% of human essays), so the checks are tuned for business writing, not school writing. Rhythm cannot fail a piece with fewer than 20 sentences of prose. ### What it is not A sixth section, **AI-style signals**, is also outside the shape score. It gives a Low, Medium or High reading from a small logistic model over eight structural measures: parenthetical asides, names and numbers per 100 words, the share of long sentences, average sentence length, variety in section length, variety in paragraph length, colons and contractions. We trained it on 32 human-written company blog posts published before 2022, 32 human-written handbook and guide pages from before 2022, and 32 AI-written posts on matching titles from Claude Haiku and Claude Sonnet, then tested it on texts it had never seen: 16 posts on the remaining titles from Claude Opus, a model it was not trained on, 10 more Claude Sonnet posts on general business topics, and 48 human texts held back from training. High caught 18 of 26 AI-written texts and 2 of 48 human-written ones. Medium or High caught 23 of 26 AI texts but also 9 of 48 human texts, so Medium is weak evidence on its own. Stock AI phrases such as "delve" are shown for reference but do not change the reading, because current models rarely use them. Em dashes are left out because our AI samples contained none, which would have taught the model the opposite of what is true of many AI tools. The limits: small samples, one family of AI models, business writing only, punchy copy with many short sentences can read as AI-style, and there is no reading under 300 words. This is not a fact-checker, and the AI-style signals are not proof of authorship. It does not tell you whether a claim is true or who wrote a piece. Human writers fail these checks, and edited AI text can pass them. Treat a result as an editor's notes on structure. The test nobody can game is still a cold reader: where did they get bored, and did anything ring false? ### Privacy All analysis runs in your browser. Text you paste or files you upload are never sent to our server or anyone else's. Cite this tool Content Shape Checker. AI Visibility Census, 2026. https://aivisibilitycensus.com/content-shape-checker/ Free to link and reference. Census data is published under CC BY 4.0. ## Check one draft before you publish it Free, no signup, and your text never leaves your browser. [Check your content](#checker)