PhD in natural language processing, with years spent building NLP applications end to end. Moe works on text analysis: lexical and syntactic structure, and what separates machine-generated prose from human prose statistically. He has been experimenting with computational linguistics since the early days of NLTK, spaCy and WordNet, and still writes most of his tooling in Python.
Job applications are the most entity-dense documents most people write, so the humanizer that edits them has to keep job titles, company names, skills, and dates verbatim. Why TextPulse is the recommendation, plus a five-step checklist for testing any alternative.
AI-drafted emails read as templates, and templates get skimmed. A type-by-type guide to humanizing outreach, internal updates, support replies, and follow-ups while ticket numbers, figures, and product terms stay exactly as written.
Compilatio is the detector European universities license where American roundups assume Turnitin: French-built, used by more than 1,100 schools across over 50 countries, and the only major institutional detector a student can run on their own work before submitting. It claims a 94 to 99% reliability rate at catching AI text and under 1% false positives. The one peer-reviewed study to include it ranked it second out of 14 tools, and concluded the tools it tested were neither accurate nor reliable. Here is what the evidence supports.
Sapling's AI detector comes from an NLP tooling company, not an integrity company, and it shows: a public API, per-sentence perplexity scores and a Chrome extension built for triage work. Its independent record is the widest we have reviewed, from zero false positives in Scribbr's test to a 90 percent false positive rate in a 2025 Texas A&M study. That volatility, not any single score, is the finding.
Winston AI advertises 99.98% accuracy, the highest self-claimed figure in the detector industry, measured on an internal test set the company has never published. The peer-reviewed RAID benchmark put its overall accuracy at 71% at a fixed 5% false positive rate, still second among the four commercial detectors tested. Here is what each number actually measures, and who this tool genuinely suits.
Scribbr's free AI checker posts the best free-tier numbers around: 78 percent accuracy, tied for first, in Scribbr's own published comparison, with the premium tier at 84 percent. The catch is the word "own": the test was designed, run, and graded by the vendor, and the other top free tool belongs to the same parent company. Here is what those numbers rest on, what the premium tier adds, and why a Scribbr pre-check cannot forecast a Turnitin result.
Copyleaks is the AI detector most students meet without choosing it, bundled with plagiarism checking inside Canvas, Moodle, and Blackboard. The vendor advertises over 99 percent accuracy and a .03 percent false positive rate. In the June 2026 Vrije Universiteit Brussel study, it failed to fully flag a single one of 40 entirely AI-generated papers. Here is what the evidence actually shows.
Pangram publishes a false positive rate of roughly 1 in 10,000, and unlike most detector vendors it now has recent independent research pointing in the same direction. A peer-reviewed 2026 study from Vrije Universiteit Brussel found it was the only tool of four that produced satisfactory results on master's level papers. Here is what the studies measured, where the evidence is still thin, and what no detector score can prove.
The step that decides everything happens before any rewriting tool opens: reading what your own module actually permits, which is almost never set at university level. Then comes the editing method for an essay specifically, which is a different document from a journal manuscript and rewards a different order of work.
A self-diagnostic for writers who did the work themselves and now cannot hear their own prose. Why competent, taught, heavily revised academic writing sets off the same signals as machine output, which five things are worth checking, and which of them to stop apologising for.
A working definition of AI slop, the three conditions that have to hold before the word applies, where the term came from, and the four symptoms that make a paragraph read as slop, each shown next to what an editor does with it.
Article choice is one of the last things a fluent second-language writer masters, because English requires a determiner in places many languages leave the noun to speak for itself. Here is the decision path: countable versus uncountable, first mention versus later mention, generic reference, and the fixed cases that follow no rule.
An AI detection appeal letter only works if it reads as an argument rather than a protest. Here is a full template you can adapt, the evidence checklist that goes with it, and the single most common reason appeals fail.
ZeroGPT's own FAQ claims a rate pushing toward 98 percent on internal testing, and separately admits that formulaic writing can read as AI regardless of who wrote it. Here is what the company states about its method, what independent testing on this exact product has found, and what a free score is actually useful for.
A practical walk through what AI use is almost always fine, what is almost never fine, the grey area in between, and how to keep a simple record of what you did, starting from the one policy that actually governs an assignment: the course's own.
Turnitin publishes two false positive figures, not one, and neither describes the paper sitting in front of one particular professor. Here is the document-level versus sentence-level split, the sub-20-percent asterisk, and the arithmetic that turns a fraction of a percent into a real count of students.
Nursing, medicine and allied health programs run academic misconduct and fitness-to-practise as two separate systems, so an integrity finding does not always stop at a grade. Here is how that two-track structure works, and why reflective writing and care plans are exactly where it gets tested.
Vanderbilt, Michigan State and UT Austin have each published their own reasoning for switching off Turnitin's AI writing indicator. None of them dropped Turnitin itself. Here is the distinction, the arithmetic behind it, and which other names get repeated without an institution's own page to confirm them.
Common App calls AI-generated essay content fraud, UCAS requires a declaration and runs its own similarity checks, and Brown allows AI for spelling and grammar only. Here is what admissions offices have actually published, and why a personal statement is judged on voice long before any scanner gets involved.
There is no single fact that proves how an essay was written. What exists instead is corroboration: independent, hard-to-fake traces that converge on the same account. Here is what that evidence actually looks like, and what to start doing now if you have not been accused of anything yet.
A classifier trained to react to statistical pattern does not need to have seen a specific chatbot to flag its output. Here is why that matters for DeepSeek, what Turnitin itself publishes about Claude, Gemini and other models by name, and what it does not publish for any of them.
An academic integrity email naming AI use is frightening precisely because there is no single fact that disproves a percentage. Here is what to do, in order, in the first 48 hours: what to say, what never to admit, what evidence to request, and how to find out whether a score is even allowed to stand alone.
Yes, Turnitin detects paraphrased text, and it was built to. Its similarity engine catches reworded passages that still overlap with published and previously submitted work, and its AI writing report carries a labelled category for AI-generated text revised with a paraphrasing tool or word spinner. Here is how each system catches it, and why swapping words for synonyms clears neither one.
Stilted non-native academic prose is rarely a grammar problem. It is collocation, hedging pitched at the wrong strength, sentence length that never varies, and nouns doing a verb's job. Here are concrete substitutions and a self-editing routine that catches all five without hiring an editor.
Turnitin's AI writing detector does not read a document as a whole. It segments a submission into chunks, scores each one, averages the results, and only then produces the percentage on a report. This covers that pipeline, what it flags across ChatGPT, Claude and Gemini, what independent testing found it misses, and who actually gets to see the number.
General-purpose grammar tools flag passive voice a methods section needs, hedging language a reviewer expects, and readability scores an argument should not chase. Here is why, plus what Trinka, Paperpal, Writefull, ProWritingAid and LanguageTool are each actually built for, priced from their own sites.
A thesis produces three different kinds of error, and reading it once, start to finish, is the method most likely to miss all three. Here is why proofreading your own work defeats normal reading, what each of three separate passes actually catches, and the tricks that make familiar sentences look unfamiliar enough to see.
Six grammar errors show up in research papers more than anywhere else, because academic sentences are built to hold more qualification than an everyday paragraph can carry. Here is the mechanism behind each one, a wrong-and-fixed example, and what a free grammar checker can and cannot catch on its own.
Oxford University Press spells organize with a z in British English, on etymological grounds, which is why the convention is called Oxford spelling rather than an Americanism. A closed set of verbs, advertise and surprise among them, never takes -ize under any style, for a completely different reason worth knowing before a proofread turns a correct spelling into an error.
Colour and organize and analyse can each be correct and still be wrong together in the same document. Four recurring patterns explain most of what separates British and American spelling, a short list of words breaks the patterns outright, and one rule decides whether a submission passes: consistency with whatever convention the target actually specifies.
Turnitin publishes no acceptable similarity percentage, and institutions set their own guidance instead. This post explains what actually drives a high or low score, why quoted material and reference lists inflate it without anything being wrong, and how to read the match breakdown behind the number rather than the number itself.
A spell checker reads letters, not meaning, so its and it's, a plural and a possessive, and a decade with a stray apostrophe all sail through clean. Here are the rules that catch what spell check cannot: singular possessives ending in s, plural possessives, joint ownership, and the marks that should never appear at all.
A similarity score and an AI score answer different questions, so a paper can score low on one and high on the other without either number being wrong. What each measures, what triggers it, and what a high number does and does not prove.
A semicolon joins two independent clauses of equal weight. A colon announces that what follows explains or completes what came before. Here is the structural difference, where each belongs in academic prose, and the one style question that splits APA from Chicago.
Raising perplexity and burstiness on purpose usually just means writing better: varying sentence length, mixing clause structures, and choosing the word that actually fits instead of the one that merely scans. Concrete before-and-after examples, plus where this stops being craft and starts being a bad idea.
APA requires the serial comma without exception, and so do Chicago and MLA. Here is what each style guide actually mandates, why the comma exists structurally, and the cases where leaving it out changes what a sentence means.
GPTZero made perplexity and burstiness famous, then quietly retired both. Here is what its classifier measures now, what a report actually shows, and where the tool says its own limits are.
APA says spell it out. Chicago says it depends on the discipline. MLA says nothing is wrong with it at all. What three major style guides actually say about contractions, and why the real fix for stiff prose is not adding them back in.
Arabic academic writing has its own AI tells: calque openers, stacked hedges, and a flattened rhythm that Modern Standard Arabic does not normally use. Here is how those tells show up, how universities across the Arab world are responding, and what a humanizer needs to preserve when the source language is Arabic.
TextPulse runs as a single engine across more than 60 languages, but detector adoption, register norms, and citation habits differ from country to country. This hub explains why language-specific humanization matters and links to in-depth guides for 15 languages, from Indonesian to Hindi.
Vietnamese formal writing has its own AI tells: a fixed rotation of Sino-Vietnamese transition openers that ChatGPT reaches for far more often than a person would. This guide breaks down what flags Vietnamese AI text, how Vietnamese universities handle detection, and what a humanizer has to preserve to keep citations and terminology intact.
Academic Spanish has its own AI tells: calqued phrases, stacked hedges, and a flatter rhythm than a trained reader expects. This post covers what triggers a flag, what universities across Spain and Latin America currently say about disclosure, and what a humanizer needs to preserve to survive review.
Japanese academic writing runs on a single consistent register, and most AI-generated drafts break it within a paragraph. This post walks through the specific grammatical tells, how Japanese universities and detectors actually check for them, and what a Japanese AI humanizer needs to preserve along the way.
Most Hindi academic writing in India is actually produced in English, but Hindi-medium scholarship still faces its own AI-detection risks. This guide breaks down the calques and stiff openers that give away ChatGPT-drafted Hindi, what Indian and Nepali universities currently say about generative AI, and how a humanizer needs to preserve citations, terms, and register.
Thirty minutes, five phases, one rule: fix the parts worth the most marks while the attention is still fresh, and save commas for last.
Perplexity gets the attention, but the arithmetic underneath it is token probability: what the model's distribution over the next word actually contains, why detection math runs in log space instead of raw probability, and how a string of per-token scores becomes one number.
Presentation is what students polish the night before a deadline. A published marking rubric says it is usually the narrowest of the categories that decide a grade. Here is where the weight actually sits, and why.
GPTZero made perplexity and burstiness famous, then quietly retired both in autumn 2023 for a deep-learning classifier. This post traces what changed, what GPTZero and Turnitin actually document today, and why the two statistics still explain why machine text is detectable even though neither vendor computes them directly anymore.
A comma splice is two independent clauses joined by a comma alone. Here is what that comma cannot structurally do, the four repairs that fix it, and why conjunctive adverbs cause most academic splices.
Getting caught is rarely a single moment triggered by a percentage score. It is a defined process with stages, evidence standards and appeal rights, most of which never make it past a first conversation. Here is what that process actually looks like end to end, using a published university procedure rather than a guess.
Detection software matters less here than most students assume. This walks through what a professor who has already read your work actually notices when a draft stops sounding like you, from a landmark blind study of examiners to the specific, checkable signs that show up long before anyone runs a scan.
A 2023 Stanford study found AI detectors misreading fluent non-native English writing as machine-generated at a much higher rate than native writing. Here is the mechanism behind that gap, and what happened when the same essays were rewritten with richer vocabulary.
A reviewer's judgment of your argument is less separate from your comma splices and apostrophes than you would like. Here is what controlled research says mechanics actually costs a paper, and the specific rules worth getting right before anyone else reads it.
Burstiness measures how widely sentence lengths swing across a passage, not how long the average sentence runs. This post explains where the term comes from, how the spread is actually calculated, and why a language model's output tends to settle into a narrow band of lengths.
Perplexity is the number a language model produces to describe how surprised it was by your word choices. This post traces the metric to its 1977 origin, works through the formula, and explains why the same passage can score differently depending on which model is doing the reading.
Vendors publish false positive rates under 1 percent. Independent researchers testing the same detectors on non-native English writing found rates above 60 percent. Here is what the evidence shows.
AI detectors do not recognise machine writing the way a reader does. They measure how predictable your text is, then turn that into a score. Once that clicks, both the technology and its failures make more sense.