humanflow

Blog

Notes on writing naturally, understanding AI detection, and keeping your own voice.

Latest · AI detection14 min read

How to test an AI detector yourself: our open replication framework

A reproducible AI detector test needs a published corpus, dated model versions, fixed scoring rules, and reruns. Here is our full open protocol.

Read the post

More posts

AI detection13 min read

AI Detector Minimum Words: Why Short Text Can't Be Reliably Detected

Most AI detectors need 100–300+ words to work — Turnitin requires ~300. Below that, scores are statistical noise. Minimums by tool and what to do instead.

Read
AI detection16 min read

Can you trust a 98% accuracy claim? How to read AI-detector marketing

"98% accurate" usually means "98% under conditions in the footnote." What accuracy claims really measure, the base-rate math, and a checklist for any vendor page.

Read
AI detection14 min read

AI Detection Thresholds: The Dial That Decides Who Gets Flagged

An AI detection threshold is a business decision trading false accusations against missed AI. How Turnitin, Originality.ai and others set the dial.

Read
AI detection14 min read

The AI detector bias study: what Liang et al. actually found, number by number

Liang et al. (Patterns, 2023) found 7 AI detectors falsely flagged non-native English TOEFL essays 61.22% of the time — near-perfect on native writers.

Read
AI detection13 min read

Why OpenAI Shut Down Its Own AI Detector

OpenAI retired its AI text classifier in July 2023 after it caught just 26% of AI text and falsely flagged 9% of human writing. The full story, dated.

Read
AI detection14 min read

AI Image Detection vs. Text Detection: Why They're Completely Different Problems

Text detectors read word-choice statistics; image detectors read pixel artifacts, frequency patterns, and C2PA metadata. Why skill in one doesn't transfer.

Read
AI detection16 min read

Are AI Detectors Reliable in 2026? What Every Published Study Actually Shows

No peer-reviewed study supports the accuracy vendors imply — the largest put all 14 tools under 80%. Every published reliability finding since 2023, reviewed.

Read
AI detection16 min read

Enterprise AI detection: the B2B layer most people never see

Enterprise AI detection means API-first tools like Sapling, Copyleaks, and Pangram — bought on volume, compliance, and audit trails, not free-scan demos.

Read
AI detection12 min read

Scribbr AI Detector Review: The Academic Tool That Publishes Its Own Report Card

Scribbr's free AI detector, reviewed: 1,200-word checks, no sign-up, and the rare vendor that publishes research ranking its own tool at 84%.

Read
AI detection12 min read

QuillBot AI Detector Review: A Good Free Tool From a Company Playing Both Sides

QuillBot sells a paraphraser and an AI detector. Our review of its free limits, 99% RAID claim, and what that dual position means for you.

Read
AI detection13 min read

Grammarly AI Detector Review: The Mainstream Brand That Refuses to Overclaim

Grammarly's AI detector is free, cautious, and refuses to overclaim. Our review of its 99% claim, Authorship provenance, and real limits.

Read
Turnitin16 min read

Reported for AI Use: What an Academic Integrity Case Actually Looks Like, Step by Step

Reported for AI use? The typical process: report, initial meeting, response window, hearing, finding, appeal — your rights and how to prepare at each stage.

Read
Turnitin15 min read

The History of Turnitin: From Berkeley Side Project to the Judge of Student Writing

Turnitin history in full — founded 1998 as iParadigms, sued by students, sold for $1.75B in 2019, pivoted to AI detection April 2023. The whole story.

Read
AI detection15 min read

AI Detectors and Translated Text: The Blind Spot Nobody Warns You About

Machine-translating human writing often trips AI detectors — one 14-tool study saw accuracy fall ~20 points on translated text. Here's why, and what to do.

Read
AI detection13 min read

Can AI Detectors Detect Edited AI Text? The Mixed-Authorship Problem Nobody Has Solved

Lightly edited AI text usually still gets flagged. Deeply rewritten text usually doesn't — because its authorship genuinely changed. Blends are where detectors fail.

Read
Turnitin15 min read

Can AI Detectors Detect Code? An Honest Answer for CS Students

Prose AI detectors don't reliably work on code — Turnitin says so itself. What CS departments actually use, and how AI-written code really gets caught.

Read
Turnitin14 min read

SafeAssign vs Turnitin: What Each One Actually Checks

SafeAssign checks similarity only — it has no AI detector. Turnitin does both. Why "my school uses SafeAssign, so AI is safe" is a risky bet.

Read
AI detection13 min read

Can AI Detectors Detect GPT-5 and the Newest Models? Yes — After a Lag, and the Lag Is the Story

Mostly yes — after a lag. Detection dipped when GPT-5 launched in Aug 2025, vendors retrained within weeks, and the gap reopened with GPT-5.5. Here's the cycle.

Read
AI detection13 min read

Can AI Detectors Detect Claude? What the Evidence Shows — and What It Doesn't

Yes. Detectors flag Claude's unedited output about as often as ChatGPT's, because they detect machine-typical statistics, not brands. Editing is what changes scores.

Read
Turnitin14 min read

Can Google Classroom detect AI? What originality reports actually check

No — Classroom's originality reports check web and student-work matches, not AI. The real signal teachers use is Docs version history. Here's how.

Read
Turnitin13 min read

Can Canvas detect ChatGPT? What the platform actually sees

No — Canvas has no AI detector. It logs quiz activity and submission data; AI flags come from Turnitin integrations. What professors really see.

Read
AI detection14 min read

AI Watermarking, Explained: SynthID and the Third Way to Spot Machine Text

AI watermarking hides a signal in text as it's generated — no guesswork needed. How SynthID works, what it covers today, and why it can't catch everything.

Read
Turnitin13 min read

How to Cite AI in an Essay: APA, MLA, and Chicago, With Real Examples

APA cites the company (OpenAI, 2023); MLA cites the prompt with no author; Chicago uses a footnote only. Exact templates and disclosure statements inside.

Read
AI detection13 min read

Burstiness in AI Detection: What the Rhythm Test Actually Measures

Burstiness is how much your sentence lengths vary. Humans write in bursts; AI keeps even rhythm. Here's how detectors use it — and where it misreads.

Read
AI detection14 min read

What Is Perplexity in AI Detection?

Perplexity measures how surprised a language model is by your next word. Low surprise reads as AI. Here's how that works — and where it breaks.

Read
Turnitin14 min read

Is Using AI for a First Draft Cheating? An Honest Answer

Using AI for a first draft is cheating only if your course policy forbids it — the rule, not the tool, defines the offense. Here's how to know where you stand.

Read
AI detection15 min read

Why Do AI Detectors Give Different Results on the Same Text?

AI detectors disagree because each uses its own threshold, training data, and segmentation. Same essay, five tools, five scores — here's why, and what to do.

Read
Turnitin16 min read

University AI Policies in 2026: From Full Bans to Required AI Courses

University AI policies in 2026 fall into four types — ban, disclose-and-cite, instructor discretion, and AI-integrated. Most default to "no permission, no AI."

Read
Turnitin16 min read

Turnitin Draft Coach: what it checks, what it doesn't, and the one thing students get wrong

Draft Coach checks similarity, citations, and grammar in Google Docs and Word — 3 similarity checks per draft. It does not show Turnitin's AI score.

Read
AI detection14 min read

The Best AI Detector in 2026 Depends on Who's Asking

There's no single best AI detector in 2026. Turnitin for institutions, Copyleaks for teams, QuillBot free for students — here's the decision framework.

Read
Turnitin13 min read

Does Turnitin Detect Humanized AI? Sometimes — and Anyone Quoting a Percentage Is Guessing

Sometimes — and anyone quoting a pass rate is guessing. Turnitin added humanizer detection in 2025. Why results vary and bypass claims can't be checked.

Read
AI detection14 min read

The Best Free AI Detectors in 2026, Ranked by What You Actually Get

Scribbr and QuillBot offer the most generous free AI checks; Copyleaks allows 25,000 characters per scan. Full free-limit table and honest rankings.

Read
AI detection12 min read

Copyleaks vs Originality.ai: Two Detectors Built for Different Jobs

Copyleaks suits institutions (LMS, compliance, $16.99+); Originality.ai suits publishers (~$0.07/1k words, site scans). Neither score is proof.

Read
Turnitin12 min read

How to Check Your Turnitin AI Score Before Submitting: The Honest Answer

You can't check Turnitin's AI score yourself — it's institution-only. Here's what actually works: directional checks, Draft Coach, and process evidence.

Read
AI detection13 min read

GPTZero vs Turnitin: consumer detector, institutional gatekeeper, and why their scores disagree

GPTZero is a consumer detector; Turnitin is institution-only with a 20% reporting threshold. Here's why they disagree and what each score means.

Read
Turnitin15 min read

Does Turnitin Save Your Papers?

Yes, by default — submissions join Turnitin's student paper repository unless your instructor chose "no repository." Your rights, retention, and deletion.

Read
Turnitin14 min read

Turnitin vs GPTZero: What Actually Separates the Two Detectors

Turnitin is institution-only; GPTZero is public. Both claim ~98-99% accuracy under narrow conditions, and they often disagree on the same text. Here's why.

Read
AI detection12 min read

GPTZero vs ZeroGPT: they're not the same tool, and the differences matter

GPTZero and ZeroGPT are unrelated products from different companies. Here's who makes each, how their claims differ, and which fits your use case.

Read
AI detection13 min read

Winston AI review: what's behind the 99.98%?

Winston AI claims 99.98% accuracy. That figure comes from its own 10,000-text self-test on older models. Full review — pricing, OCR, evidence, limits.

Read
Turnitin13 min read

How Do Professors Use Turnitin? The View From the Other Side of the Report

Professors see a similarity score, an AI percentage, and highlighted passages — then decide what to do. Here's how careful instructors handle a flag.

Read
Turnitin13 min read

Turnitin AI Score Meaning: What the Number Actually Tells You

Turnitin's AI score is the share of prose its model thinks is AI-written — a probability estimate, not proof. Here's what 0%, *, 20%, and 100% mean.

Read
AI detection14 min read

Originality.ai review: the marketer's detector, read closely

Originality.ai claims 99% accuracy for marketers at ~$14.95/month. What the claim covers, what it doesn't, and who the tool actually fits.

Read
Turnitin14 min read

What Turnitin Similarity Score Is Too High?

No universal Turnitin similarity score is "too high." Many instructors look closer above ~25%, but what's matched matters far more than the number.

Read
AI detection14 min read

Copyleaks AI Detector Review: The Education Heavyweight, Examined

Copyleaks claims 99%+ accuracy and a 0.03% false-positive rate. Strong tool, real third-party studies — but the fine print matters. Full review.

Read
AI detection13 min read

ZeroGPT Review: What the Free Detector Does Well, and Where It Breaks

ZeroGPT scans 15,000 characters free — a usable rough signal on unedited AI text, but it once flagged the Declaration of Independence as 97.93% AI.

Read
Turnitin12 min read

Does Turnitin Check PowerPoint?

Yes — Turnitin converts .pptx files to PDF and checks slide text for similarity. Speaker notes are stripped, and AI detection rarely applies. Details inside.

Read
AI detection14 min read

Is GPTZero accurate? Every published number, plus how to test it yourself

GPTZero claims 99% accuracy, and 95.7% detection at 1% false positives on RAID. Here's what those numbers mean, who measured them, and how to test it yourself.

Read
Turnitin14 min read

Can Turnitin Detect AI in a PDF?

Yes. Turnitin extracts the text from any readable PDF and runs the same AI detection as on Word files. Format changes nothing. Here's what does.

Read
Turnitin13 min read

Can Turnitin Detect Grammarly? Separating the Three Things Grammarly Does

Grammar fixes won't give you an AI score. Grammarly's rewrite and generative features can. Here's the line — and how to prove your work is yours.

Read
Turnitin14 min read

Can Turnitin Detect QuillBot? What the Evidence Actually Shows

Sometimes, yes. Turnitin added AI paraphrasing detection in July 2024, and QuillBot output carries two separate risks. Here's what the evidence shows.

Read
Turnitin13 min read

Can Turnitin Detect Paraphrasing? Three Different Questions, Three Different Answers

Sometimes. Word-swapped sources often still match; AI-paraphrased AI text is increasingly flagged; deep genuine rewriting isn't. Here's each case.

Read
Turnitin13 min read

Can Turnitin Detect Gemini? Yes — and Google Docs Adds a Second Trail

Yes — Turnitin flags unedited Gemini text like any AI output. And in Google Docs, version history leaves a second trail most students forget about.

Read
Turnitin14 min read

Can Turnitin Detect Claude? What the Evidence Actually Says

Yes. Turnitin flags unedited Claude output at rates similar to ChatGPT, because it detects the statistics of AI prose, not any specific model.

Read
Turnitin13 min read

Can Turnitin Be Wrong? Yes — In Both Directions

Yes. Turnitin's AI detector falsely flags real students and deliberately misses ~15% of AI text. Here's the evidence, and how to challenge a score.

Read
Turnitin13 min read

Turnitin False Positives: How Often Innocent Writing Gets Flagged

Turnitin claims under 1% false positives — but only above 20% AI scores. At its scale, even that rate means thousands of innocent students flagged.

Read
Turnitin16 min read

Turnitin AI detection accuracy: every published number in one place

Turnitin claims 98% accuracy and <1% false positives — only for docs over 20% flagged. Every independent result, and how to reconcile the gap.

Read
Turnitin12 min read

Does Turnitin detect AI writing? Inside the indicator, step by step

Yes — a separate AI indicator inside the Similarity Report scores your prose. How segmentation, classification, the percentage, and the asterisk work.

Read
Turnitin13 min read

Can Turnitin detect ChatGPT? What actually happens when you submit

Usually, yes. Unedited ChatGPT text gets flagged most of the time. The submission timeline, what your instructor sees, and outcomes for 3 scenarios.

Read
Turnitin13 min read

Turnitin and Non-Native English Speakers: The False-Positive Problem Nobody Warned You About

Research found AI detectors falsely flagged 61% of non-native English essays. Why ESL writing trips Turnitin-style tools, and how to protect yourself.

Read
Writing15 min read

Writing in your own voice when AI drafts it

AI drafts fast and flattens voice. Here is what voice is made of mechanically, why models erase it, and the five edits that put it back — with the before and after.

Read