How to Tell If a Blog Post Was Written by AI (2026)
No single test is proof. Use a detector with a tested low false positive rate, check Google Docs version history, then verify every number in the draft.

You can't prove a blog post was written by AI from the text alone, but you can get close with three checks: run it through a detector that has a published, independently tested false positive rate, ask for the Google Docs version history, and verify every number and quote in the draft. In Epoch AI's July 2026 test of 495 human-written passages, Pangram and GPTZero wrongly flagged none of them, while Originality.ai flagged 19 (3.8%). The same test found all three tools missed 10% to 18% of AI text when the model was told to imitate a real writer's style.
This guide is for the person paying for the writing: a founder, an owner or a marketing lead who hired a freelancer or an agency and wants to know what they're getting. It covers what the detectors actually catch, what they cost as of 14 November 2026, and the checks that work when a detector can't settle it.
Disclosure: EchoPulse Media sells blog writing and sells no detector, so we have a stake in writing standards and none in which tool you pick.
Can an AI detector tell if a blog post was written by AI?
Mostly yes for lightly edited AI text, and much less reliably once someone has worked to hide it. A false positive is a human-written post the tool labels as AI. A false negative is an AI-written post the tool labels as human. For a buyer, the false positive is the expensive mistake, because it means accusing a writer who did the work.
Here are the most recent independent numbers we could find. Epoch AI, a research group that doesn't sell a detector, published them on 15 July 2026. It tested 495 human passages of about 500 words (from sources dated before 2022) and 594 AI passages written by Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro.
| Detector (version tested) | Human text wrongly flagged as AI | Plain AI text missed | AI text imitating a writer's style missed | Style-imitation misses on blog posts only |
|---|---|---|---|---|
| Pangram (v3.3.2) | 0 of 495 (0.0%) | 0 of 297 (0.0%) | 30 of 297 (10.1%) | 4 of 99 (4%) |
| GPTZero (2026-05-11 base) | 0 of 495 (0.0%) | 2 of 297 (0.7%) | 32 of 297 (10.8%) | 6 of 99 (6%) |
| Originality.ai (Turbo 3.0.2) | 19 of 495 (3.8%) | 1 of 297 (0.3%) | 53 of 297 (17.8%) | 19 of 99 (19%) |
Source: Epoch AI, AI detectors false negatives, 15 July 2026. Epoch counted only a clean "AI" verdict as a detection, so a "Mixed" result counted as a miss.
Three things in that table matter for a blog buyer.
Plain AI drafts are easy to catch
If a writer pastes a prompt into a chatbot and sends you the output, all three tools caught it more than 99% of the time in Epoch's test. That's the case most buyers worry about, and it's the one the tools handle well.
Style imitation gets through about one time in ten
When the model was shown five real passages by the author and asked to write like them, every tool missed more AI text. On blog-style writing specifically, Pangram missed 4% and Originality.ai missed 19%. A writer who feeds your old posts into a model is doing exactly this.
Nobody tested edited or "humanized" text
Epoch says it did not test paraphrasing tools, humanizer tools or human editing. The best evidence on humanizers is older. In a University of Chicago Becker Friedman Institute working paper summarised on 6 October 2025, Brian Jabarian and Alex Imas tested 1,992 passages and found GPTZero's false negative rate rose to around 50% or more across most genres after the text went through the StealthGPT humanizer, while Pangram's stayed low. The same paper found the open-source RoBERTa detector flagged roughly 30% to 78% of human text as AI.
Why shouldn't I trust a single detector score?
Because the error rates above are averages from one test, and your writer is one person. Older studies show how badly the tools can fail on specific groups.
Debora Weber-Wulff and colleagues tested 14 detection tools, including Turnitin, and published in the International Journal for Educational Integrity in 2023 (preprint revised 10 July 2023). Their conclusion was that the tools were "neither accurate nor reliable" and that obfuscation made them worse.
Weixin Liang and colleagues at Stanford ran 91 TOEFL essays by non-native English writers through seven detectors, published in Patterns in July 2023. The detectors labeled more than half of those essays as AI, and one tool flagged nearly 98% of them, while more than 90% of essays by US eighth graders were correctly read as human (Cell Press release, 10 July 2023). The tools tested then are not today's versions, but the lesson holds: plain, simple English can look machine-written to a model.
If your writer works in a second language, a high AI score is weak evidence on its own.
What do AI detectors cost in 2026?
We read these off each vendor's own pricing page on 14 November 2026.
| Tool | Free option | Entry paid plan | Words included | Next plan up |
|---|---|---|---|---|
| Pangram | 7-day trial | $20/month | 300,000 words a month | Professional, $65/month, 5x the allowance |
| Originality.ai | 100 credits once (about 10,000 words) | Pro, $14.95/month, or $12.95/month billed yearly | 2,000 credits (200,000 words) a month | Enterprise, $179/month, or $136.58/month billed yearly |
| GPTZero | "Get started for free" button | Not published in readable form | Not published in readable form | Not published in readable form |
GPTZero's pricing page loaded without plan prices when we checked on 14 November 2026, so we've left it blank rather than copy a figure from a review site. Check gptzero.me/pricing yourself before you buy.
For scale: a 1,500 word post is about 15 Originality.ai credits. Checking eight posts a month uses about 12,000 words, which fits inside any paid plan above.
How else can I tell if my writer used AI?
These checks don't depend on a detector, and they tell you more about whether the post is any good.
Ask for the Google Docs version history
A post written by a person in Google Docs builds up over hours or days: a messy outline, paragraphs that appear and get rewritten, a pass of edits. A post that arrives in one or two large pastes didn't happen in that document. Google's help centre says you need edit access to see a file's version history, so ask for edit access to the working doc, not a view link or a PDF. A single paste is not proof of AI, since some writers draft elsewhere, but it's a fair thing to ask about.
Check every number and quote
This is the check we'd run first. Open every source the post cites and find the number in it. AI drafts often carry statistics with no source, a source that doesn't contain the number, or a figure that was true in 2022. A human writer can get this wrong too, and either way you've found the problem that will hurt you with readers.
Look for first-hand detail
Ask what the writer knows that the top five Google results don't. A real price they checked, a screenshot they took, a customer quote you gave them, a correction of a number everyone repeats. Google's February 2023 guidance points creators to its "Who, How, and Why" questions about how a page was made. A post with none of that reads as generic whatever produced it.
Ask directly
Put the question in writing before work starts: what tools do you use, and for which steps? Many good writers use AI for outlines, research summaries or a grammar pass. What you want to rule out is undisclosed, unedited output sold at a human-writing rate.
Will Google penalize my site for AI-written posts?
Not for using AI. Google said on 8 February 2023 that "appropriate use of AI or automation is not against our guidelines" and that it rewards content on quality, not on how it was produced. What its spam policies do target is "scaled content abuse": generating many pages "without adding value for users", with generative AI named as one way to do it.
That's the real risk for a buyer. Ten thin, unsourced posts a month can do more harm than none. Google began rolling out its November 2026 core update on 10 November 2026, saying it may take up to two weeks to complete, per Search Engine Roundtable's report the same day. Updates like that are when thin content tends to lose ground. Our post Does AI Content Rank in 2026? The Studies Disagree goes through the ranking studies in detail.
What should a blog writing contract say about AI?
Put it in writing before the first post, not after a detector flags one. A short clause covers most of it.
| Clause | Example wording |
|---|---|
| Disclosure | "Writer will state which AI tools were used on each post and for which steps." |
| Sources | "Every statistic links to its original source and is checked against it before delivery." |
| Working file | "Drafts are written in a shared Google Doc with edit access for the client." |
| Detector check | "Client may run a detector. A high score triggers a conversation and a review of the version history, not automatic rejection." |
| Remedy | "Any post that fails the source check is rewritten at no charge." |
The fourth row protects good writers from a false positive and protects you from paying for a pasted draft.
If you're comparing what agencies charge for this standard of writing, Content Agency Cost Per Month: $2,000 to $15,000 in 2026 lists published prices, and What Is a Done-For-You Content Agency? A Founder's Guide explains what you should get for the money. Our own ladder is on the pricing page: a free sample, then a $299 14-day Pilot that includes one sourced 1,500 word blog, a $399/month Presence Engine, and a $1,997/month Growth plan that includes four sourced long-form posts a month, all month to month. The blog production service page sets out how those posts are researched.
When does this not apply?
- You asked for AI drafts. If your agreement allows AI drafting with human editing, a detector score tells you nothing useful. Judge the post on sources and first-hand detail instead.
- Short copy. Product descriptions, meta descriptions and social captions are far shorter than the 500 word passages Epoch tested, so the error rates above don't tell you how a tool does on them.
- Heavily edited text. If a human rewrote an AI draft line by line, no tool in the tests above was shown to tell reliably, because none of them tested human editing.
- Non-native writers. As the Stanford study showed, simple plain English can score as AI. Use version history and a conversation, not a score.
- Academic or legal decisions. This guide is for buying marketing content. It's not a basis for disciplining a student or an employee.
FAQ
What is the most accurate AI detector in 2026?
In Epoch AI's July 2026 test, Pangram and GPTZero both had zero false positives on 495 human passages, and Pangram missed the fewest style-imitating AI blog posts (4%). In the Becker Friedman Institute paper from 2025, Pangram also held up best against a humanizer tool. No tool was tested on text a human had edited by hand.
Can a detector prove my writer used ChatGPT?
No. A detector gives a probability from patterns in the text, not a record of how it was made. Treat a high score as a reason to ask for the version history and talk to the writer.
Is AI-assisted writing bad for SEO?
Not by itself. Google said in February 2023 that appropriate use of AI is not against its guidelines. Its spam policies target many pages generated "without adding value for users", so the risk is thin, unsourced posts at volume.
How much does it cost to check my blog posts for AI?
As of 14 November 2026, Pangram starts at $20 a month for 300,000 words and Originality.ai Pro is $14.95 a month for 200,000 words. Eight 1,500 word posts a month is about 12,000 words, well inside either plan.
Should I fire a writer whose post scored 100% AI?
Not on the score alone. Ask for the Google Docs version history, check the sources, and ask how they worked. If the history shows one paste, the sources don't hold up and they can't explain their process, you have your answer.
Sources
- Epoch AI, AI detectors false negatives, 15 July 2026.
- Jabarian, B. and Imas, A., Artificial Writing and Automated Detection, Becker Friedman Institute research brief, 6 October 2025 (BFI Working Paper 2025-116).
- Weber-Wulff, D. et al., Testing of detection tools for AI-generated text, International Journal for Educational Integrity 19, 26 (2023); preprint revised 10 July 2023.
- Liang, W. et al., GPT detectors are biased against non-native English writers, Patterns, July 2023; Cell Press release via EurekAlert, 10 July 2023.
- Google Search Central, Google Search's guidance about AI-generated content, 8 February 2023.
- Google Search Central, Spam policies for Google web search, read 14 November 2026.
- Google Docs Editors Help, Find what's changed in a file, read 14 November 2026.
- Search Engine Roundtable, Google November 2026 core update, 10 November 2026.
- Pricing pages read 14 November 2026: Pangram, Originality.ai, GPTZero.
Want the first three fixes for your own site?
15 minutes with Lakshya, the founder. He looks at your site and tells you the first three things he would fix. No pitch.
Book a free 15-minute callSee what this did for MagicBNBGet the weekly what's-working note
One short email a week: what is working in content for small businesses right now. No spam, unsubscribe any time.


