Free AI Text Detector

Paste any post, comment, or article and see how likely it is to be AI-written. Runs two independent open-source detection models plus a stylometry and quote-similarity scan — built to catch the exact tricks that fool single-score tools, right in your browser.

🔒 Private — your text never leaves your device 🧠 3 checks — not just one score 🆓 Free — no account needed
📝 Text to check
0 words
💬 Post it's quoting / replying to
We'll measure text overlap between the two boxes. High overlap can mean the reply is a reworded ("spun") version of this post rather than an original response.
🤖 Detection Results
Downloading detection model…
Waiting for text…
Signal breakdown
⚠️ Runs fully in your browser via open-source detection models plus a local stylometric heuristic. Each signal below is shown on its own, not merged into one score — they can disagree, and it's up to you to weigh them. This is a best-effort estimate either way, not proof — it's most reliable on English text of 50+ words and can be wrong on very short replies, non-English text, or text that's been heavily paraphrased after being AI-written.
📊 After each check, we log anonymous aggregate stats (word count, which signals ran) to power the live usage stat above — never the text you checked. See the FAQ below for details.
Models: ModernBERT AI-detector (fast first result, trained on ChatGPT/Claude/DeepSeek + human text) + TMR AI-text-detector (RAID-benchmarked, loads after the first result) + an in-browser stylometry heuristic (sentence-length variance, vocabulary diversity, common AI transition phrases). Each one's result is shown independently below — nothing is combined into a single score. Both models are third-party open-source projects — accuracy depends on how each was trained and isn't guaranteed.
🧾 Manipulation checks
These are pattern checks (punctuation density, overlap with a quoted post), not proof of manipulation. A clean result here doesn't guarantee the text is human-written — it just means these two specific patterns weren't detected.

What it can catch

It doesn't identify which specific model wrote a text — it estimates the overall likelihood of AI involvement, and flags a couple of specific manipulation patterns common on forums:

💬

ChatGPT-style replies

The generic, evenly-structured paragraphs typical of ChatGPT and similar chat assistants.

📰

AI-written articles

Longer-form blog posts, guides, and "explainer" content produced end-to-end by an LLM.

🔡

Punctuation-stripped text

Removing commas and apostrophes is a known trick to fool single-score detectors — this tool flags it instead of being fooled by it.

🔁

Quoted-post rewriting

Replies that are really just a spun/reworded version of the post they're quoting, via a dedicated similarity check.

💭

Short forum replies

Built to score short-form text natively, without needing to pad it to a minimum character count.

✍️

Mixed human + AI edits

Text that's mostly human-written but AI-polished may show up as a partial or inconclusive signal rather than a clean pass or fail.

Why this catches what single-score detectors miss

Most free AI text checkers — the kind commonly linked in forum "AI spam" reports — give you one percentage from one model. That approach has well-documented blind spots. Here's how this tool is built differently:

CapabilityTypical single-score detectorCointikka AI Text Detector
Detection signals shownOne model, one score2 independent ML models + a stylometry heuristic, shown separately in green/red
Robust to punctuation-strippingScore can swing 50+ points from removing commasFlags abnormally low punctuation density as a manipulation signal
Catches quoted-post rewritingNot checkedDedicated quote-similarity scan (optional 2nd text box)
Handles short forum repliesSome need padding to a character minimumBuilt for short-form text natively
Shows its workOne opaque percentageFull signal breakdown with each model's raw label
Your text leaves your deviceUsually, for the ML scoringNever — models run locally via WebAssembly
Matching toolkitStandaloneSame site also has a free AI Image Detector and a Bitcointalk Rank & Merit Tracker

One documented example of the blind spot this is built to close: in community testing on a well-known crypto forum, one popular detector's score for the same post swung from over 90% "AI-likely" down to roughly 30% simply by removing the commas and apostrophes — nothing about the actual writing changed. Any detector that scores raw text fluency alone can be gamed this way; this tool treats an unusually low punctuation density in long, evenly-structured sentences as a signal in its own right, not something that silently lowers the score.

How to check if text is AI-generated

Three steps, no software to install.

1

Paste your text

Drop in a post, comment, or article — 50+ words gives the most reliable result.

2

Optionally add the quoted post

If it's a reply, paste the original post to check for rewording.

3

Analysis runs locally

Two models download once (cached after that) and run entirely inside your browser.

4

Get a full breakdown

See a plain-language verdict plus every signal that fed into it, not just one number.

Understanding AI text detection

A little context on how detection works, why one score isn't enough, and where it falls short.

How the detector works

The tool runs three independent signals and shows each one on its own, rather than blending them into a single score. Two are open-source language models — one trained on early ChatGPT-era text, one trained on the more recent RAID benchmark dataset across many modern generators — each downloaded once from a public model hub and cached by your browser afterward. The third is a stylometric heuristic that runs instantly, with no download: it looks at sentence-length variance ("burstiness"), vocabulary diversity, and the density of stock transition phrases (like "moreover" or "it's important to note") that show up disproportionately often in AI-generated prose. Each signal gets its own green (human-like) or red (AI-likely) badge so you can compare them and judge for yourself, and separate pattern checks flag things like abnormally stripped punctuation or high overlap with a quoted post.

Why combining signals matters

A single model scoring raw text fluency has a specific, well-documented weakness: it can be gamed by superficial changes — like stripping punctuation — that don't change the actual meaning of the text at all. Combining multiple, differently-trained signals and explicitly checking for known manipulation patterns makes the overall result harder to fool with any single trick, and showing the breakdown (rather than one bare percentage) lets you judge for yourself when the signals disagree.

Why local, browser-based detection matters

Every check on this page happens directly on your device using WebAssembly — the text itself is never sent to Cointikka or any third party. That matters if you're checking something sensitive: a draft you haven't published yet, a message someone sent you, or a post you're deciding whether to report. There's no server-side copy of your text and results come back in seconds. The only thing that does leave your device is an anonymous result summary — verdict, AI-probability %, word count — used solely to power the live usage stat on this page.

Limitations to keep in mind

No AI detector is perfect, and this one is no exception. It's most reliable on English prose of 50 or more words; very short replies, non-English text, and text that's been substantially paraphrased after being AI-written are all harder to score confidently, and the tool will show an inconclusive result rather than a forced verdict when its signals disagree. It also can't tell you which specific AI model produced a piece of text, only how likely AI involvement is overall. Treat the result as a helpful signal to combine with your own judgment and context, not as definitive proof either way.

Frequently asked questions

Is this AI text detector really free?
+
Yes. Cointikka's AI Text Detector is completely free with no sign-up, no watermark, and no limit on how many checks you run.
Is my text uploaded to a server?
+
No. Both detection models run entirely inside your browser using WebAssembly, and the stylometry and quote-similarity checks are plain JavaScript running locally. Your text is never sent to Cointikka or anyone else. After each check, we do log an anonymous result summary — the verdict, the AI-probability %, the word count, and which flags fired — to power the live "texts checked" stat on this page. That summary never includes the text itself, and isn't linked to you or your IP address.
How accurate is the AI text detector?
+
It gives a strong best-effort estimate, not a certainty — and it's shown as three separate signals rather than one blended score. Two independently-trained open-source models plus a stylometric heuristic each report their own read, in green (human-like) or red (AI-likely), and any signal that's genuinely borderline is shown as "inconclusive" rather than forced either way. It's most reliable on English text of 50+ words, and less reliable on very short replies, non-English text, or text that's been heavily paraphrased after being AI-written. Treat every result as one signal among others, not definitive proof.
What makes this different from tools like GPTZero or Copyleaks?
+
Most popular AI text detectors give a single percentage from a single model. This tool combines two independent models with a stylometry scan, is designed to not be thrown off by simple punctuation removal, and includes a dedicated quote-similarity check to catch replies that are just a reworded version of the post they're quoting — a pattern single-score detectors don't check for at all. It also runs fully in your browser instead of uploading your text to a server.
Can it tell me which AI wrote the text?
+
No. It estimates the overall likelihood that a piece of text is AI-generated, but it doesn't fingerprint which specific model (ChatGPT, Claude, Gemini, etc.) produced it.
Does it work on short forum replies?
+
Yes, it's built to handle short-form text without needing to pad it to a minimum length. That said, very short text (under roughly 15-20 words) is inherently harder for any detector — human or automated — to score reliably, so treat short-text results with extra caution.
What is the quote-similarity check for?
+
If you paste in the post someone was replying to, the tool measures how much word-level overlap their reply has with it. A very high overlap can indicate the reply is a rewritten/spun version of the original post rather than an independent response — a common pattern in low-effort AI-assisted forum posting that single-score AI detectors don't check for.

🎨 Appearance Settings

Choose how the app looks. Your preference is saved automatically and will be remembered the next time you open this tool.

🖌️ Theme
Select a colour scheme that suits your environment. Bitcoin orange branding is preserved across all themes.
Bitcoin Dark
Deep navy blacks with Bitcoin orange accents. The original tracker aesthetic.
Default
Soft Dark
Warmer grey tones that reduce eye strain in low-light environments.
Light
Clean, bright interface for well-lit environments and daytime use.
ℹ️ About Themes
Bitcoin orange is always preserved. All three themes keep the Bitcoin orange accent colour (#F7931A) to maintain the tool's identity regardless of the background scheme.
Preference is saved locally. Your chosen theme is stored in your browser and will persist between sessions.