Was This Written by ChatGPT? AI Text Detector Guide
ChatGPT and Claude are everywhere. Learn how AI text detectors work, why they fail, and how to run a free, private analysis in your browser.
The debate over AI in education and publishing is raging. But before we can decide *what* to do about AI-generated text, we have to solve a harder problem: how do we actually prove that a human did not write something?
Human intuition is terrible at this. We often mistake dense, academic human writing for AI, and we mistake casual, error-prone AI writing for humans. To get an objective answer, we have to look at the math behind the words.
How AI text detection actually works
AI language models (like ChatGPT) are essentially giant probability engines. They generate text by predicting the next most likely word. This creates two specific mathematical signatures:
- Perplexity. This measures how "surprised" a model is by the text. Because AI generates predictable, statistically likely text, AI writing has very low perplexity. Human writing jumps around, uses unusual words, and has high perplexity.
- Burstiness. Humans write in bursts. We use a long, complex sentence, followed by a short one. AI tends to write sentences of uniform length and complexity. This lack of variation is a dead giveaway.
Why you should use a local detector
Most online AI detectors upload your text to their servers. For a student essay, this is a massive privacy violation. The text is stored, analyzed, and sometimes used to train their own models.
ToolDeck’s detector uses a RoBERTa-based neural network that downloads directly to your browser. It runs entirely on WebAssembly.
Analyze text locally. No upload, no limits.
Detect AI Text NowHow to check if text is AI
- Copy the text you want to analyze.
- Open the AI Text Detector.
- Paste the text (it works best on texts over 100 words).
- Click analyze and review the confidence score.
What to do if you find AI text
If a tool flags a text as 90% AI, it does not mean the person did not write it. It means the text shares statistical properties with AI. Always combine tool results with context. Did the student suddenly change their writing style? Does the text use a lot of "fluff" words like "delve" or "testament"?
If you are an editor, run the text through a Prompt Sanitizer to clean up invisible formatting, and then ask the author for their drafts. The process is usually more revealing than the final product.
Tools mentioned in this article
Frequently asked
Can AI text detectors be fooled?
Yes. Paraphrasing tools (like QuillBot) can take AI text and make it look human by increasing the perplexity and burstiness. However, heavy paraphrasing often makes the text sound unnatural in other ways.
Is it legal to use an AI detector on student work?
It depends on your school policy and local laws. However, because false positives are common, you should never punish a student based solely on an AI detector score. Use it as a starting point for a conversation.
Does this tool use GPTZero or Originality.ai?
No. We use an open-source, browser-based model that runs on your device. It is completely free and private, unlike those paid services.
How accurate is the AI Text Detector?
It is highly accurate on longer texts (over 150 words) generated by standard models like GPT-3.5 and GPT-4. However, it can struggle with highly technical human writing or heavily edited AI text.