Under the Hood

It’s Not Magic. It’s Mathematics.

We strip away the marketing hype to explain exactly how our NLP engine detects ChatGPT, Claude, and Gemini — and why our approach is different.

The 4-Step Analysis Pipeline
1

Tokenization & Vectorization

When you paste text into our free scanner, the engine breaks it into “tokens” (sub-words). These are converted into mathematical vectors that represent semantic relationships between words.

2

Perplexity Measurement

We test how “surprised” a language model is by your word choices. If the model can predict every next word you write, your Perplexity score drops — a strong signal of AI-generated content.

3

Burstiness Evaluation

Humans are chaotic writers. We naturally vary sentence length — some short, some long, some interrupted. AI tends to be rhythmically monotonous. We measure this variation, known as Burstiness.

4

Final Probability Calculation

We combine Perplexity and Burstiness scores with additional semantic markers to produce a final confidence percentage — for example, “95% Human-written.”

Why AI Writes the Way It Does

To understand how we detect AI, you must first understand how AI writes. Large Language Models like GPT are essentially advanced auto-complete engines — trained to select the most statistically probable next word at each step.

Because they always aim for the “average” or “safest” word choice, their output tends to be smooth, grammatically perfect, and predictably structured. This creates a specific mathematical signature that our detector looks for.

✦ Human writing

“The pizza was absolutely bomb. Like, I’d drive an hour for this.”

◈ AI writing

“The pizza was delicious. It had an excellent flavor profile and crispy crust.”

Why this matters: The word “bomb” in that context has high perplexity — a language model wouldn’t predict it. “Delicious” has low perplexity — it’s exactly what a model would choose. This difference is measurable.
How We Handle False Positives

No detector is 100% perfect. Sometimes a human writes in a very formal, structured way — especially in academic or legal contexts — which can resemble AI output to a naive algorithm.

Unlike aggressive tools that prioritize catching AI at all costs, our algorithm is tuned for High Specificity. We would rather let an AI-written text pass undetected than falsely accuse a human writer. This design decision reflects a core belief: the cost of a false accusation is higher than the cost of a missed detection.

If your text comes back as “possibly AI” but you wrote it yourself, try adding more personal examples, contractions, or varied sentence rhythm — the hallmarks of human writing.
Privacy Architecture: Why We’re Different

Most commercial detectors — including paid Winston AI and Turnitin — require account creation and cloud uploads. Your document is stored on their servers, creating a permanent record of your work.

🔒

Client-Side Processing

Our analysis runs entirely in your device’s RAM using JavaScript and WebAssembly. The moment you close the tab, the data vanishes. We literally cannot see or sell your text — because we never receive it.

🚫

No Account Required

No sign-up, no email, no tracking. Paste your text, get your result, leave. That’s the entire flow.

Technical FAQ
Can this detect the latest GPT models?

Yes. While the newest GPT models produce more human-like output than earlier versions, they still follow probabilistic patterns our model is trained to recognise. The detector has been updated to account for the smoother syntax of these advanced models.

Does it work on languages other than English?

Currently, our algorithm is optimised for English. Analysing Perplexity in other languages requires different training datasets — this is something we’re actively developing.

Why do I get different results on different detectors?

Each tool uses different training data, temperature settings, and weighting for Perplexity vs Burstiness. There is no single “ground truth” for AI detection — results will vary between tools, especially for borderline cases.

What’s the minimum text length for an accurate result?

We recommend at least 150–200 words. Shorter samples don’t provide enough data to calculate meaningful Perplexity or Burstiness scores, which increases the margin of error significantly.

See the Technology in Action

Test your text against our NLP engine — no account, no upload, no storage. Results in seconds.

Run Free Analysis →