How Decopy AI Works
No magic. No black boxes. Just advanced mathematical modeling and linguistic pattern recognition designed to protect human authenticity.
The Two Pillars of Detection
AI doesn’t “think” like a human; it predicts. Large Language Models (LLMs) like ChatGPT construct sentences by calculating the most mathematically probable next word. Our algorithm detects these robotic patterns by analyzing two core metrics.
1. Perplexity (Predictability)
Perplexity measures how “surprised” an AI model is by a sequence of words. AI tends to choose the most common, expected words to form a sentence.
Low Perplexity (AI): “The sky is blue.” (Very predictable)
High Perplexity (Human): “The canopy above hung like bruised sapphire.” (Unexpected, creative)
2. Burstiness (Variation)
Burstiness measures the variation in sentence length and structure throughout a document. AI models are trained to be consistently readable, resulting in uniform sentence structures.
Low Burstiness (AI): Medium sentence. Medium sentence. Medium sentence.
High Burstiness (Human): Short sentence! Then, a much longer, complex, and meandering thought processes that spans several lines. Short again.
The Analysis Pipeline
What happens in the 3 seconds after you click “Analyze”?
Tokenization & Cleaning
Your text is instantly broken down into “tokens” (words and sub-words). We strip out formatting and normalize the text so our algorithm can read the underlying structure without bias from fonts or spacing.
Cross-Model Scoring
Instead of relying on a single model, we run your tokens against an ensemble of baseline language models. We calculate the mathematical probability of every single word being generated by an LLM versus a human writer.
Sentence-Level Mapping
We aggregate the token scores into sentence-level metrics. This is how we can pinpoint exactly which sentences exhibit low perplexity and flat burstiness, allowing us to highlight specific areas of concern rather than just giving a vague overall percentage.
Zero-Retention Purge
The moment the results are rendered on your screen, the original text data and the tokenized arrays are permanently purged from our system memory (RAM). Nothing is written to a disk database.
The ESL Calibration
One of the biggest failures of early AI detectors (like GPTZero and Turnitin’s V1) was the devastating false positive rate for Non-Native English Speakers (ESL).
Because ESL students often rely on standardized vocabulary and follow strict grammatical rules to avoid mistakes, their writing naturally exhibits lower perplexity. Standard detectors look at this and falsely declare: “Robot!”
How we fixed it: We explicitly trained the Decopy AI classification model on vast datasets of human-written ESL essays (TOEFL/IELTS practice exams). Our algorithm is calibrated to distinguish between the “predictability” of a non-native speaker and the “predictability” of an LLM. This significantly reduces unfair academic flagging.
De-Biased Engine
Optimized to protect international students and neurodivergent writers from false positives.