How a 22‑Year‑Old Built GPTZero to Spot AI‑Written Text

A 22‑year‑old Princeton senior, Edward Tian, developed GPTZero during his winter break to help people determine whether a text was authored by a human or by OpenAI’s ChatGPT.

He uploaded the app to Streamlit, expecting only a few users, but within hours more than 2,000 people signed up, causing the hosting platform to crash under the sudden traffic.

GPTZero evaluates text using two statistical indicators: perplexity, which measures the randomness and complexity of language, and burstiness, which assesses the variability of sentence structures. Low scores on both suggest AI‑generated content, while higher scores indicate human writing.

Perplexity reflects how predictable word choices are; AI‑trained models tend to produce uniformly predictable text, whereas human writers exhibit more varied and surprising phrasing. Burstiness captures the alternation between long, complex sentences and short, simple ones, a pattern common in human prose but rare in machine‑generated output.

Tian trained GPTZero using the open‑source GPT‑2 model and acknowledges that the detector is not flawless, continuing to refine its accuracy.

He emphasizes that AI should remain transparent and ethically used, noting that while he leverages tools like GitHub Copilot for coding, he does not oppose AI itself.

GPTZero has attracted attention from educators, HR professionals, and venture capital firms such as A16Z and Menlo Ventures, and remains freely available while Tian plans further improvements.