A college student built an app to catch ChatGPT cheaters

While high schoolers were busy plotting how to let ChatGPT do their English homework, a Princeton senior spent his winter break building the ultimate countermeasure.

22-year-old Edward Tian, majoring in computer science and minoring in journalism, coded an app called GPTZero over the holidays. Its job is simple: quickly determine whether an essay was penned by a human or ChatGPT.

Measuring the rhythm of human thought

Ever since OpenAI launched ChatGPT on November 30, 2022, educators have been on high alert. The bot racked up over a million users in just five days, demonstrating a knack for producing convincing prose. It could emulate Shakespeare or a New Yorker essayist with ease, and a Wall Street Journal reporter even used it to pass an AP English exam.

Panic followed fast. New York City public schools blocked access to ChatGPT on school networks, and top AI conferences barred fully machine-written research papers.

Tian's GPTZero detects automated writing using two core linguistic markers: perplexity and burstiness.

Perplexity tracks how random or surprising words in a sentence are to the model. Bots stick to predictable phrasing, meaning lower perplexity, while humans tend to choose words more unpredictably.

Burstiness measures variance across sentences. Human writers naturally alternate between long, tangled clauses and punchy statements. AI, by contrast, keeps a steady, uniform tempo.

In testing, a ChatGPT-crafted summary of Hamlet scored a 10 in perplexity and a 19 in burstiness, easily triggering a bot label. When tested on human journalism, burstiness spiked to 387, confirming human authorship.

Crashed servers and divided internet

It took almost no time for GPTZero to blow up. After Tian posted demos on Twitter, teachers from Switzerland to France flooded his inbox with gratitude.

Over 10,000 people rushed to try the tool on Streamlit, causing both the hosted app and gptzero.me to crash under unexpected traffic.

The reaction was fiercely divided. Venture capitalists filled Tian's DMs with funding offers, while students took to social media to brand him a "narc." Accuracy remains a work in progress, too: some users noted the detector missed certain AI snippets, though Tian rolled out an update to curb false positives.

Tian told reporters he plans to keep the tool free for now to help freshman English teachers cope.

Why it matters

GPTZero captures the frantic absurdity of this early AI wave. Frontier labs pushed hyper-capable language models out the door, leaving teachers with zero official verification tooling.

OpenAI floated the idea of cryptographic watermarking back in December, but until that arrives, a student's dorm-room hack is literally the best shield educators have. For teachers drowning in suspicious essays, it is certainly worth bookmarking as an initial sniff test.

Still, we think relying on statistical detectors like GPTZero as definitive proof of cheating carries massive risks. Metrics like perplexity can be easily defeated by a student who edits a few adjectives or tweaks prompt parameters. False accusations are inevitable, and turning grading into a forensic investigation will only erode classroom trust.

The broader takeaway is that bans and detection tools are temporary roadblocks. As Tian himself pointed out, AI writing is here to stay and will only improve. The conversation has to move from catching students in the act to redefining what original writing even means in an automated world.