A Short History of AI Text Detection

Before ChatGPT

Researchers built early tools to spot machine text from models like GPT-2 around 2019. They mostly looked at how predictable each word was, the same idea behind perplexity scoring today. Few people outside research used them.

2023: the rush

• January: GPTZero launched, built by a Princeton student, and drew millions of users within months.
• January: OpenAI released its own AI text classifier, with a published true positive rate of 26% and a false positive rate of 9%.
• Spring: Turnitin added AI writing detection to the software many schools already used for plagiarism checks.
• July: OpenAI withdrew its classifier, citing low accuracy. A Stanford study reported that detectors flagged more than half of essays by non-native English speakers as AI.
• August: Vanderbilt University disabled Turnitin’s AI detector, citing false positives, lack of transparency, and bias concerns.

2024: new methods, new questions

• Researchers published zero-shot methods such as Binoculars, which detect AI text without detector-specific training.
• Writing-process tools appeared. Grammarly Authorship, for example, records whether text was typed, pasted, or AI-assisted.

2025: better detectors, process tools go mainstream

• Independent research found the best commercial detectors had become far more accurate on unedited text, while quality varied widely between products, especially against humanizer tools.
• Turnitin launched Turnitin Clarity, a writing space that shows teachers revision timelines and pasted versus typed text.

Where things stand

Detection is better than it was, but it still estimates. The trend is toward evidence of how something was made: process records for writing, and content credentials for images and video. Detectors are becoming one signal among several rather than a verdict.