ThePrompt
Posts
OpenAI launches app to detect AI-text 🔊

OpenAI launches app to detect AI-text 🔊

with many limitations

Anita Kirkovska
February 01, 2023

Hi folks! 👋🏻 This is The Prompt, the AI newsletter that's like a juicy burger from In-N-Out's not-so-secret menu. Hidden in plain sight, full of flavor, and always served up fresh. 🍔

Here's what we're serving today:

GPT model that detects AI-text
Misuse of voice cloning
Smart prompting
GraphGPT
SingSong & more

OpenAI launches app to detect AI-text 🔊

Yesterday, OpenAI released a fine-tuned GPT model that is trained to detect AI-generated text from 36 different models, including ChatGPT.

They do stress that their model is not reliable, has many limitations, and should not be used as a primary decision-making tool.

Its success rate is actually around 26% - but can be improved if used with other methods.

Limitations

Two of the biggest limitations are that this model is very unreliable on short texts (below 1,000 characters) & there is a high chance that human-written text can be detected as AI-generated. There are others limitations as well.

How does it work

This classifier is trained on a dataset of pairs of human-written text and AI-written text on the same topic. Collected from various other sources, but also from the data that millions of users created by chatting with ChatGPT.

Based on the confidence of the model, each text can be labeled as either very unlikely, unlikely, unclear if it is, possibly, or likely.

About human+AI written text

According to OpenAI, they have not conducted a comprehensive evaluation of the classifier's ability to identify content created with the assistance of human writers.