OpenAI released an AI text classifier on 31 January 2023, designed to distinguish between human-written and AI-generated content [source]. The tool was positioned as a response to concerns about AI-generated misinformation, academic dishonesty, and automated spam.

The classifier returned one of five confidence levels: "very unlikely", "unlikely", "unclear if it is", "possibly", or "likely" AI-generated. OpenAI acknowledged significant limitations at launch. The tool correctly identified only 26% of AI-written text as "likely AI-generated", while incorrectly labelling human-written text as AI-generated 9% of the time [source].

Performance degraded further on short texts. The classifier required a minimum of 1,000 characters—roughly 150–250 words—and was unreliable below that threshold [source]. It also performed poorly on text written by children or non-native English speakers, and could be easily evaded by editing AI output [source].

OpenAI stated the classifier was trained on pairs of human-written text and AI-generated text on the same topic, sourced from 34 text-generation models [source]. The company noted it would be "significantly less reliable" on text produced by newer AI systems not included in the training data [source].

The tool was offered free of charge and did not require an OpenAI account. OpenAI described it as part of ongoing research into AI-generated content detection, while cautioning that it "should not be used as a primary decision-making tool" [source]. The company invited feedback to improve the classifier's accuracy and utility.

The classifier was later withdrawn in July 2023 due to its low accuracy rate.