OpenAI AI Text Classifier Release (January 2023)

TL;DR

OpenAI launched a free, publicly‑available AI text classifier on January 31 2023 to help detect AI‑written English content, but the tool is limited: it correctly flags only 26% of AI‑generated text and produces a 9% false‑positive rate on human text, and it was later withdrawn due to low accuracy.

What the classifier does

The classifier is a fine‑tuned language model that distinguishes between human‑written and AI‑written passages. It was trained on paired datasets where the same prompt was answered by humans and by a variety of language models from OpenAI and other providers. For the web demo, OpenAI set a high confidence threshold to keep false positives low, labeling text as "likely AI‑written" only when the model is very certain.

Reported performance

  • True positive rate: 26% on a challenge set of English AI‑generated texts.
  • False positive rate: 9% of human‑written texts are incorrectly labeled as AI‑written.
  • Length dependence: Accuracy improves with longer inputs; the classifier is especially unreliable on texts under 1,000 characters.
  • Comparison to prior work: The new model is significantly more reliable than OpenAI’s earlier detector released with the GPT‑2 output dataset.

Key limitations

  • Not a decision‑making tool: The classifier should be used only as a supplemental signal, not as a definitive verdict.
  • Short text weakness: Inputs shorter than 1,000 characters yield highly unreliable predictions.
  • Language and domain scope: It is calibrated for English only and performs poorly on other languages and on code.
  • Predictable content: Text that is deterministic (e.g., a list of prime numbers) cannot be distinguished.
  • Adversarial editing: Human edits can evade detection; ongoing attacks may outpace retraining.
  • Calibration issues: Outside the training distribution, the model can be over‑confident in incorrect predictions.

Training data and methodology

OpenAI assembled a dataset of human‑written excerpts drawn from pre‑training corpora and human demonstrations submitted to InstructGPT. Each excerpt was split into a prompt and a response. For each prompt, responses were generated by multiple language models, including OpenAI’s own models and those from other organizations. The classifier was then fine‑tuned on these human‑AI pairs.

Intended use cases and outreach

OpenAI highlights several potential applications:

  • Detecting automated misinformation campaigns that masquerade as human discourse.
  • Identifying academic dishonesty where AI tools are used to produce essays.
  • Preventing the misrepresentation of AI chatbots as human interlocutors.

The lab also released a preliminary resource for educators on the use of ChatGPT, inviting feedback from teachers, administrators, students, and parents via a public form. OpenAI emphasizes that the classifier is part of a broader effort to develop provenance techniques for text, audio, and visual content.

Current status

As of July 20 2023, the AI text classifier was taken offline because of its low accuracy. OpenAI indicated that it is researching more effective provenance methods and remains committed to deploying tools that help users determine whether content was AI‑generated.

Authors: Jan Hendrik Kirchner, Lama Ahmad, Scott Aaronson, Jan Leike

Sources