release

New AI classifier for indicating AI-written text

Developers and solopreneurs can integrate this classifier to automatically identify AI-written content in their applications, aiding transparency and moderation, but must account for its imperfect accuracy.

OpenAI Blog·January 31, 2023·1 min readrelease

releaseNew AI classifier for indicating AI-written text

openai.com

What happened

OpenAI has launched a classifier intended to differentiate between human-written and AI-generated text, as announced on the OpenAI Blog. The model is trained to detect outputs from GPT models and similar large language models by analyzing linguistic patterns common in machine-generated content. To use it, text must be at least 1,000 characters long, and the tool is most reliable on English text. The classifier is accessible through a web interface and an API, but OpenAI notes its limitations: it may mislabel both human and AI text, especially short or creative passages, and it should not be used as the sole arbiter of authorship. For developers and solopreneurs building AI workflows, this tool offers a way to automatically flag potential AI-generated content in user submissions, chatbot logs, or publishing pipelines. However, given its constraints, it is best employed as a probabilistic indicator rather than a decisive filter.

Key takeaways

OpenAI released a classifier that predicts whether text was written by AI or human.
It works best on English text longer than 1,000 characters, especially from GPT models.
The tool is available via web app and API, but has known false positive and negative rates.
OpenAI advises using it as a supplementary signal, not a standalone detection method.