New AI classifier for indicating AI-written text
OpenAI
OpenAI built a tool that guesses whether a chunk of text came from a bot or a person. It's a shaky first swing at a problem that's about to get a lot bigger.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has released a classifier designed to flag whether a piece of writing was generated by an AI system or typed out by a human. The company trained the model on pairs of text covering the same topics and prompts, one version written by a person and the other spun up by a language model, then taught the classifier to spot patterns that separate the two.
The use case OpenAI is clearly aiming at is education. Teachers, professors, and school administrators have been scrambling since ChatGPT showed up, worried that students would quietly outsource essays to a chatbot with no way to prove it. A detector, even an imperfect one, gives institutions something to point to.
But the tool comes with real caveats baked right into its release. OpenAI itself says the classifier isn't reliable on short passages, under a thousand characters, and it can be fooled or produce false positives, meaning it sometimes labels human writing as machine-made and vice versa. Text that's been edited after generation, or written in languages other than English, trips it up even more.
That honesty matters, because the stakes for getting this wrong are high. A student wrongly accused of cheating because a probabilistic model misfired isn't a hypothetical, it's the most likely failure mode. OpenAI is framing this less as a finished product and more as an early, rough instrument in what's going to be a long-running arms race between generative models and the tools built to catch them.
My take — AI-written commentary, not fact-checked reporting
I'll say the obvious thing nobody wants to hear: a company selling the pen and the lie-detector for that pen is not a stable business model. OpenAI shipping a shaky classifier feels less like a solution and more like a liability shield, something to wave at regulators and schools while the underlying generation tech keeps getting better and harder to catch. If detection can't keep pace with generation, and right now it clearly can't, the honest answer is that we need new norms around writing, not another leaky filter.
Read more about this at: OpenAI
