OpenAI discontinues its AI writing detector due to “low rate of accuracy”

On Thursday, OpenAI quietly pulled its AI Classifier, an experimental device designed to detect AI-written textual content. The decommissioning, first noticed by Decrypt, occurred with no main fanfare and was introduced via a small notice added to OpenAI’s official AI Classifier webpage:

As of July 20, 2023, the AI classifier is now not obtainable because of its low charge of accuracy. We’re working to include suggestions and are at the moment researching more practical provenance methods for textual content, and have made a dedication to develop and deploy mechanisms that allow customers to know if audio or visible content material is AI-generated.

Launched on January 31 amid clamor from educators about college students doubtlessly utilizing ChatGPT to put in writing essays and schoolwork, OpenAI’s AI Classifier at all times felt like a performative Band-Support on a deep wound. From the start, OpenAI admitted that its AI Classifier was not “totally dependable,” accurately figuring out solely 26 p.c of AI-written textual content as “probably AI-written” and incorrectly labeling human-written works 9 p.c of the time.

As we have identified on Ars, AI writing detectors equivalent to OpenAI’s AI Classifier, Turnitin, and GPTZero simply don’t work with sufficient accuracy to depend on them for reliable outcomes. The methodology behind how they work is speculative and unproven, and the instruments are at the moment routinely used to falsely accuse college students of dishonest.

People can write like AI fashions, and AI fashions can write like people if correctly prompted. Typically, all it takes to evade AI detectors is to easily ask ChatGPT to put in writing within the fashion of a identified writer. However this hasn’t stopped a small business of economic AI detectors from sprouting up over the previous six months.

“If OpenAI cannot get its AI detection device to work, no one else can both,” tweeted AI author and futurist Daniel Jeffries. “I’ve stated earlier than that AI detection instruments are snake oil bought to individuals and that is simply additional proof that they’re. Do not belief them. They’re nonsense.”

These statements have to date been backed up by recent studies (Sadasivan et al., 2023) and testimonials from educators who usually discover that their very own human-written work is flagged as AI-composed. Moreover, AI writing detectors have been discovered to unfairly punish non-native English writers and presumably neurodivergent writers.

Analysis remains to be underway to find out if AI-generated textual content will be watermarked (by purposely manipulating the frequency of phrases in an AI-generated output), however the examine cited above reveals that textual content watermarking can simply be defeated by AI fashions that paraphrase the output.

For now, evidently AI writing is right here to remain. Going forward, AI-augmented textual content will probably circulate among the many nice works of mankind undetectably if deployed with talent. It could be time to look past how textual content consists and make sure that it correctly represents what a selected human needs to say, which is the purpose of all efficient communication.