OpenAI releases a tool to detect AI-generated text, including from ChatGPT.
The AI-generated text detector tool by OpenAI is now available.
OpenAI has released a tool that tries to differentiate between human-written and AI-generated language, such as the material created by the company’s own ChatGPT and GPT-3 models, after teasing the move in media appearances. Although the classifier’s success rate is only about 26%, according to OpenAI, it can be helpful in preventing the misuse of AI text generators when used in conjunction with other techniques.
As enthusiasm for generative AI, especially text-generative AI, has grown, critics have called on the developers of these tools to take steps to mitigate their potentially harmful effects. Some of the largest U.S. school districts have banned ChatGPT on networks and devices due to concerns about the impact on student learning and the accuracy of the content generated by the tool. Sites such as Stack Overflow have also banned users from sharing content generated by ChatGPT, and AI has made it all too easy for users to flood her discussion threads with questionable answers.
“The classifier aims to help mitigate false claims that AI-generated text was written by a human. However, it still has a number of limitations — so it should be used as a complement to other methods of determining the source of text instead of being the primary decision-making tool,” an OpenAI spokesperson told TechCrunch via email. “We’re making this initial classifier available to get feedback on whether tools like this are useful, and we hope to share improved methods in the future.”
When determining if a certain piece of text was produced using AI, the detector offers a somewhat ambiguous response. Depending on its degree of confidence, it will identify the text as “extremely improbable”. Other are (less than 10% chance), “unlikely” (between 10% and 45% chance), “unclear whether it is” (45% to 90% chance). Another says “possible” (90-98% chance), or “likely” (above 98% probability) AI-generated.
For me, concerning this AI-generated text detector, at a practical level, I didn’t find the classifier particularly useful for evaluating short texts.
In fact, reaching 1,000 characters is a difficult threshold for messages such as e-mail (at least the ones I receive on a regular basis). And this limitation becomes material for consideration. OpenAI emphasizes that the classifier can be bypassed by changing some words or phrases in the generated text.