OpenAI’s new AI classifier tool will detect if the text was written by a human or AI

OpenAI's AI classifier is in its early stages and aims to distinguish text that is written by using artificial intelligence. Here's how it works and how you can use it.

OpenAI, the creator of the popular chatbot ChatGTP, has released a software tool to identify text generated by AI (Express photo)

Listen to this article

00:00

1x 1.5x 1.8x

OpenAI has released a new AI classifier trained to distinguish between AI-written and human-written text. The release comes after ChatGPT, the company’s AI chatbot created to mimic human speech, went viral. Obviously, it is already being used to auto-generate homework essays and assignments. This has led many educational institutions around the world to outright ban the tool. That’s no foolproof solution, though, considering students can always choose to continue using the tool for homework. Then there are also wider concerns about copyright and plagiarism.Here’s everything you need to know about OpenAI’s latest tool to detect text written by a machine.

How does the AI classifier work?

The AI classifier is “fine-tuned on a dataset of pairs of human-written text and AI-written text on the same topic,” reads OpenAI’s press release for the same. These texts were gathered from various sources believed to be written by humans. Each text was divided into a prompt and a response.

On these prompts, OpenAI generated responses from a variety of different language models trained by the company and other organisations. Based on its training, it labels text as very unlikely, unlikely, unclear, possibly, or likely written by AI.

Story continues below this ad

How accurate is the AI classifier?

OpenAI says that for the public beta version that’s put up for all to use as a web app. The confidence threshold has been kept high, meaning that the classifier only marks text as AI-written if it is very confident. This does lead to some caveats, which OpenAI has noted.

Firstly, the classifier doesn’t work as well on texts shorter than 1,000 characters. Accuracy is better with longer texts, with reliability improving as the length of the input text increases – although even lengthier texts can be labelled incorrectly. The AI classifier can even be tricked into believing that an AI-generated text is written by a human if it’s sufficiently tweaked. The reverse also holds true because sometimes human-written text can be incorrectly labelled as AI-written.

OpenAI also notes that the classifier works best with English text, as it performs significantly worse in other languages. It’s also unreliable on code and text that’s “predictable.”

How to use the AI classifier?

You need an OpenAI account to use the AI classifier, so if you haven’t already created one, head to its official website for the same. The tool can be accessed directly via platform.openai.com/ai-text-classifier. Simply paste your text into the text field on the page and hit submit. The tool works fast and will tell you in no time if the text is AI-generated or not. Just keep the following bits in mind before using the tool:

Story continues below this ad

Make sure that the text you’re pasting has a minimum of 1,000 characters, which is approximately 150-250 words
Be more careful with non-English text as the classifier’s accuracy falls with those
The classifier is likely to get things wrong on text written by children or code

Is the AI classifier the first of its kind?

No, a computer science student at Princeton previously created an app called GPTZero that can help detect whether a text was written by a human or using ChatGPT, driven by plagiarism concerns. GPTZero can be accessed via gptzero.me. It seems to work a little differently compared to OpenAI’s tool in that it focuses on the perplexity of the text by measuring randomness. Human-generated content is assumed to be more random. Like the OpenAI classifier, GPTZero is also in its early stages, so its output should be taken with a pinch of salt.

Tags:
ChatGPT