OpenAI is testing a new tool to help distinguish between AI-written and human-written text. The “AI Text Classifier,” as the company calls it, is a fine-tuned GPT model that predicts how likely it is that a piece of text was generated by AI from a variety of sources.
While there is no foolproof method for identifying AI-generated content, OpenAI hopes its new tool will help dispel the myth that such writing (i.e. an AI-generated content) must have been done by a human.
The AI company claims that its new AI Text Classifier will make it harder to conduct automated misinformation campaigns, commit academic fraud using AI tools, and pass off as a human in online chats.
The tool achieved 26% accuracy when testing on a corpus of English texts in determining whether or not the text was created by artificial intelligence. But it also mistook human-written text for being written by AI 9 percent of the time.
OpenAI claims that the longer the text is, the better its tool performs, which may explain why it requires at least 1,000 characters to run a test.
Other limitations of the new OpenAI Text Classifier include the following:
- Can mislabel both AI-generated and human-written text.
- AI-generated text can evade the classifier with minor edits.
- Can get things wrong with text written by children and on text not in English because it was primarily trained on English content written by adults.
Having said that, let’s take a look at how well it operates.
OpenAI’s AI Text Classifier is user-friendly and easy to work with. Simply sign in, paste the text you want to check, and hit the “Submit” button. The tool will assess how likely it is that it was created by artificial means. Results can be anything from:
- Very unlikely
- Unlikely
- Unclear if it is
- Possibly
- Likely
To test text classifier, I copied an article written on Innovation Village and submitted. The detection software then described the text as “possibly AI-generated.” Conversely, it labelled ChatGPT-made content as “possibly AI-generated.” This tool may or may not be accurate, but my (very small) test suggests it has some validity. OpenAI warns against relying solely on the tool to establish content’s veracity, and emphasises that texts of 1,000 words or more are required for optimal results.
Innovation Village learnt that ChatGPT was recently banned by NYC public schools due to concerns that it facilitates cheating by allowing users to generate articles and essays that are difficult to differentiate from authentic student work. The natural-language model can write essays in seconds based on simple text prompts. These essays can even pass a graduate business and law exam, giving students a new way to cheat.
The launch of ChatGPT has generated considerable interest recently, and this has spread to the media as a whole as publications look for ways to automate the production of SEO-friendly content. Microsoft reportedly invested $10 billion in the company earlier this month and has plans to incorporate it into Bing and Office 365. OpenAI was rumoured to have discussed a share offering at a $29 billion valuation late last year, making it one of the most valuable US startups at the time.
While ChatGPT is the best natural language AI model available to the public at the moment, competitors are in development by Google, Baidu, and others. Earlier this year, a disgruntled Google researcher lost his job after publicly declaring that the company’s LaMDA chatbot is conscious.
Due to ethical concerns, Google has only released a very limited beta version of its chatbot. It will be fascinating to see how long the restraining order holds after the genie has been let out of the bottle.
1 Comment
Pingback: How ChatGPT is taking over the self-publishing industry with over 200 AI-written books on Amazon - Innovation Village | Technology, Product Reviews, Business