After telegraphing the move through media appearances OpenAI has thrown out a tool that attempts to distinguish between human-written and AI-generated text, such as text produced by the company's own ChatGPT and GPT-3 models. The classifier isn't particularly accurate: its success rate is around 26%, OpenAI notes, but OpenAI argues that, when used in conjunction with other methods, it could be useful in helping prevent abuse of AI text generators. .
“The classifier is intended 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 to determine the source of the text rather than being the main decision-making tool,” said an OpenAI spokesperson. We are making this initial classifier available for feedback on whether tools like this are useful and we look forward to sharing improved methods in the future.
As fervor grows around generative AI, particularly AI that generates text, critics have called on the creators of these tools to take steps to mitigate their potentially damaging effects. Some of the largest school districts in the United States have banned ChatGPT from their networks and devices, fearing the impacts on student learning and the accuracy of the content the tool produces. And sites that include Stack Overflow has kicked users out who shared content generated by ChatGPT, saying that the AI makes it too easy for users to flood discussion threads with dubious answers.
The OpenAI classifier, aptly named OpenAI AI Text Classifier, is challenging from an architectural point of view. Like ChatGPT, it is an AI language model trained on many, many examples of publicly available text on the web. But unlike ChatGPT, it is tuned to predict the probability that the AI generated a piece of text, not just from ChatGPT, but from any AI model that generates text.
More specifically, OpenAI enabled the OpenAI AI Text Classifier in texts from 34 content generation systems from five different organizations, including OpenAI itself. This text was combined with similar (but not exactly similar) human-written text from Wikipedia, sitios web excerpted from links shared on Reddit and a set of "human demos" collected for an older OpenAI text generation system. (OpenAI supports in a supporting document however, that he might have inadvertently misclassified some AI-written text as human-written "given the proliferation of AI-generated content on the Internet").
The text classifier OpenAI it will not work with any text, which is more important. You need a minimum of 1000 characters, or about 150-250 words. It does not detect plagiarism, an especially unfortunate limitation considering that text-generating AI has been shown to regurgitate the text on which he was trained. And OpenAI says it's more likely to err on text written by children or in a language other than English, because of its English dataset.
The detector somewhat covers its answer by evaluating whether a given piece of text is AI generated. Depending on your confidence level, it will label the text as AI-generated “highly unlikely” (less than 10% probability), AI-generated “unlikely” (10-45% probability), “unclear” if it is ” AI Generated (45-90% chance), “Possibly” AI-Generated (90-98% chance) or “Probably” AI-Generated (greater than 98% chance) .
Out of curiosity, we ran some text through the classifier to see how it worked. While he correctly and confidently predicted that several paragraphs of a TRPlanes article on Meta Horizon Worlds and a snippet of an OpenAI support page were not AI generated, the classifier had more difficulty with the ChatGPT article length text and ultimately failed to classify it. total. However, it successfully detected the ChatGPT output from a Gizmodo piece about – what else? —ChatGPT.
According to OpenAI, the classifier incorrectly labels human-written text as AI-written 9% of the time. That error did not occur in the tests, but it can be attributed to the small sample size.
Credits: OpenAI
On a practical level, the classifier is not particularly useful for evaluating shorter writings. 1000 characters is a difficult threshold to reach in the field of messages, for example, emails (at least the regular ones). and has several limitations such as OpenAI emphasizes that the classifier can be circumvented by modifying some words or clauses in the generated text.
That does not mean that the classifier is useless, far from it. But it certainly won't stop committed scammers (or students, for that matter) in its current state.
The question is, will there be other tools? A sort of cottage industry has sprung up to meet the demand for AI-generated text detectors. ChatZero, developed by a Princeton University student, uses criteria including “stumpiness” (the complexity of the text) and “bursts” (the variations of sentences) to detect whether the text might be AI-written. The plagiarism detector Turnitin is developing its own AI generated text detector. Beyond that, a Google search turns up at least half a dozen other applications that claim to be able of separating the AI-generated wheat from the human-generated chaff, to twist the cliché.
It is likely to become a game of cat and mouse. As text-generating AI improves, so will detectors, a never-ending back-and-forth similar to that between cybercriminals and security researchers. And as OpenAI writes, while classifiers can help in certain circumstances, they will never be a single reliable test to decide whether the text was generated by AI.
That's a way of saying there's no silver bullet to solve the problems posed by AI-generated text. It is very likely that there never will be.
