Move in after telegraphing Media presenceOpenAI is started A tool that tries to distinguish between human-written and AI-generated text – similar to text produced by the company’s own ChatGPT and GPT-3 models. The classifier isn’t particularly accurate — its success rate is about 26%, OpenAI notes — but OpenAI argues that it, when used in conjunction with other methods, can be useful in helping prevent AI text generators from being misused.
“The purpose of the classifier is to help reduce false claims that AI-generated text is written by humans. However, it still has many limitations – so it should be used as a supplement to other methods of determining the source of text rather than as a primary decision-making tool,” an OpenAI The spokesperson told TechCrunch via email. “We are providing this initial classifier to get feedback on whether these types of tools are useful, and hope to share improved methods in the future.”
As excitement around generative AI — especially text-generating AI — grows, critics have called for the creators of these tools to take steps to mitigate their potentially harmful effects. Some large school districts in the US have banned ChatGPT on their networks and devices, fearing it could affect student learning and the accuracy of the content the device produces. and including sites Stack Overflow has banned users From sharing content generated by ChatGPT, saying that AI makes it too easy for users to flood discussion threads with questionable answers.
OpenAI’s classifier – aptly called the OpenAI AI Text Classifier – is architecturally interesting. It, like ChatGPT, is an AI language model trained on many, many examples of publicly available text from the web. But unlike ChatGPT, it’s better at predicting how likely a piece of text is to be generated by AI—not just from ChatGPT, but any text-generating AI model.
More specifically, OpenAI trained the OpenAI AI text classifier on text from 34 text-generation systems from five different organizations, including OpenAI. This text was combined with similar (but not exactly identical) human-written text from Wikipedia, websites extracted from links shared on Reddit, and a set of “human performances” collected for the previous OpenAI text-generation system. (OpenAI admits Supporting documentsHowever, it may have inadvertently misclassified some AI-written texts as human-written “given the proliferation of AI-generated content on the Internet”.)
The OpenAI text classifier doesn’t work on any text, crucially. It requires a minimum of 1,000 characters, or about 150 to 250 words. It doesn’t detect plagiarism—an especially unfortunate limitation considering that it features text-generating AI. Reorganize Lessons in which this training was given. And OpenAI says it’s more likely to get things wrong with text written by children or in a language other than English, due to its English-forward data set.
The detector hedges its answer slightly when evaluating whether a piece of text is AI-generated. Based on its confidence level, it will label the text as “very unlikely” AI-generated (less than 10% chance), “unlikely” AI-generated (between 10% and 45% chance), “unclear.” AI- generated (a 45% to 90% chance), “probably” AI-generated (90% to 98% chance) or “probably” AI-generated (greater than 98% chance).
Out of curiosity, I fed some text through the classifier to see how it could manage. With several paragraphs from a TechCrunch article about Meta’s Horizon Worlds and a snippet from the OpenAI support page confident that the AI had not been generated, correctly predicted, the classifier had a hard time with the article-length text from ChatGPT, ultimately failing to classify it. completely However, it successfully detected the ChatGPT output from Gizmodo a piece About – what else? – ChatGPT.
According to OpenAI, the classifier incorrectly labels human-written text as AI-written 9% of the time. That mistake didn’t happen in my testing, but I chalk it up to the small sample size.
Image credit: OpenAI
On a practical level, I found the classifier not particularly useful for evaluating short pieces of writing. 1,000 characters is a difficult threshold to reach in a range of messages, for example emails (at least the ones I receive regularly). And the limitations give pause – OpenAI insists that the classifier can be circumvented by modifying certain words or phrases in the generated text.
That’s not to suggest that classifiers are useless – far from it. But that certainly doesn’t stop committed fraudsters (or students, for that matter) in the current situation.
The question is, other tools? Something of a cottage industry has developed to meet the demand for AI-generated text detectors. ChatZero, developed by a Princeton University student, uses criteria including “paraplexity” (complexity of text) and “burstiness” (variation of sentences) to detect whether text may be AI-written. Plagiarism detector Turnitin is developing its own AI-generated text detector. Besides those, a Google search yields at least half a dozen other applications that claim to be able to separate the AI-generated wheat from the human-generated chaff, to torture the metaphor.
It will probably become a cat and mouse game. As text-generating AI improves, so will the detectors — never-before-seen ones like those between cybercriminals and security researchers. And as OpenAI writes, while classifiers can help in certain situations, they will never be the sole piece of evidence in deciding whether a text was AI-generated.
This is all to say that there is no silver bullet to solving the problems of AI-generated text. Probably, there never will be.
