The Microsoft-backed company says the new classifier, as its name suggests, has been developed to combat malicious use of AI content generators, such as its very popular ChatGPT, in “run automated disinformation campaign (opens in a new tab)… academic dishonesty and positioning an AI chatbot as a human.”
So far, it claims that the classifier has a 26% success rate in identifying AI-generated content, correctly labeling it as “likely AI-written”, and false positive is 9% when misrepresenting human work as artifical. create.
Spot the difference
OpenAI notes that the classifier performs better when the text is longer, and that, compared to previous versions, the newer version is “significantly more reliable” in detecting auto-generated text from the engines. Newer AI.
The classifier is now publicly available, and OpenAI will use the feedback it receives to determine its usefulness and help improve the further development of AI detection tools in the future. .
OpenAI is keen to point out that it has limitations and should not be relied upon as a “primary decision-making tool”, a view shared by most participants in all fields of AI.
As mentioned, the length of the text is critical to the success of the classifier, with OpenAI claiming it’s “very unreliable” for pieces with less than a thousand characters.
Even longer texts can be incorrectly identified, and human-written content can be “incorrectly labeled but certainly written by AI”. Also, it performs worse on text written in non-English languages as well as computer code.
Predictable text where the content can only be written realistically nor can it be reliably labeled, such as a list of the first thousand primes, to give an example of OpenAI.
Furthermore, OpenAI shows that AI text can be edited to fool the classifier, and although the classifier can be updated and learn from being tricked like this, interestingly, the company says knows the “unclear whether detectability has an advantage in the long run” term.”
Text that is also very different from the text it was trained on can also cause classifier problems, with “sometimes [being] extremely confident in a false prediction.”
For this training data, OpenAI says it used pairs of texts written on the same topic, one generated by the AI and the other supposedly written by a human – some gathered from feedback. human to the prompt used to train InstructGPT, the AI model from the company mainly used by researchers and developers.
The development of the classifier comes amid much concern and debate surrounding the use of AI chatbots, such as OpenAI’s ChatGPT, in academic institutions such as high schools and universities.
Cheating allegations are on the rise because students are using chatbots to write assignments for them. Essay submission platform Turnitin even developed its own AI writing detection system (opens in a new tab) in response.
OpenAI acknowledged this fact and even produced separate set of guides for educators (opens in a new tab) to understand the uses and limitations of ChatGPT. It hopes its new classifier will benefit not only the organization but also “journalists, misinformation researchers and other groups”.
The company wants to interact with educators to hear about their experience with ChatGPT in the classroom, and they can use this sample (opens in a new tab) to send feedback to OpenAI.
AI writing tools have also caused a stir elsewhere. technology page CNET was recently criticized for using AI tools to write articles (opens in a new tab)as part of the trial, but was accused of failing to distinguish the articles from those written by real people. Such articles were also found to contain some fundamental errors of fact.