Two months after OpenAI upset some educators with the public release of ChatGPT, an AI chatbot that can help students and professionals generate shockingly persuasive essays, the company is unveiling a new tool to help educators adapt.
OpenAI announced Tuesday a new feature, called an “AI text classifier”[ads1];, which allows users to check whether an essay is written by a human or AI. But even OpenAI admits it’s “imperfect.”
The tool, which works in English AI-generated text, is powered by a machine learning system that takes an input and assigns it to multiple categories. In this case, after pasting a text as a school assignment into the new tool, it will give one of five possible outcomes, from “probably generated by AI” to “very unlikely”.
Lama Ahmad, policy research director at OpenAI, told CNN that educators have been asking for a ChatGPT feature like this, but cautioned that it should be “taken with a grain of salt.”
“We really don’t recommend taking this tool in isolation because we know it can be wrong and will be wrong at times — much like using AI for any kind of assessment purpose,” Ahmad said. “We emphasize how important it is to keep a human in the loop … and that it’s just one data point among many others.”
Ahmad notes that some teachers have referred to previous examples of student work and writing style to gauge whether it was written by the student. While the new tool may provide a different reference point, Ahmad said “teachers need to be very careful how they include it in academic dishonesty decisions.”
Since being made available in late November, ChatGPT has been used to generate original essays, stories and song lyrics in response to user prompts. It has produced research reports that fooled some researchers. It even recently passed law exams in four courses at the University of Minnesota, another exam at the University of Pennsylvania’s Wharton School of Business and a medical licensing exam in the United States.
In the process, it has raised alarms among some teachers. Public schools in New York City and Seattle have already banned students and teachers from using ChatGPT on district networks and devices. Some teachers are now moving with remarkable speed to rethink their assignments in response to ChatGPT, although it remains unclear how widespread the use of the tool is among students and how detrimental it may actually be to learning.
OpenAI now joins a small but growing list of efforts to help educators detect when a written work is generated by ChatGPT. Some companies such as Turnitin are actively working with ChatGPT plagiarism detection tools that can help teachers identify when assignments are written by the tool. Meanwhile, Princeton student Edward Tuan told CNN that more than 95,000 people have already tried the beta version of his own ChatGPT detection feature, called ZeroGPT, noting that there has been “incredible demand among educators” so far.
Jan Leike – head of the OpenAI adaptation team, which works to ensure that the AI tool is aligned with human values – listed several reasons why detecting plagiarism via ChatGPT can be a challenge. People can edit text to avoid being identified by the tool, for example. It will also “be best at identifying text that is very similar to the type of text we’ve trained it on.”
plus company said it is impossible to determine whether predictable text — such as a list of the first 1,000 prime numbers — was written by AI or a human because the correct answer is always the same, according to a company blog post. The classifier is also “very unreliable” on short texts under 1000 characters.
During a demo with CNN ahead of Tuesday’s launch, ChatGPT successfully tagged several working bodies. An excerpt from the book “Peter Pan”, for example, was deemed “unlikely” to be AI-generated. However, in the company’s blog post, OpenAI said it incorrectly labeled human-written text as AI-written 5% of the time.
Despite the possibility of false positives, Leike said the company aims to use the tool to spark conversations around AI competence and possibly dissuade people from claiming AI-written text was created by a human. He said the decision to release new feature also stems from the debate about whether people have the right to know if they are interacting with AI.
“This question is much bigger than what we are doing here; society as a whole has to deal with that question, he said.
OpenAI said it is encouraging the public to share their feedback on the AI check feature. Ahmad said the company continues to talk to K-12 teachers and those at the collegiate level and beyond, such as Harvard University and Stanford Design School.
The company sees its role as “an educator for the teachers,” according to Ahmad, in the sense that OpenAI wants to make them more “aware of the technologies and what they can and can’t be used for.”
“We’re not teachers ourselves — we’re very aware of that — and so our goals are really to help teachers equip teachers to deploy these models effectively in and out of the classroom,” Ahmad said. “That means giving them the language to talk about it, helping them understand the possibilities and limitations, and then secondarily through them, equipping students to navigate the complexities that AI is already introducing into the world.”