The maker of ChatGPT has launched a new tool designed to help teachers detect if the text was written by a student or artificial intelligence.
Key points:
- The tool to detect AI-written text “is imperfect and it will be wrong sometimes”
- ChatGPT has been banned in public schools in New South Wales, Queensland, Western Australia and Tasmania
- OpenAI says it has been working for several weeks to craft new guidelines to help educators
The new text classifier by OpenAI follows discussions at schools and universities over fears that ChatGPT’s ability to write just about anything on command could fuel academic dishonesty and hinder learning.
However, the method for detecting AI-written text “is imperfect and it will be wrong sometimes”, according to the head of OpenAI’s alignment team, Jan Leike.
“Because of that, it shouldn’t be solely relied upon when making decisions.”
Millions have experimented with ChatGPT since it launched in November as a free application on OpenAI’s website.
While many found ways to use it creatively and harmlessly, the ease with which students could answer take-home test questions and assist with other assignments has sparked panic among some educators.
There are concerns the tool will be used to cheat and plagiarise, with some universities moving quickly to rewrite exams, essay questions and integrity procedures.
ChatGPT has already been banned in public schools in New South Wales, Queensland, Western Australia and Tasmania.
However, University of Western Australia’s Julia Powles said she felt the cheating concern was “overblown” noting that “if you’re setting assessments that could be addressed simply by drawing on web resources, then you may have a problem.”
“Ever since we’ve had the ability to search the web or access material on Wikipedia, people have been able to draw on digital resources,” she said.
OpenAI emphasised the limitations of its detection tool in a blog post, but said that in addition to deterring plagiarism it could help to detect automated disinformation campaigns and other misuse of AI to mimic humans.
The longer a passage of text, the better the tool is at detecting if an AI or human wrote something.
The tool labels text as either “very unlikely, unlikely, unclear if it is, possibly, or likely” AI-generated.
But much like ChatGPT itself, it is not easy to interpret how it came up with a result.
“We don’t fundamentally know what kind of pattern it pays attention to, or how it works internally,” Ms Leike said.
“There’s really not much we could say at this point about how the classifier actually works.”
In France, Sciences Po has prohibited the use of ChatGPT warning anyone found to be using it other AI tools to produce written or oral work could be banned.
In response to the backlash, OpenAI said it had been working for several weeks to craft new guidelines to help educators.
“Like many other technologies, it may be that one district decides that it’s inappropriate for use in their classrooms,” said OpenAI policy researcher Lama Ahmad.
“We don’t really push them one way or another. We just want to give them the information that they need to be able to make the right decisions for them.”
France’s digital economy minister Jean-Noël Barrot recently met in California with OpenAI executives, including chief executive Sam Altman.
After the meeting he told the world economic forum in Switzerland that he was optimistic about the technology.
“If you’re in the law faculty, there is room for concern because obviously ChatGPT, among other tools, will be able to deliver exams that are relatively impressive,” he said.
“If you are in the economics faculty, then you’re fine because ChatGPT will have a hard time finding or delivering something that is expected when you are in a graduate-level economics faculty.”
He said it would be increasingly important for users to understand the basics of how these systems work so they know what biases might exist.
AP/ABC