UNSW reveals new measures to deal with AI cheating
Key points
- UNSW is expanding its ChatGPT Edu rollout to 80,000 students and staff after a limited deployment to 10,000 staff late last year.
- New academic policy measures will include an “assessment transformation” dividing assessments into “can”, “can’t” and “must” AI modes.
- The university says AI watermarking and detection tools aren’t yet reliable enough alone to determine misconduct, and will only support educator judgement.
UNSW Sydney says it will introduce new administrative measures to address “academic integrity” as its moves to expand its rollout of OpenAI’s ChatGPT Edu to tens-of-thousands of students.
The AI assistant is being made available to 80,000 students and staff following a limited rollout to 10,000 fixed-term and permanent staff that began late last year.
The deployment, which OpenAI described as a “strategic collaboration”, will also see OpenAI’s technology incorporated into teaching materials, including its AI coding tool, Codex.
A UNSW spokesperson told iTnews that the rollout would be accompanied by a range of academic policy measures, including what it described as an “assessment transformation” that would see it set new “AI modes”.
It will also revise guidance around its expectations “in relation to academic integrity”.
The new AI modes will divide assessments into three categories designed to give students clear directions on when to use the technology: “can”, “can’t” and “must”.
“This will ensure we manage the balance of maintaining trust in our qualifications while supporting the development of AI capability,” the spokesperson said.
At this stage, the university doesn’t view technological tools such as AI watermarking as sufficiently reliable on their own to provide a basis for identifying instances of misconduct or cheating.
“We will continue to monitor the development of technical safeguards noting our position is that any detection tools and watermarking will not be relied upon in insolation for determining misconduct. They represent a tool to support educator judgement,” the university’s spokesperson said.
Sources familiar with OpenAI’s AI development for education settings told iTnews that the company was working on new technological measures to steer students away from cheating behaviours and encourage learning.
For instance, it’s understood that the company is working on measures to detect when students might be trying to take shortcuts in assignments and redirect them back to a safer learning path.
The rollout to students could also bring the university and OpenAI within the reach of the Australian Government’s proposed digital duty of care legislation released earlier this week.
The new legislation mainly targets social media platforms and aims force them to give Australian users the ability to opt out of feeds informed by what the government considers to be harmful algorithms.
However, it also captures AI chat bots and assistants, and requires their makers to protect minors from negative content that promotes eating disorders, misogyny or causes serious mental distress.
The university said that ChatGPT Edu would include administrative controls to work alongside OpenAI safety systems designed to “reduce the generation of illegal, dangerous or otherwise harmful content”.
Nevertheless, it said it would keep a close watch on the government’s digital duty of care legislation.
“UNSW is following these developments closely and will keep reviewing and strengthening our systems for identifying, monitoring and responding to issues affecting student and staff wellbeing — including where AI use may be a contributing factor,” UNSW’s spokesperson said.
OpenAI also told iTnews that it was watching the legislation closely and indicated it was willing to work with the government on the bill.
“Providing users, particularly those aged under 18, with a positive AI experience is a priority for OpenAI. We are reviewing the Government’s exposure draft of the digital duty of care bill and look forward to constructive engagement on protecting Australians while enabling them to benefit from AI,” a spokesperson for OpenAI said.
By default, ChatGPT Edu is configured to prevent OpenAI from collecting information into the assistant to train its models and UNSW said that the arrangement would be carried over from the original deployment to staff last year.
UNSW has opted for a revenue model based on tokens for its arrangement with OpenAI rather than licenses.
“UNSW will pay for consumption of tokens, within a fair use limit,” UNSW’s spokesperson said.
OpenAI said that its collaboration UNSW was the first of its kind it had struck within Australia’s academic sector.
The dean of the UNSW Business and AI lead Paul Andon framed the partnership as a national skill-building exercise.
“The collaboration is intended to provide a model for institution-wide, responsible AI capability development – helping more Australian graduates compete globally and ensuring AI-driven innovation flows beyond the campus into businesses, startups, government and critical sectors,” Andon said.
About Author
