The Australian Research Council has faced allegations that some of its peer reviewers may have used ChatGPT to assess research proposals, prompting a warning from the education minister and concerns about possible academic misconduct.
Several researchers have reported that some assessor feedback provided as part of the latest Discovery Projects round of grant funding included generic wording suggesting they may have been written by artificial intelligence.
One academic, who wished to remain anonymous, told Guardian Australia that one of the assessor reports they received included the words “Regenerate response” – text which appears as a prompt button in the ChatGPT interface.
“It’s quite a positive report, but it’s quite bland also, and it quotes back the proposal at you,” the researcher said. “It’s almost like reading something you’ve written yourself.”
After they submitted a complaint to the ARC, the report was removed.
The researcher said the apparent use of AI pointed to the time pressures faced by academics in Australia and also a possible lack of quality control internally by the ARC.
“I think it’s a sign of someone being overworked and trying to cut corners … If you’ve used artificial intelligence to generate a response, you lose the ability to engage in a proper academic cut and thrust.”
Detailed assessor reports are typically written by academics in closely related fields and are used by the ARC’s College of Experts to decide which grant proposals should ultimately receive government funding. Only 19% of Discovery Projects in last year’s funding round were ultimately successful. The outcomes of the 2023 grant round have not yet been announced.
The affected researcher called for greater transparency from the ARC. Academics receive assessor reports on their grant proposals but are not concurrently given their scores for each corresponding report.
“If you suspect this is a ChatGPT report, but you don’t have the proof that I did, you have no way to respond to it. You should be able to … [point out if] the scores are inconsistent.”
The federal education minister, Jason Clare, told Guardian Australia in a statement: “The use of AI in this way is not acceptable.”
Clare said he had instructed the ARC to “put in place measures to ensure it doesn’t happen”.
Researchers who receive ARC money are required as a formal condition of their grant funding to write assessor feedback for other academics’ proposals. In a given year, researchers are asked to assess up to 20 proposals, which are each typically 50 to 100 pages long.
Andrew Francis, a professor of mathematics at Western Sydney University, said if information from grant proposals was being put into ChatGPT, that would constitute “a violation of confidentiality agreements that the assessor has signed on to”.
“If actual judgments are being generated by [Chat]GPT then it’s excruciatingly dishonourable on the part of the assessor,” Francis said. “To my mind, it’s academic misconduct worthy of being denied future funding.
“The ARC must make it extremely clear that using AI to make assessments is completely unacceptable.”
An Australian academic who runs the Twitter account ARC Tracker said they had read four assessor reports received by researchers “where it was just absolutely clear [and] no one could conclude anything else but that the assessments had been done by ChatGPT”. They were aware of four other instances of suspected generative AI use.
“Quality control of assessments has been something that researchers have been talking to the ARC about for a long time, and they’ve done basically nothing about it,” ARC Tracker’s administrator said.
In 2021, a pre-budget submission co-signed by more than 1,000 academics suggested that the ARC introduce consequences for inappropriate and unprofessional reviewer feedback.
“There’s enormous pressure on the peer review approach to assessing research in Australia,” ARC Tracker’s administrator added. “Most universities don’t give their researchers time in a formal and documented way to review anything – whether other people’s papers, grant proposals, proposals for using infrastructure … that’s counted in your research time.
“As any researcher will tell you, you can spend a lot of time assessing other people’s research while not getting any time for your own.
“The ARC should have seen this coming.”
In a public statement, the ARC advised that “peer reviewers should not use AI as part of their assessment activities”.
An ARC spokesperson told Guardian Australia that more than 7,000 assessors contributed to ARC peer review processes in 2021-22.
They said: “The ARC has a conflict of interest and confidentiality policy which outlines the requirements around confidentiality in the conduct of ARC business, including peer review. All ARC assessors confirm their acceptance of this policy when undertaking assessments.
“While generative artificial intelligence (AI) tools such as ChatGPT are not explicitly named in this policy, the common principles of confidentiality apply across both existing and emerging channels through which confidential information may be inappropriately disclosed.
“Developments in generative AI are fast-moving and bring complex considerations including the balance of opportunities and risks. The ARC is closely monitoring these developments and is engaging with other research funding agencies both in Australia and overseas on these issues.”
