The award-winning Australian science magazine Cosmos, published by the CSIRO, published AI-generated articles on its website over a two-week period in July.
Each article included a disclaimer stating that it was created using OpenAI and that the content was reviewed by a trained science communicator and edited by the Cosmos publishing team.
Topics covered in the article included painkillers, decomposition, black holes, and cybersecurity.
The Science Journalists Association of Australia criticised the articles for containing incorrect or oversimplified information and said the use of AI in this way was worrying.
Jackson Ryan, president of the association, pointed out that these inaccuracies would damage people’s perception of Cosmos and undermine their trust in Cosmos’ reliability.
The CSIRO issued a statement supporting the “experimental project” that will examine the “opportunities and risks” of AI. The experiment is scheduled to run until February 2025.
The project was funded by a 2023 grant from the Meta Australian News Fund through the Walkley Foundation. The Walkley Foundation claims that it “sets the industry standard for excellence and best practice journalism”.
According to former editors, Cosmos did not inform the editorial staff about the AI project before its launch.
Many former staff members disagreed with the proposal, with former editor-in-chief Ian Connellan saying that if he had known about it, he would have said it was a bad idea.
Cosmos co-founder Kylie Ahern, who served as CEO until 2013, stated that AI-generated articles were “not the right direction” for Cosmos.
Cosmos employees also criticized the project, saying they were not consulted about whether their work could potentially be used to train an AI model.
Bianca Nogrady, a freelance science journalist who has written for Cosmos, said she was “incredibly disappointed” and had “contacted CSIRO and the Cosmos editors twice in the last week and received no response”.
A spokesperson for CSIRO Publishing dismissed the contributors’ concerns in a statement, saying that “the experiment did not involve training OpenAI’s GPT-4 model.”
GPT-4 has been criticized for being trained on huge datasets containing billions of web pages and copyrighted material, likely including the archives of Cosmos itself.
ChatGPT is currently being sued by several authors and the New York Times for misusing copyrighted works to train its AI model.
Cosmos has not published an AI-generated article since the end of July.
