Posted: 2024-08-07 20:00:00

A decision by popular Australian science magazine Cosmos to use a Walkley Foundation grant to publish articles generated by artificial intelligence (AI) has drawn criticism from its own contributors and former editors, including two co-founders.

The CSIRO, which publishes Cosmos, says it backs the "experimental project", which is designed to investigate the "opportunities and risks of using AI", and scheduled to run until February 2025.

But critics say the AI service undermines journalism and was built without proper consent.

The Walkley Foundation, which supports excellence in journalism, knew of Cosmos's plan to publish AI-generated articles when it oversaw and awarded the grant in February 2024.

The September 2023 Cosmos Magazine cover.

Cosmos' 100th edition, published in September 2023.(Supplied: Cosmos Magazine)

The controversy is an example of growing anxieties around the role of AI in journalism as publishers experiment with new productivity tools. Hundreds of journalists employed by Nine Entertainment went on strike last week, partly over AI protections.

Cosmos ran into financial difficulties and lost half its staff earlier this year, having won dozens of journalism and industry awards over 20 years of publishing. National science agency CSIRO took over the publication in June

During two weeks last month, Cosmos published six AI-generated explainer articles on its website on topics ranging from black holes to carbon sinks.

Each article stated, "This article was generated by our custom AI service."

"Our service was built to focus on our archive of more than 15,000 factually correct science news stories and features. It also uses Open AI to help create the content. All generated content is fact checked by a trained science communicator and edited by our publishing team."

Cosmos contributors reacted to the move with disbelief, saying their work had been used to develop an AI service that generated articles and, in their opinion, undermined their role as journalists.

Many were angry they had not been consulted and said their calls to the publisher had gone unanswered.

"I've contacted CSIRO and the Cosmos editors twice in the last week and had no response," Bianca Nogrady, a freelance science journalist and Cosmos contributor, said.

A collection of Cosmos Magazine covers.

Founded in 2004, Cosmos set out to be "the New Yorker of science".(Supplied: Cosmos Magazine)

Editorial staff at Cosmos were also not told about the proposed custom AI service, two former editors said.

According to Ian Connellan and Gail MacCallum, Cosmos's former publisher, the Royal Institution of Australia (RiAus), did not tell them it was applying for funding for a custom AI service in late 2023, even though they were in charge of Cosmos's editorial decisions.

"We had no knowledge of the proposal to employ AI as a background writer creator," Mr Connellan, who was the RiAus editor-in-chief until February, said.

"As editor-in-chief, I would have said this is a bad idea."

Ms MacCallum, Cosmos's managing editor at the time, said there were questions to answer about the ethics of using AI in such a way.

"I'm a huge proponent of exploring AI, but having it create articles of fact is a little past my comfort zone."

Kylie Ahern, who co-founded Cosmos in 2004 and served as chief executive until 2013, said AI-generated articles were "not the right direction" for her former employer. 

Wilson da Silva, another co-founder who edited Cosmos from 2004 to 2013, said it was "definitely not what [he] would've done."

AI not specifically trained on Cosmos articles: CSIRO

Cosmos's acting editor, Gavin Stone, referred questions from the ABC to CSIRO Publishing.

In a statement, a CSIRO Publishing spokesperson dismissed concerns that Cosmos's AI service was trained on contributors' articles.

"This experiment does not involve training OpenAI's GPT-4 model (which was pre‑trained by OpenAI)," the spokesperson said.

The service works by using OpenAI's GPT-4 to generate text on various topics. It then automatically fact-checks this against the publisher's large, well-researched database of 15,000 stories and feature articles using a process called retrieval-augmented generation (RAG).

GPT-4 was trained on vast datasets including CommonCrawl, a repository of content scraped from billions of web pages, including those of Cosmos Magazine. The legality of this approach is still being decided, and The New York Times and several authors are currently suing OpenAI for using scraped copyrighted work to train its AI.

View More
  • 0 Comment(s)
Captcha Challenge
Reload Image
Type in the verification code above