Scott Farquhar: AI should train freely on creative content, ignoring copyright
Atlassian co-founder Scott Farquhar has urged Australia to overhaul its copyright laws, advocating for a US-style “fair use” exemption that would permit artificial intelligence models to train freely on copyrighted creative content. Farquhar, who also serves as chief executive of the Tech Council of Australia, argues that current Australian law likely renders most AI data mining activities illegal, thereby stifling investment in the burgeoning AI industry within the country.
Speaking on ABC’s 7.30 program, Farquhar asserted that the economic benefits of large language models far outweigh concerns about AI leveraging existing creative works without compensation. His core argument hinges on whether the AI’s output is “transformative”—meaning it creates something “new and novel” rather than merely replicating an existing work. He posits that if an AI assists in generating an entirely new song or piece of software, for instance, such usage should be deemed fair. Farquhar expressed no issue with his own intellectual property being used in a transformative way, provided it doesn’t lead to direct competition. He believes that using the world’s collective software knowledge to improve future software development falls squarely within the bounds of fair use.
However, Farquhar’s proposal overlooks a critical point: the legal landscape surrounding AI training and fair use in the United States is far from settled. Despite calls from major AI developers like Atlassian, Google, and Meta for broad exemptions allowing perpetual, unpaid training on human works, the issue is currently the subject of dozens of lawsuits. The US Copyright Office, in its May pre-print report on generative AI training, explicitly noted the significant legal challenges to AI companies’ claims of fair use.
US copyright law considers several factors when determining fair use, including whether the use is commercial, the nature of the original copyrighted work, the amount of the work used, and, crucially, the effect of the use on the market for or value of the copyrighted work. The latter has been repeatedly identified by the US Supreme Court as “undoubtedly the single most important element” of fair use. The Copyright Office report underscores this, warning that AI training “threatens significant potential harm to the market for or value of copyrighted works.” It explains that if an AI model can produce outputs substantially similar to, or even stylistically akin to, works in its training data, it could lead to lost sales or dilute the market for original creations.
For instance, in the news industry, AI-generated summaries already reduce the need for users to click through to original articles, severely impacting traffic and revenue for publishers. While the US Copyright Office has stopped short of recommending legislative intervention, it acknowledges that voluntary licensing agreements are emerging, offering a path for AI innovation to proceed without undermining intellectual property rights.
Farquhar’s vision for AI’s free access to creative works would only hold water if there were ironclad guarantees that all AI usage would be genuinely transformative and would not negatively impact the markets from which the data is drawn. Given the current legal battles and the demonstrated potential for market disruption across various creative sectors, rushing to adopt an unsettled US legal concept could come at a severe cost. Granting tech companies broad exemptions in the name of innovation for one industry risks undermining the viability and sustainability of many others.