Australian creatives should be terrified that Amazon is taking their work to train the online retailer’s AI, a federal parliamentary inquiry was told.
Amazon representatives were questioned during a hearing of the Select Committee on Adopting Artificial Intelligence.
Matt Levey, head of public policy, Australia and New Zealand, Amazon, was asked about demands by the Association of Voice Actors and the union, the Media, Entertainment & Arts Alliance, to prohibit companies using AI to create synthesised voice or likenesses of an actor.
The creatives want contracts with Amazon to be changed to put permissions for use of any content in the hands of the content owners.
Levey replied that Amazon's Audible, while continuing to work through issues, isn't planning to make changes to current contracts
“It still defies belief that you're unable to tell us whether Amazon has used the work of the Australian creatives across Audible, Kindle and Prime Video to train your AI models,” said the inquiry chair, Labor Senator Tony Sheldon.
“Your answers provide no comfort to people worried about getting ripped off as a result of their life's work. If I was a writer or a voice actor listening to this, I'd be terrified.”
Levey said he would be happy to come back to the committee on the specific query about whether that content is being used to train AI models.
Senator Sheldon said the issue is a critical and fundamental piece of information.
“You've demanded consent to all current —I'm using the terminology in the language that's used in the contracts — and future forms of exploitation of the rights granted to the producer under this agreement,” he said.
“You've actually said that you're going to have the full rights to everybody's creative presence to use in any fashion you like, and if you don't take it then leave it and we will find somebody else.
“That's not a negotiation, and you've just said it now. Amazon has no intention, as you understand it, to make any change to that. I find it reprehensible and it does require governments to act.”
Nicole Foster, director, Global AI/ML Public Policy, Amazon Web Services, told the inquiry that replicating somebody's content obviously would be a copyright infringement,
“But reading a book or hearing a song and learning from that is not against what copyright is intended to protect,” she said via video link.
“I think that the technology has sort of thrown some new questions at us to consider in terms of what the meaning of copyright is and how it was intended.
“As policy-makers who are thoughtful about the role of Australia in light of this AI technology, I wanted to share a couple of thoughts for you about some of the considerations that you might want to have as you think about changes to the copyright regime.
“I'm Canadian, not American. As you probably well know, the vast majority of technology that is AI technology is either American based or it is Chinese based.
“For those of us that live in parts of the world that are not the US and China but are interacting with this technology, I'm very conscious of the fact that some of our content, whether it's our cultural references or minority language, are going to be left out.
“As we think about how we want to participate in this technology and the impact it's going to have on our citizens, we need to ensure it is actually representative of our own society.
“There may be trade-offs that you want to consider about the availability of content that is specific to your region versus not just having a highly protective regime.”
Senator Sheldon: “Isn't it also that part of the Australian culture should be reflected in AI outputs?
“We would say if you're stealing without paying for recognition—and that's financial recognition as well as personal recognition—that's still stealing. That's part of the Australian culture.”
Amazon’s Foster said there are different copyright regimes around the world.
“This isn't relevant to Australia, but there's a fair use doctrine,” she said.
“Some jurisdictions have actually specific exemptions for text and data mining.
“Thinking about copyright regime, I think you need to think about both the input end of copyright and the output end of copyright. You may want to think about more flexibility on the learning and the training of how that data is used or how that content is used, and then focus more effort on suppressing what is outputted from the model.
“There are two ends of how you can think about copyright and AI, but the availability of content is going to be, I think, key in ensuring that non-dominant cultures are represented in this technology, and this technology is becoming very influential in a lot of aspects of our world.”
Foster said that there is a difference between replicating creative content and learning from that content to produce something new.
“if you photocopy a textbook rather than purchase the textbook book, there is an infringement of copyright versus I'm reading the textbook,” she said.
“I'm taking notes from the textbook, I'm learning from that textbook and then applying that knowledge elsewhere. It's an analogy, but it's really a very similar concept.
“The output of the model should not, in this analogy, copy the inputted information. It should learn from it and extract from it and then produce something different. The issue from a copyright standpoint in directly copying that information would be a copyright infringement.”
Have something to say on this? Share your views in the comments section below. Or if you have a news story or tip-off, drop us a line at adnews@yaffa.com.au
Sign up to the AdNews newsletter, like us on Facebook or follow us on Twitter for breaking stories and campaigns throughout the day.