Audiobook creation service Findaway Voices has become a popular alternative to Audible’s ACX, especially in the wake of #Audiblegate (the controversy over ACX’s author-penalizing returns policies that has generated at least one lawsuit).
In the past few days, though, authors and narrators have been drawing attention to this paragraph from Findaway’s Digital Distribution Agreement, which grants Apple–a third party–a license to use the rights holder’s audiobook files for “machine learning”, aka AI training:
This paragraph doesn’t appear to be a new addition: I’ve seen an agreement from September 2022 that includes it. But I can’t find anything to indicate when it may have been incorporated, or any discussion of it prior to a few days ago. (UPDATE: In the comments and also on social media, people are confirming that the paragraph was part of the agreement as far back as 2019.)
If you’ve been paying attention over the past couple of years to the increasingly intense debate over generative AI (algorithms that can generate humanlike content in response to prompts from users)–especially recently, as programs like DALL-E, Midjourney, and ChatGPT have been released to the public–you may be aware of the wide range of opinion on the benefits and dangers of this new technology, as well as its implications for the future.
For example, ChatGPT, a tool for creating text, can help a dyslexic business owner to compose well-written emails and scientists to improve research papers–but also raises the specter of academic cheating, writers replaced by machine-written books and articles, and the dissemination of false content–since ChatGPT and similar chatbots are text-creators but not fact-checkers, and can spit out erroneous answers or disinformation as convincingly as true and accurate responses.
In the creative industries, writers and artists are experimenting with ways of working with AI tools–but there’s also much concern about displacement. Voice actors and audiobook narrators worry about being supplanted by machine voices, like those in the catalog of AI-narrated books recently released by Apple. Artists and designers fear losing their livelihood to art-generating programs like Midjourney and Stable Diffusion that can create beautiful art in seconds, while the companies behind them make a fortune on new technology and publishers and others save a bundle by not employing humans.
There’s also–and importantly–the underlying issue of rights. Generative AI “learns” by ingesting massive amounts of whatever content it’s designed to mimic: chatbots scrape text from the internet, art generators gobble up millions of images, voice AI trains on human voices. All of that material has been created by humans, who in most cases never gave permission for such use. That’s obviously not an issue for content not covered by copyright–paintings by Renaissance artists or the poems of Edmund Spenser–but it’s not like the companies behind generative AI make that distinction. In-copyright work can be, and is, as readily scraped as anything else.
Unsurprisingly, there are now multiple lawsuits. Microsoft, Github, and Open AI are being sued for copyright-related issues over Github’s AI-powered coding assistant, Copilot. A group of artists has filed a class action suit against Stability AI (owner of Stable Diffusion), Midjourney, and DeviantArt for copyright violation and unlawful competition. Getty Images is also suing Stability AI, alleging that it scraped millions of copyright-protected images from Getty’s database. On a different side of the issue, computer scientist Steven Thaler is suing to overturn the US Copyright Office’s determination that AI art can’t be copyrighted. There will no doubt be much more legal action to come.
The above barely brushes the surface of this extremely complex and rapidly developing field (for a much deeper dive, see Jason Sanford’s article on what generative AI portends for artists and authors). It’s an altogether thorny tangle of questions: copyright, infringement, moral rights, fair use, the nature of creativity, and more. In terms of AI training, we may be seeing the emergence of a whole new class of rights that creators and their representatives will one day routinely negotiate into contracts. It’s reminiscent of the early days of ebooks, when some people argued that they weren’t really books at all, there was no consensus on contract language or royalty rates, and publishers were claiming that electronic rights weren’t separately salable but rather an additional print format grandfathered in under the existing rights grants in pre-digital publishing contracts.
Regardless, rights holders need to be fully informed–and properly compensated. And relative to that, Findaway Voices’ machine learning license poses several issues of concern.
- Rights holders are being asked to license their audio files for unlimited use–and, by implication, commercialization–by Apple, free of charge. Such licenses should be paid.
- Findaway’s narrator agreement includes no mention of machine learning. Narrators aren’t rights holders, unless they are authors narrating their own books: Findaway’s narrator agreement is work-for-hire. But in terms of knowing what you’re signing up for, having your voice used for AI training is an important data point, and it’s possible that some narrators might choose not to sign up with Findaway if they were aware of this requirement. It should be incorporated into the narrator agreement, or otherwise disclosed.
- The Findaway rights holder agreement is a strictly a distribution agreement: there’s no language empowering Findaway to license subsidiary rights to third parties. Yet that’s exactly what’s happening here: the licensing of a subsidiary right to Apple via the rights holder’s agreement with Findaway. Am I wrong in thinking that possibly presents legal questions?
- The machine learning clause should appear in the body of the agreement, rather than way down at the end under Program Policies (where, whether by accident or design, it very clearly can be easily missed). A rights license is hardly just a “program policy”.
As allowed by the license language, rights holders have been contacting Findaway to opt out. They receive this response:
As far as I know, no one has heard from Apple yet. Some writers are also trying to find out if ACX has a similar machine learning policy, but so far ACX isn’t playing.
The fact that the license has been included in Findaway’s distribution agreement for at least several months, but is only now generating discussion, really points up how important it is to read and understand every word of a contract before entering into it–even addendums.
I’ll be following the situation closely, and will post updates as I receive them.
UPDATE 2/14/23: Responding to questions from members, National Director for Audiobooks Jane Love and others from SAG-AFTRA met with Findaway Voices last week to discuss the machine learning clause. Here is their update, which reports some limited good news.
Over the last week or so, we’ve heard from a number of you regarding a clause in Findaway’s terms of service agreement with authors/license holders that authorized Apple to use the material for machine learning training and models unless the author/license holder specifically “opted out”. First, many thanks to all of you who brought this to our attention. We met with the Findaway team last week, outlined SAG-AFTRA’s position that performers’ voices may not be used beyond the distribution of the audiobook(s) they consent to narrate. Further, we made clear that any language attempting to acquire permission from performers or in this case the author/publisher for machine learning, or digital simulation, is unenforceable until bargained directly with SAG-AFTRA. And that the terms and conditions of these rights are mandatory subjects of bargaining. As such, we told them that we do not consider any such rights transferred. The dialogue was very positive, and we are pleased to report that Findaway and Apple immediately agreed to halt any/all use of files for machine learning purposes. This “halt” covers all files dating back to the beginning of this practice. We will now focus on bargaining with Findaway over fair terms around any future use of material for digital training purposes to include appropriate consent.
Further, we are seeking information from our other collective bargaining partners in this regard. And while we can’t say more given confidentiality, we can assure you that Audible/ACX has not granted any AI rights (including machine learning rights) to any party.
Finally, we saw a number of communications around this – blogs, social media posts etc that contain quite a lot of misinformation both around what was occurring and performers’ rights in this area. We would encourage you to be discerning consumers of online content, and to reach out to us with any concerns.
This is a challenge for sure, but together, we truly believe that we can continue to grow human storytelling in audiobooks while also embracing new technology. And SAG-AFTRA is here to protect you. We are doing this on multiple fronts, including collective bargaining, legal advocacy, legislation, and education.
On the collective bargaining side, we are continuing to work with our partners to negotiate the right contract for use of synthetic voices in audiobooks and are ready to negotiate in good faith with any outside company interested in licensing our members’ voices.
On the legal and legislative front, SAG-AFTRA tracks relevant cases, educates artists about existing image and voice rights (rights of publicity), and advocates for legislative improvements in image and voice rights as needed.
Thank you to all o the members of SAG-AFTRA whose membership and commitment to working under union contracts make these efforts possible.
In Unity, Jane Rich and Andrea
National Director, Audiobooks
UPDATE 3/2/23: The Authors Guild has added an AI training clause to the Grant of Rights section of its Model Contract, prohibiting publishers from using or licensing the works they publish for machine learning.
UPDATE 3/4/23: Today I was contacted by an author who just uploaded an audiobook to Findaway, and discovered that the distribution agreement still includes the machine learning clause. According to SAG-AFTRA, Findaway and Apple agreed to halt all use of files for machine learning purposes, a halt that “covers all files dating back to the beginning of this practice”–but while apparently halting use for files already shared with Apple, could this statement provide a loophole for the use of newly uploaded files?
UPDATE 5/3/23: As of this writing, the machine learning clause is still in the Findaway distribution agreement (link above). One author who contacted Findaway about this reports the following timeline of questions and responses (I’ve seen the emails paraphrased in the timeline):
AUTHOR (Feb 7): I was made aware of this section of the contract where Findaway gives Apple the right to train their AI narrator on my audiobooks — I do not want this. Please remove that from the contract. You do not have my permission to allow Apple (or anyone else) to train their AI narrators on the audiobooks which I distribute through Findaway.
FINDAWAY (Feb 9): Thank you for your request. We have submitted your opt-out request to Apple per the Machine Learning clause in the distribution agreement.
AUTHOR (Feb 9): I would like a new contract with Findaway that explicitly says those rights are not granted to Apple (or anyone). The current contract says I may revoke the licensing of these rights to Apple. It does not say that I may “request” it (implying that Apple has the option to deny that request).
**May 1**: Findaway sends out a notice that they’ve updated their Distribution agreement for various things unrelated to machine learning with Apple. I checked: machine learning language still there. I’m supposed to “view and agree” to this new contract.**
AUTHOR (May 1): I requested removal of machine training in my contract back in February, but received no response — it’s not adequate to say that you’ve submitted an “opt out request” to Apple. The contract needs to reflect that. And now that I see you’ve updated the contract for changes with Spotify, I’m renewing my request that you update my contract to not include the language about machine learning.
FINDAWAY (May 1): Thank you for your feedback. Our contract is not negotiable, but I can assure you that your audiobooks are still opted out with Apple.
**I check and see that my audiobook distribution to Apple has been discontinued**
AUTHOR (May 1): Are you saying that the only way to not have my audiobooks subject to machine training by Apple is to not have them distributed to Apple?
FINDAWAY (May1): Nope, not at all. All Findaway Voices audiobooks are opted out of Apple’s Machine Learning Program. This does not impact distribution to Apple Books.
AUTHOR (May 1): If all Findaway Voices audiobooks are opted out of Apple’s Machine Learning Program, then why is the machine learning clause still in the (newly updated) distribution agreement?
The author reports that they have not received a response to their final question.