This is not an official policy.
Rather, it is a statement explaining FAWM’s perspective on artificial intelligence (AI) and machine learning (ML) technologies in the platform, on AI-based tools for music creation, and expectations for how we think these could and should be used in the context of our songwriting challenges.
FAWM was actually founded as a musical side-project by an AI/ML researcher, so we are quite excited about AI and its potential for helping you make music. As far back as 2010, we even created "The Muse" at muse.fawm.org as a (nascent) exploration into AI-based songwriting tools for the FAWM community (research paper here).
We plan to integrate more AI-based personalization into the FAWM platform in your home feed, recommendations, lists/rankings, creativity tools, and so on.
Most online platforms today train AI systems to maximize your attention and time spent on the platform. FAWM is different. We are funded by your support (not ads or venture capitalists), so our systems will be trained and deployed to try and optimize your songwriting output, collaborations, and other goal-oriented and pro-social behaviors (see our previous research on this here and here).
As AI systems are deployed at fawm.org, this document will be updated and we plan to provide model cards (think “nutrition labels” for AI) to be as fair and transparent as possible about how your data are used, and what role AI plays in your experience.
More complete documentation to come, but for now FAWM uses two ML models:
A user spam classifier available to moderators.
This is trained on historical data to predict whether and account was suspended or not, as a function of its profile contents. The area under the ROC curve (AUC) for this model is 0.98, which is near perfect. Account suspension is still a manual process, but this model helps mods prioritize.
A song popularity classifier used for "magic" sort on the songs page.
This is trained on historical data to predict whether a song will receive 5 or more comments (5 is the median comment count for songs in both FAWM and 5090 challenges). The AUC for this model is 0.86, and rank correlation with actual final song count is 0.70, both are statistically significant (p < 0.001). However, this "popularity" score is not used as a ranking — that would be against the ethos of FAWM! Instead, "magic" sort is a function of the AI's probability score and the actual number of comments received. High probability makes songs bubble up, more comments make them bubble down. In other words, "magic" sort helps prioritize songs that so far have gone “under-appreciated.” You can always switch back to "chronological" sort if you prefer.
FAWM sends no data to third-party AI APIs, and uses no third-party models, except for occasional (offline, local) use of Stable Diffusion for help generating visual assets for social media content and weekly challenge illustrations.
Our position is that AI-based music tools are exactly that: tools. As with pitch correction, drum plugins, MIDI libraries, rhyming dictionaries, samplers, or even a capo — there are varying opinions about whether using a particular tool is “cheating.” In principle, FAWM encourages you to use any tool at your disposal that helps you grow as a songwriter and a musician (including AI).
That said, we expect fawmers to act honestly. This means that you should not:
We intend to develop new “computational creativity” tools, and expand offerings at muse.fawm.org, including offerings based on more modern AI than what is currently there. Our philosophy for these is to make interactive tools to help you initiate or develop your own musical and lyrical ideas, rather than “end-to-end” AI music creation bots.
During the February 2023 challenge, we saw a surprising split within the FAWM community over the use of tools like ChatGPT. The capabilities of publically-available AI-based products for generating lyrics, music, and even entire audio tracks will only continue to grow.
We ask anyone using AI-based tools to explain how you used them in your “liner notes” and perhaps use an appropriate hashtag. In fact, we encourage fawmers to discuss any and all tools (AI or otherwise) used in the liner notes, so that fellow fawmers can learn from your process!
There is an invisible line where, on the one side you use AI as a legitimate tool, and on the other side it is a crutch. FAWM cannot tell you where that line is, but we are asking fawmers to reflect on how they're engaging with such tools when participating in these challenges.
This is perhaps the murkiest and most controversial topic in generative AI and art right now. For a good primer, we recommend this testimony from the U.S. House Judiciary Committee hearing on AI and Intellectual Property.
In short: FAWM can and will remove any media that violates copyright per our existing copyright policy. Furthermore, we reserve the right to remove any content that uses AI to try and represent the likeness (including the voice) of another artist or copyrighted work/entity (e.g., “deep fakes”). While one can make the argument that such works are part of one’s own creative practice, there are also many other outlets on the Internet for such things, so why use them for FAWM?
So we ask that you remain focused on original songs or compositions that are yours, and that you represent to be your own, for the purposes of FAWM challenges.
If you are concerned about your music or lyrics being scraped by bots to train their generative AI systems, we made two changes in 2023:
noai directive (more info here) to all pages of fawm.org, to tell webbots not to crawl or cache FAWM pages for training AI systems. Of course, these directives work on the honor system and those developing AI tools must voluntarily opt to follow protocol...