In the last couple months, there's been so much written about AI generally and large language models specifically (ahem, Time Magazine cover...) that the basic logic of these tools is not much of a mystery. In my layman’s understanding, the way a lot of popular AI apps work is that they’re trained on huge datasets of text, audio, images, etc., and then used to generate new content based on the patterns they've learned to recognize and predict. This result is the impressive applications that’ve captured the zeitgeist: automated translations, generative music and image apps, and of course the lifelike conversations and cogent college term papers outputted by the likes of ChatGPT.
Like everyone else, I've been fascinated by these developments, and it led to the following 3 a.m. thought: If you're looking for an AI to generate content related to music—from an artist biography to a list of songs to more sophisticated requests—to what extent is the resulting content that gets hurled out based on work that I or my colleagues have had a hand in creating?
Perhaps this seems absurd on its face—who would ever claim to have made such a significant contribution to a knowledge base as vast as Recorded Music? But bear with me for a second... For completely unrelated reasons, TBC recently decided to tally up the number of content pieces we've created in the nearly eight years (as of this writing) that we've been doing this work. Get a load of these numbers:
- 12,571 Artist biographies, totaling approximately 4.6M words.
- 21,637 Original Playlists and Playlist Updates
- 58,504 track evaluations, which means our team listened to each track, adding metadata and assessing it for inclusion in various playlists, stations, and collections
- 2,275 blog posts and feature articles
- 32,783 in-product descriptions of playlists, albums or tracks
Now maybe you're impressed or maybe not, but that's just the work that our company itself has delivered to clients. What if you consider the collective output of our contributors? Electronic music shaman and founding TBC contributor Philip Sherburne publishes in the ballpark of 130 articles a year for Pitchfork, and has since 2014. Pop savant Maura Johnston — who's created 3,195 individual assignments (!) for TBC since our founding — is similarly prolific, publishing upward of 200 pieces a year for places like The Boston Globe, Entertainment Weekly, and Rolling Stone. These are just two examples of the roughly 300 folks around the world who work on our projects. And those are just the online articles: Several of our contributors are published authors. So yeah, do the math... I don't think it's a stretch to suggest that the work of Third Bridge combined with its contributor network comprises a non-trivial percentage of all contemporary music writing. I'm not saying it's a large percentage, just that it's statistically significant.
There is currently a series of class action lawsuits being filed against a handful AI companies alleging copyright violations, and these could have serious implications on the future of AI-generated content. In the case of a painter like Kelly McKernan, the legal grounding for such a claim is more obvious: People are using a generative AI tool such as Midjourney to request images in McKernan's style using McKernan's actual name. No one's out there asking ChapGPT to "Write a Kendrick Lamar bio in the style of Third Bridge Creative," not least because we nearly always provide our services on a white-label basis, meaning anonymously as far as anyone scraping the internet is concerned. And look, I'm not trying to compare apples and oranges, or suggesting any kind of infringement in our case. I'm merely pointing out that I'm not the only one wondering about the provenance and underpinnings of some of these datasets.
It gets even weirder when I think about how long we’ve been doing this work. Long before Sam and I started Third Bridge, we worked together at a company called Rhapsody, which as Wikipedia will tell you "was the first streaming on-demand music subscription service to offer unlimited access to a large library of digital music for a flat monthly fee." At Rhapsody, we were part of a staff of music experts whose whole job was to catalog the entire universe of recorded music. The colorful history of this group and the way it essentially wrote and curated (and partied) its way to laying the foundation of streaming music is a story for another day, but the gist is we collectively wrote millions of words, programmed thousands of tracks, made zillions of genre and artist associations (which today would be called "tagging"). Some of this data made it out into the ether of the larger internet and some of it remains entombed on some server, and we'll probably never know which is which. But then, who knows how these mysterious datasets that get fed to the AIs are themselves unearthed and organized?
Back in the Rhapsody days, there was this beloved writer named Mike McGuirk, a genius wordsmith and passionate music fan who by his own admission would have remained a line cook had a friend not recruited him to join an early incarnation of the team. McGuirk was one of the most prolific writers we had, who wrote upwards of 30 blurbs a day every day for several years, covering everything from Cher to Lightnin' Hopkins to Florida death metal legends Deicide. And Mike was kind of your '80s-Bill Murray-type irreverent joker, and so every now and then he'd insert either a subtle wink or blatant non sequitur into one of his blurbs, which was certainly not allowed but which we all secretly got a kick out of. One of my favorites is the line he appended to a review of a record by the oft-derided '90s nu-metal group Creed, beautiful in its simplicity: "Wrestling is fake."
Now, I'm certainly no expert in how large language models work, let alone AI algorithms writ large. But based on everything I've just explained, it seems at least possible that me and Sam and the cast of characters we've had the great pleasure of working with these last 20 years have had some kind of hand in shaping the output of whatever music-related request or question someone asks an app like ChatGPT. And if that's true then it seems likewise possible that deep inside the gray matter simulacrum that is this emerging new technology, there exists the sensibility of Mike McGuirk. So if you detect a hint of sarcasm the next time you ask a chatbot a question about Cher, then perhaps that's where it comes from.
.png){kind=small}