Why AI can’t make a tusk or a unique playlist.
In a recent New York Magazine article, the author comes to the realization that Spotify’s Discover Weekly playlists are not, in fact, particularly unique to the listener after hearing a good portion of their “personalized” songs played in a local bar. Discover weekly seemed like a service that could magically take ALL the music a listener had ever played, dump it into a black box, and spit out a one-of-a-kind list of new music, perfectly matched their unique listening DNA. In fact this is not how Spotify Discover weekly, or any of the other algorithm-driven applications actually work. They work by taking massive amounts of data and averaging it in order to provide the “best” results. The problem is that averages are weighted against novelty.
Discover Weekly is unlikely to ever recommend a brand new song by a brand new artist, even if it’s something that a certain listener would 100% love, because Spotify has no data to back up that decision. Their recommendations are not based on the contents of the songs, they’re based who else likes the songs, and those numbers are always going to veer towards mass because mass equals consensus in the eyes of a bot. Users are finding similar issues when they query Chat GPT whose responses are simply an average of whatever is in its database, regardless of veracity. If there are enough blog posts in that database questioning whether the moon landing happened, Chat GPT will also question whether the moon landing happened. Something to keep in mind when using Chat GPT to write your term paper on the Apollo missions. The same issues apply to AI image generators like Mid-journey. They’re great at creating images that look like other things we’ve all seen, but they’re terrible at originating more novel images. Walrus tusks, for example, are very difficult for Mid-journey to draw. I know this for obvious reasons. Mid-journey has no clue that there are lots of walruses in captivity who have had their tusks removed and that they are not the platonic ideal of the way a walrus should look. It just knows it has x number of walrus pictures and some have white things on their faces and some don’t. As a result, it kind of splits the difference. All of these algorithms can be tweaked, and weighted to make them more efficient, but fundamentally they exist to provide consensus answers, not novelty and that is why they will never be a substitute for human creativity.
Human creative choices and tastes are hard to replicate because they are so individual and unique. For years at the agency we’ve run into a similar issue whenever we need to use stock photos to generate comps for ads. The images available in stock libraries are generic and not up to the task. It’s often impossible to pull together a decent representation of what we’re envisioning in Photoshop because the visual is something entirely new. For a photograph to make it into a stock library, it has to be useful across lots of applications so the stock library can make money off of it. This makes most of those images useless to someone with a new idea. We end up using a drawing instead, because it’s much easier than cobbling something together from disparate existing photographs. In fact, we’ve found that this is a good check on the freshness of an idea: if you can find really good reference for it, it’s probably not a very original thought. Initially we had thought that Ai would be a boon to this type of photo comping because we could finally generate quality representations of visually unique ideas. Sadly it’s not the case – no matter how hard you try and manipulate the query, it’s impossible to generate a good image of, say, a 900 foot walrus rampaging through Union Square like we have on our holiday t-shirt. Ultimately, AI is extremely useful in surfacing the aggregate of what’s already been thought, photographed, drawn, discovered, but if you’re looking for the NEXT thing, you’re better off talking to a human.