Identifying a good metaphor is a superpower that all good product marketers have. (See what I did there?) Being able to abstract a complex technical concept and find a metaphor or mental model that grounds it in something familiar is how we figure out how to talk about it. I’ve picked apart a bad one before, and I’ve sat in meetings where the metaphors were actively making things worse.

For the past few years, I’ve been trying to find the right metaphors for various AI concepts. Most recently, I’ve been thinking about how if the vast majority of content that is being created is AI-generated, at what point will there be no new raw material for AI to use? At what point will all its content just be built on its own generated content?

The first thing I learned is that there’s a word for this (“model collapse”) and a metaphor (“Habsburg AI,” coined by researcher Jathan Sadowski, referring to the European dynasty whose Spanish line produced offspring with progressively greater levels of genetic frailty due to inbreeding). There are differing opinions among AI researchers as to how much of a threat model collapse really is, but it is debated as part of the long-term impact AI will have on our society.

What’s interesting is that some researchers think that as long as AI-generated content is created alongside net-new human content, the threat is low. One paper found that if you keep the real data and pile synthetic content on top of it, rather than swapping one out for the other, the degradation has a ceiling it doesn’t climb past. Other researchers see it as a strong concern as the volume of new AI-generated content significantly outpaces that of new human-generated content.

I wanted a more thorough metaphor so I could form my own point of view.

A sourdough starter. The first metaphor I thought of was a sourdough starter (shout-out to 2020). You keep a jar of flour and water alive by feeding it forever, and every loaf comes from the last loaf’s starter. It never runs out, and over years it picks up the wild yeast in your ingredients and environment. But it fails as a model for content quality, because a starter has no standard it can fail to meet. Whatever yours becomes is, by definition, yours. It might get crustier or sweeter, but there’s no measure of if it gets worse.

Dollar-cost averaging. The next metaphor I thought of was this financial concept, where you buy the same amount of a stock or other instrument every month regardless of price, so that no single peak or valley in the market determines your outcome. At first blush it’s a good model, because its purpose is to smooth out variation over time. But it only describes the shape of the result. There’s no mechanism inside it to inspect, and nothing I can point at to say what would go wrong, or when.

A solera. I think I’ve landed on this one, which is a system describing how sherry is made. Imagine a structure of barrels stacked in rows by age. Each session, you take sherry out of the lowest barrels to bottle, and add new wine to the top. When you remove the sherry, that barrel refills from the row above, and that one from the row above it, all the way up. The bottom barrel is never emptied, only diluted, which means from year to year, the composition of a new bottle of sherry is similar. As a machine for cancelling variation, a magnificent year and a miserable year can taste the same in the glass.

Which brings me to the one part of the solera that doesn’t map to our content metaphor: the capataz. The capataz (“cellar master”) walks the rows of the bodega and tastes the barrels. One working capataz describes wearing out the soles of his shoes going through each solera again and again, tasting every cask of every style, marking them in chalk as his predecessors did.

Because theoretically, even though a solara is designed for consistency, over time it can drift. Ten years in a row of a different harvest will eventually show up in the glass, though it will happen gradually enough that no single bottle gives it away. The system can’t account for that, but a person tasting can notice it. The same is true for content. If we want AI-generated content to stay good, we need to keep creating, publishing and paying for new human work in enough volume that it’s still shaping what comes out. And someone has to be paying attention to whether it is.

Now here’s an uncomfortable part. Right now, with AEO and GEO, many of us are optimizing content for the LLMs that index it. We’re writing for what the model will pick up, cite and repeat. In effect, we’ve made the system its own capataz. The taster deciding what’s good is the same machine whose output we’re worried about. I understand why we’re doing it; it’s where buyers are starting to look. But I’m not sure it’s what we want. A solera works because the person tasting cares about what the wine should be.