Don't build a business on AI getting worse

The premise is real. I'm not disputing it. Code assistants produce subtly broken code that still compiles. Writing tools produce plausible prose with factual errors nobody catches on the first read. Image generators still do strange things with hands. Many of today's "AI detection" and "AI output cleanup" products address genuine, frustrating, high-volume problems.

The bet that has to fail

The trouble is that a cleanup business requires the failure mode to persist. It needs the model to keep hallucinating, the coding assistant to keep inserting subtle bugs, the image generator to keep producing its signature tells. The business dies the moment the frontier model gets noticeably better at the specific class of failures your product targets.

Every year since 2020 the underlying bet has got worse. The error rates that founded an entire generation of "AI fact-checker" startups in 2023 are roughly an order of magnitude lower in 2026. The specific classes of bug that "code review AI" tools are catching today are the ones the next major model release will not introduce. Each generation of the underlying platform shrinks your total addressable market from the inside.

You don't need the model to be perfect for the cleanup pitch to fail. You just need it to get good enough that the cost of cleanup becomes difficult to defend against the cost of simply using the better model.

Ride the wave of AI's improvement, not the gaps of its early stage.

What to build instead

The more durable version of the same insight is almost the opposite. Don't build the feature that AI today fails to deliver. Build the feature that AI today just-about delivers, and make it reliable, governable, and embedded in a real workflow that someone would pay to keep.

There's a useful analogy from the early cloud era, roughly 2010. A wave of startups pitched themselves as "disaster recovery for ephemeral cloud instances". Their thesis was that cloud VMs were fragile and would keep falling over. They were right, until they weren't. The durable businesses from that era were the ones who assumed the underlying infrastructure would become reliable, and who bet instead on orchestration, observability, managed services on top. The ones who kept pitching "cloud is fragile, we catch the pieces" disappeared into a market that shrank beneath them.

The AI equivalent today is to assume the models will improve faster than your startup can iterate, and to build a product where the model's current behaviour is the floor of what you offer, not the ceiling. Everything that gets better as the model gets better is free roadmap. Everything that gets worse is existential.

A filter for your roadmap

One useful test to apply when you're evaluating an AI feature you're considering shipping: does this feature rely on the model staying bad at something, or does it get better as the model gets better?

If the first, ship it carefully and with a short roadmap horizon. Assume it will look dated in eighteen months and plan the pivot already. If the second, invest seriously. The second category is the category that compounds.

The slop is real and it's worth cleaning up in the short run. You don't want to be the specialist cleaning up a mess that the people making the mess are actively, quickly, un-making. Build on the improvement curve. The mess is temporary. The capability isn't.