“AI Slop” - When Optimization Metrics Replace Human Readability

There’s a growing pattern of using tokens to generate AI code and documentation slop. Then use even more tokens to understand and review that slop.

Then judge engineers by token usage instead of how empathetic and clear their docs and code actually are

At some point, the system starts optimizing for the wrong thing. Instead of asking “Can a human actually work with this?”, we continue asking “How much did we generate?” or “How many agents did we spin up today?” - are those the success metrics we want?

A clear example of this is what we’re seeing in AI-generated UIs for landing pages. Tools like Claude (and others) can produce interfaces quickly, but they often converge into a very recognizable template. Same layout patterns, same spacing, same visual language. It becomes less “design” and more “average of all designs the model has seen.” 

Maybe we should be measuring success differently:

• How readable is the documentation without context?

• How long does it take a new engineer/PM to understand and contribute?

• How often do humans need to rewrite or simplify what the AI produced?

• Does the output reduce or increase cognitive load?

Instead of another “50 agents built before breakfast” post, it’s time to also ask: what is the human cost of maintaining all of it? what should we actually optimize for? Token efficiency, or human understanding?