How to read this.
"AI slop" is the texture of writing that has the shape of thought without the substance. It is not that the model is wrong, exactly. It is that the model is producing the silhouette of careful prose (the rhythms, the bolded keywords, the reflexive hedges) without the underlying decisions a person would make to produce that prose for a reason.
The simplest frame is from machine learning: the model has overfit to surface features that the training data and human raters rewarded, and underfit to your actual intent. It learned that em-dashes correlate with sophistication, that "delve" correlates with seriousness, that hedging correlates with safety, that bullets correlate with organization. So it produces all of those features regardless of whether they are warranted by the specific question you asked.
From philosophy of mind: the model is doing extremely competent pattern-matching against the distribution of human writing, but the patterns are not anchored in a model of you, the topic, or the world. It is mimicry of expert register without an expert's mental model of why expert register is shaped the way it is. The cadence is real; the noticing that produces the cadence is missing.
This is not an indictment. It is a description of the failure mode that comes with the architecture: prediction over a high-dimensional surface, shaped by reward over short-context human ratings. Every entry in this glossary names one specific way that failure mode shows up, explains the training-time mechanism behind it, and gives you a sentence to drop into a system prompt that suppresses it.
How to use the glossary
- Browse the glossary. Filter by category or severity. Each card opens a detail page with the full breakdown.
- Hit the + on any card (or "Add to prompt" on a detail page) to collect it. The floating button bottom-right shows your count.
- Open the collector to preview and copy a combined system prompt assembled from your selections. Paste it into any chat model.
- Adjust as you go: drop entries that don't apply to your use case, keep the ones that consistently bite.
Adding patterns
The glossary is a single TypeScript file. To add a pattern, append one object to src/data/slop-patterns.ts. The list, search, filters, detail routes, and collector all derive from that array. There is no admin UI and no database, by design, so the data stays diff-able and reviewable.