Why do specific tools like chatbots or coding assistants get their own classification systems that still fit inside one big map of AI interaction?

This explores why classification schemes for specific GenAI tools seem to slot neatly inside a larger design framework for human–AI interaction. The corpus doesn't hold a paper that argues for nesting directly. It does have the pieces to explain why it happens: broad frameworks describe *dimensions* of a design, and narrow taxonomies describe *values* along those dimensions. The clearest broad framework comes from a review of 291 papers. It found six recurring dimensions in human–GenAI interaction design: why people use the system, how the model gives feedback, how much control users have, how deeply users engage, which domain the tool serves, and how it gets evaluated How should designers structure human-GenAI interaction systems?. These are questions any GenAI system has to answer, which is why a narrower taxonomy for, say, writing assistants usually turns out to be a detailed breakdown of one or two of these axes.

The IAI model shows why nesting works better at some levels than others How do interaction signals reshape what generative AI actually receives?. Its key move is to stop asking what the interface looks like and instead ask what the model actually *receives*. That's the 'Augmented Instruction': your prompt plus all the constraints the interface quietly adds, such as sliders, selected regions and templates. Across 66 tools, this lens reduced the variety to twelve reusable interaction patterns that combine like building blocks. Here's the useful bit: earlier tool-specific taxonomies couldn't be compared with each other, because each one described surface features unique to its own tool type. Once you describe tools by what reaches the model, they become comparable, and narrow taxonomies become special cases of a shared vocabulary instead of separate islands.

There's a deeper reason hierarchies like this feel natural, and it shows up somewhere unexpected. The internal geometry of language models arranges concepts coarse to fine. The strongest directions in embedding space split broad categories first and finer sub-branches later, tracking the WordNet hierarchy level by level Do embedding eigenvectors organize taxonomy from coarse to fine?. Placing knowledge inside a taxonomy also does real work. StructTuning reached half of full-corpus performance with 0.3% of the data, because the model learned where each piece of knowledge sat in a conceptual structure Can organizing knowledge structures beat raw training data volume?. Nesting isn't just tidy bookkeeping; it reflects how concepts actually organize.

The counterpoint is that nesting has limits, and the corpus is sharp on this. Fairness research found that frameworks built for specific, structured tasks *don't* scale up to general-purpose LLMs. They either stop making logical sense or become unmanageable across endless contexts, so fairness has to be worked out case by case Can fairness frameworks extend to general-purpose language models?. Nesting runs one way. A specific taxonomy can live inside a general framework, but you can't inflate a specific framework into a general one. Generative UI is a good stress test. When the AI decides what interface to build for users who can't describe it themselves Can generative UI serve users who cannot specify interfaces?, it blurs the 'user control' and 'model feedback' dimensions together. Edge cases like this are where a broad framework may need new axes rather than new sub-branches.