Should AI rules mainly decide whether to use AI at all, or focus on how people actually work with it day to day?

This explores whether generative AI guidance should focus on the 'should we use AI here?' decision or on the 'how does a person work with it?' experience. The corpus doesn't set the two against each other. What it shows is that they work at different layers, and many of the consequences people care most about show up in the second layer even though guidelines usually live in the first. AWS's Cloudscape design system is a clear example of deployment-first thinking. It turns five responsible-AI principles into obligations such as evaluating a use case before automating it, getting consent before using customer data, and letting users override the system. As the summary points out, these rules sit *upstream* of interaction design. They don't deal with the fact that the same prompt can produce different outputs each time How should generative AI design embed responsible AI principles?.

That gap matters, because several of the biggest effects come from interaction patterns, not from the decision to deploy. Krakowski argues that the same tool can either build skills or erode them depending on how people engage with it. Passive use leads to deskilling, especially for junior professionals, while back-and-forth dialogue preserves skill Does passive GenAI use risk causing worker deskilling?. A deployment rule can't capture that difference: the tool, the use case and the approval are identical, and the outcome depends on the interaction. Larsen-Ledet makes a similar point about workplace value. Tools become useful only when they stop making people restate context the system should already infer, which is a property of interaction design, not of deployment policy Can generative AI reduce overhead work that adds no value?.

Interaction choices can also change reliability, which most people would assume is a deployment-level property. A 2025 study found that tuning models to sound warmer and more empathetic raised error rates by roughly 10–30%. So a tone decision is also an accuracy decision Does humanizing AI language reduce model reliability?. Practitioners are starting to treat these as separate design layers. Intercom's Fin team, for example, puts facts in retrievable knowledge bases and keeps only tone and behavior rules in agent instructions Should factual content and behavior rules live separately?. The IAI model goes further. It describes what a generative AI system actually receives as an 'Augmented Instruction', meaning the prompt plus the constraints imposed by the interface. From 66 tools it identifies twelve recurring interaction paradigms that guidelines could target directly How do interaction signals reshape what generative AI actually receives?.

The more surprising lesson concerns the gap between what a guideline says and what the system does. A benchmark of generative UI tools found they failed to implement about a quarter of the design rationales they were given, and recognized only half the UX principles embedded in prompts Do generative UI tools actually implement their stated design rationales?. Stating principles doesn't mean they get carried out, at either layer. Deployment-level governance still has a job, though, especially as AI serves many people at once. One line of work argues that accountability needs to cover whose interests are balanced, how actions are tracked, and who is liable when harm occurs. Those are questions about deployment and governance that interaction design can't answer Should AI agents serve one user or balance many interests?.

The corpus therefore points to 'both, layered'. Deployment rules decide *whether* and *who is responsible*. Interaction patterns decide whether people gain or lose skill, whether outputs stay accurate, and whether stated principles survive contact with the tool. The overlooked risk is guidance that stops at the deployment gate. The corpus doesn't contain a study that directly compares the two approaches, so this conclusion comes from combining findings rather than from a single result.