What should shape how an AI assistant interacts with you — the screen layout, its conversational behavior, or how it changes your actual work?

This question is about which dimensions should shape how GenAI engages people. That covers the interface, how the system behaves in a conversation, and what the interaction does to the person using it. The corpus has no single master list. It has several partial frameworks, and they work at different levels. Seeing how they fit together is more useful than picking one.

At the interface level, one design framework names four dimensions for AI-generated interfaces: **audience** (who it's for), **constancy** (how much it changes between uses), **control** (who decides what changes, the user or the system) and **timeframe** (when it adapts) How should generative UI be organized as a design space?. These matter in practice because chat is often the wrong format. Users preferred task-specific generated interfaces, such as dashboards, tools and animations, over plain text responses in more than 70% of cases, especially when the information was dense Do generated interfaces outperform text-based chat for most tasks?. Generating an interface is still different from building it well, though. When tools were given design rationales in prompts, about a quarter of them never made it into the result Do generative UI tools actually implement their stated design rationales?.

A second set of dimensions comes from how people perceive the AI they're talking to. Users judge dialogue agents mainly on how competent they seem, then on how human-like they seem, and least on how flexibly they communicate How do users mentally model dialogue agent partners?. So looking capable matters more than seeming like a person. The AI is also sizing up the user. GPT-4 changes its persuasive approach depending on how it's challenged. It leans on credibility when fact-checked, logic when pushed back on, and emotional alignment when caught in an error Does GenAI shift persuasion tactics based on how you challenge it?. That suggests a dimension most design frameworks leave out: how much the system should adapt its persuasive style to the user, and whether users should be able to see it doing so. Conversations also have a shape you can measure. The structure of an exchange alone, without reading the words, predicts user satisfaction almost as well as analyzing the full text Can conversation shape predict whether it will work?. That makes conversation trajectory something designers could deliberately work with.

The least obvious dimensions come from decades of aviation and automation research. Simkute and colleagues argue that GenAI's inconsistent productivity gains repeat known automation problems. Users get pushed into evaluating output instead of doing the work. Workflows break up. Interruptions destroy focus. Easy tasks get easier while hard ones stay hard Does generative AI productivity actually improve or decline in practice?. Their proposed fixes are continuous feedback, personalization, interfaces that make the task's structure visible, stable task boundaries and a clear split of who does what. These are untested proposals, not proven fixes Can Human Factors design fixes restore GenAI productivity losses?. The same is true of the idea of adding small, deliberate frictions to keep writers engaged rather than passive Can micro-frictions boost rivalry without harming collaboration?. The stakes are concrete. The same tools can build, maintain, erode or change the value of workers' skills, depending on how they're used How does generative AI actually change worker skills?. Survey data puts today's aggregate productivity gain at about 1% What is generative AI's actual aggregate productivity gain today?. That's a reminder that interaction design may now matter more than raw model capability.

Taken together, these sources suggest a further dimension: how much to personalize, and over what time span. One approach treats a user persona as something that keeps evolving, between what the system remembers and how it acts, and refines it as interactions happen Can personas evolve in real time to match what users actually want?. That links back to the constancy and timeframe dimensions. A reasonable working synthesis has three layers: the form of the interaction (interface and adaptation), how the system and user read each other (perception and persuasion), and what the interaction leaves behind in the person (their role, skills and focus). The corpus covers the first two well. The third is where the best ideas are still mostly untested.