Before proposing solutions, I needed to understand what research already existed. I conducted a systematic literature review of the past decade's work on AI personalization in user interfaces, searching 7 research databases and applying PRISMA screening protocols.

70 peer-reviewed papers from top-tier journals and conferences made the final selection. Each was coded across 30 parameters to identify patterns, gaps, and validated findings.

The process took 6 months and resulted in a journal paper currently under peer review at a Q1 SCI journal in HCI. The paper serves as a foundation for designers navigating this rapidly evolving field.

Key finding: While AI adaptivity can enhance user experience, it also introduces friction when not implemented thoughtfully. The application needs to be nuanced, not universal.

Academic research often happens in controlled environments. I wanted to understand how people actually use AI tools in practice, across different contexts and professions.

I designed a global survey that reached 345 participants across 46 countries. Respondents ranged from academics to artists, marketers to engineers, providing a genuinely diverse perspective on AI usage patterns.

Beyond quantitative data, we collected nearly 400 written responses detailing specific frustrations, desired features, and interaction patterns.

Key findings: Users want enhanced control over AI personalization. Prompting needs to be easier and more guided. Seeing sources matters. 80% want more control over AI functionality. 

I audited contemporary AI tools to document current interaction design patterns and identify gaps. Tools were grouped into 6 interface archetypes: multimodal chat interfaces (ChatGPT, Perplexity), image generation (Midjourney, Leonardo), video (Runway), software generation (Lovable, v0), canvas-based (Miro), and agentic AI (Zapier AI, n8n).

Each tool was evaluated against 20 AI design guidelines selected from Microsoft's Human-AI Interaction Guidelines and Google's People and AI Guidebook. I documented user flows from initial prompt to final result, creating an original audit framework and a 50-page analysis of the current state.

Key findings: Most tools lack meaningful user control mechanisms. Tools either overpromise capabilities or are difficult to prompt effectively. The "empty chat box" pattern dominates all interface archetypes.

Why existing solutions fail:
- Chat-only interfaces assume everyone is a prompt engineer.
- Parameter dashboards assume everyone wants to tinker.
- Ignoring how humans actually work: set preferences once, adjust in the moment, refine after seeing results.

The framework emerged from watching how users naturally tried to interact with AI. Three layers, each serving a different moment in the workflow.

Layer 1: Project DNA – Set it and forget it
Persistent defaults that travel with you. Your role, tone, and format preferences. The AI remembers, so you don't repeat yourself in every prompt.

Layer 2: Prompt Controls – Override in the moment
Four quick adjustments above the chat: length, tone, task type, and format. When you need formal instead of casual, detailed instead of brief, without changing your defaults.

Layer 3: Section Editing – Refine without restarting
Select any part of the response and adjust it. Expand this section, improve that paragraph, find sources for this claim. No need to regenerate everything.

The logic: Most recent control wins. Prompt Controls override Project DNA. Section Edits override the original response. This matches how humans think: "I usually want X, but right now I need Y."

Why this works:
- Novices see a simple chat interface.
- Controls reveal themselves progressively as users engage.
- Power users get granular control without UI clutter.
- Everyone gets exactly the complexity they need, when they need it.

Tests: Controlled A/B testing against standard chat interfaces. Participants complete identical tasks while we measure efficiency, adoption, and satisfaction through standardized UX methods and behavioral tracking.