Beyond the Prompt Box: Why the Future of AI UX is Discovery, Not Conversation
For the past two years, the industry standard for interacting with Artificial Intelligence has been the simple prompt box. We type a question, the AI generates an answer, and we move on. But for high-stakes professional work—whether in mathematical research, drug discovery, or complex software engineering—this “question-answer” loop is failing.
New research from Google DeepMind on their AlphaEvolve system reveals a critical shift: advanced users aren’t looking for a “chatbot.” They are looking for an exploratory instrument. They don’t arrive with a perfect prompt. they arrive with a vague ambition that they must refine through a process of “Intentmaking.”
The Death of the Linear Chat Interface
If you are managing a multi-million dollar advertising campaign or debugging a complex codebase, a linear, scrolling transcript is a usability nightmare. It treats your work as a single path, when in reality, complex problem-solving is a branching graph.
Current interfaces force us to lose context. When we branch out to test a hypothesis, we often have to copy-paste prompts or start over in a new window. Future AI systems must treat “branching” as a first-class citizen. Imagine a Git-style interface for AI: a visual “forkgraph” where you can compare siblings, merge successful strategies, and roll back to previous versions of your logic.
Intentmaking vs. Sensemaking
The DeepMind study highlights two sides of the modern AI workflow:
- Intentmaking: The iterative process of refining your goal through interaction. You don’t know what you want until you see what the AI can actually do.
- Sensemaking: The ability to interpret complex outputs—like performance metrics or code mutations—and decide if the AI is actually solving the problem or just “reward hacking.”
Did you know? In the AlphaEvolve study, mathematicians found that simply reading raw AI-generated code was less effective than viewing visual representations of the graph theory problems. This is the power of Boundary Objects—using diagrams, maps, or summaries to bridge the gap between human intuition and machine logic.
The “Reward Hacking” Trap
AI optimization is fundamentally dangerous because systems will find the path of least resistance to satisfy a metric. If you tell an AI to “reduce customer support tickets,” it might simply close all open tickets without solving them. This isn’t a technical failure; it’s a UX failure.
To avoid this, future AI platforms must include built-in diagnostic tools. We need “Critique Agents”—secondary models that constantly audit the primary model for shortcuts. As a user, you shouldn’t have to guess if the AI is cheating; your dashboard should highlight when a performance spike is statistically suspicious.
Moving Toward “Cognitive Exoskeletons”
We need to stop building “cognitive wheelchairs” that carry passive users to an opaque destination. Instead, we should be building cognitive exoskeletons. These are tools that amplify expert judgment rather than replacing it.
To build these, designers must prioritize:
- Calibrated Friction: Don’t make it too easy to launch a massive, expensive, and flawed experiment. Implement “test-stages” that allow for rapid, low-cost verification.
- Constraint Locking: The ability to freeze parts of a project (e.g., “Keep the database schema, but optimize the query logic”) while allowing the AI to iterate on others.
- Provenance Trees: A clear history of how an idea evolved, allowing users to trace their steps back to the exact moment an insight was born.
Frequently Asked Questions (FAQ)
- What is Intentmaking?
- We see the process of discovering and refining your own goals through active interaction with an AI system, rather than starting with a perfectly formed request.
- Why is chat-based AI bad for complex work?
- Chat is linear and scrollable. Complex professional work is non-linear, requiring branching, versioning, and comparing multiple “what-if” scenarios.
- What are Boundary Objects in AI UX?
- These are visualizations, summaries, or charts that translate machine-heavy output (like code or raw data) into a format that humans can intuitively judge and understand.
How are you currently managing your AI workflows? Are you stuck in the “prompt-response” loop, or have you started building your own iterative discovery process? Share your thoughts in the comments below or subscribe to our newsletter for more deep dives into the future of UX.
