Common Ground: A Framework for Human-AI Collaboration

#ai #productivity #programming #discuss

Introduction

Recently I've been thinking about how AI, developers, and end users can efficiently collaborate in content management.

At the same time, Matt Biilmann, CEO of Netlify, who introduced the term AX (Agent Experience), wrote about AX, Creativity and the Human Web. I frequently find myself aligned with Matt's vision. In this case too, I strongly agree with his perspective that we must find a balance between automation and human creativity, ensuring AI can enhance—rather than constrain—human expression.

I believe that true creativity can only emerge from genuine collaboration between humans and AI. Creativity is an iterative process grounded in a medium: think of a painter sketching on canvas or a composer playing notes on a piano and transcribing them to a music score. This is why I believe humans should not only provide abstract feedback to AI, but also be able to directly modify the creative work in progress.

To address this need, this article introduces the concept of "common ground."

Common Ground

Common Ground is a shared medium, such as a file, text, canvas, or interactive interface, where multiple parties can collaborate as peers on intellectual work. It serves as a shared format or interface enabling effective cooperation between humans, and between humans and AI.

Good Common Ground Examples

When considering effective human-AI collaboration, two examples spring to my mind: software development and copywriting.

In software development, code serves as an excellent common ground, as both developers and AI can understand and iteratively modify it.

In copywriting, the text itself acts as the common ground, allowing both content editors and AI to read and directly refine the work.

Bad Common Ground Examples

When AI produces output in formats that humans cannot easily modify, the collaboration becomes less effective and ultimately constrains human creativity.

Imagine if AI always produced low-level assembly code, or highly optimized but unreadable code without comments. The collaboration would become a one-way process: after AI creates the code, only AI could modify it. Developers would be limited to providing prompts and crossing their fingers. No common ground → no true collaboration.

Consider AI-generated music (as a "composer myself, I'll set aside my reservations about AI-generated music): while AI can create perfectly mixed instrumental and vocal tracks, humans cannot directly modify them.

The same applies to images and video: humans can only provide feedback and eventually settle for the results, even when the outcome differs from their original vision. This happens simply because they cannot make further improvements through prompts, and the absence of common ground prevents them from directly "sketching" their ideas.

Characteristics of Effective Common Ground

Finding a good common ground is straightforward in some cases, but challenging in others. When multiple approaches exist, we should aim for the one that offers the most effective common ground. A big LLM can work easily with unstructured output and refine it, but human brain’s neural network needs to work on structured schemas.

For AI-generated music, a musical score would serve as a better common ground than audio tracks. This would enable humans to directly edit harmony, melody, and lyrics and quickly test their changes, with fully mixed tracks generated only as a final step (and even for this step, it would be better to have the raw tracks as well so that compression or effects could be tweaked by a human).

Another example: now that AI can generate entire full-stack applications, I propose that an AI-driven data-intensive app generator should establish common ground through an augmented database structure. This structure would include metadata for each field, indicating whether it should appear in list or detail views, what type of editing control works best, relevant permission rules, and so on. This approach would enable developers to easily modify properties without wading through AI-generated code or relying trial-and-error prompting for simple changes.

These examples demonstrate that an effective common ground between AI and humans is typically a well-structured format—such as a JSON file, programming code, or a musical score.

A structured common ground has two advantages:

It makes complex problems more manageable for humans
It provides a set of good constraints for AI, so that the output stays predictable and manageable

As a former jazz player, I'd like to make a parallel to improvisation: when you play a jazz standard, you have a well-defined structure—the sequence of chords and their harmonic function. When you improvise over this chord structure, you follow certain rules and use specific scales for each chord. This doesn't limit your freedom; quite the opposite, it provides helpful rails and a framework to express creativity. Moreover, the shared chord structure provides all band members with a solid common ground for creating music together. Within this framework, they can listen to each other and collaboratively create something new.

This is true for art in general: good rules provide guidance, and artists need to know them even when they intend to break them. Good constraints don't limit creativity—they help it flourish. Try this simple test: invent a line of poetry. Difficult, isn't it? Now, invent a line of poetry with the word "key". Your brain immediately starts working, right?

By providing a framework within which both humans and AI can work, we create space for innovation while maintaining coherence and consistency.

Conclusion

The concept of Common Ground offers a valuable framework for understanding and improving human-AI collaboration. As AI becomes increasingly integrated into creative work, establishing effective shared mediums will be essential for a truly collaborative process.

The most effective collaborations occur when both parties work within a well-structured format that balances flexibility with constraints.

Structured formats like a component’s interface, a well-defined JSON object, or musical notation provide:

Clear syntax and rules that both humans and AI can understand
The right level of flexibility for creative expression
A schema that make complex problems manageable
Good constraints leading to predictable output that supports iterative refinement

As we design tools and workflows for human-AI collaboration, we should prioritize creating robust, structured common grounds that amplify human creativity rather than constrain it.

In a forthcoming article, I'll explore how these Common Ground principles apply specifically to content management systems, examining which approaches provide the most effective collaboration environment for developers, content editors, and AI.