Project Overview:

With the proliferation of Large Language Models, we have new opportunities for creative collaboration. However, users often struggle to form accurate mental models of AI in co-creative contexts, which can hinder effective collaboration.

We aimed to address this challenge by designing a tool that could assist in creative processes and facilitate effective human-AI collaboration. The result was a tool we now use in IBM Innovation Studio's co-creation and design thinking workshops.

Initial Problem Statement:

“How might we help users form accurate mental models of Generative AI in co-creative contexts to build effective human-AI collaboration?”
‍

Research and Initial Experiments

In July 2022, I began experimenting with GPT-3, prompting it much like participants in design thinking workshops. I discovered that GPT-3 could contribute creative artifacts useful for context-specific ideation. By conversing with AI and transferring its outputs onto sticky notes, we effectively engaged the models in co-creation.

However, initial experiments highlighted issues with inconsistent output quality in group settings. To address this, I developed a framework for prompting AI agents to actively and consistently participate in design thinking with human groups.

Designing the Collaborative Canvas

Based on these findings, the HCAI team and I developed and tested a broader, medium-fidelity collaborative whiteboard interface that generated actual sticky notes from user input:

Initial testing found that a simple "input context -> output ideas"  feature was useful for live ideation, neat! However, users wanted the option to utilize existing frameworks, the freedom to explore their own ideation through free-form input, and the ability to reference the existing artifacts.

So for both of these interaction, we landed on the following strengths and weaknesses to guide our design:

Advanced UI and Features

Building on initial feedback, I designed a high-fidelity interface with new interaction patterns, clear identification of AI-generated content, a more streamlined interface, and example prompts to guide and inspire users.

Navigation and General UI:

Left Navigation: Added options for home, creating a new board, recently visited boards, favorites, and the user’s collection.

‍Bottom Interface: Introduced text input, a single sticky note button, and a design thinking template generator.

‍AI Chat Interface: Enhanced chat for direct AI engagement, allowing sticky note generation within the same conversation.

Enhanced Prompting Panel: Embedded color and text editing, a blank prompt input area, and a menu of predefined prompts.

[AI] Slug and Gradient: Clear identification of AI-generated sticky notes.
‍
Starter Prompts: Prompts to quickly utilize AI in a way that expands creativity, such as suggesting prompts, summarizing, expanding, providing variations, and asking questions

Starting out with Empathy Mapping with plans to add more in the future, I wanted to create a clean way for users to quickly generate multiple sticky notes in response to a specific design thinking exercise, so I added a feature similar to my initial experiment from earlier, with some extras like:
‍
Generate an example: Implemented a button for users to generate an example empathy map, so they can quickly learn how to use the tool.

‍For example, Merge, which combines ideas into a single cohesive concept or proposal, or Cluster Ideas, which groups sticky notes into thematic clusters based on the content. (Plus it's a very fun thing to say.)

The previous UI required switching between a private sticky note interface and a chat interface, so in addition to providing visual consistency and a streamlined chat experience, the Canvas Agent now generates sticky notes all within the same conversation.

For comparison:

Reflections and Impact

Implementing the latest version, showcasing it to clients, and running its capabilities in workshops has proved incredibly effective. We received positive client feedback, leading to increased business and a deeper understanding of AI systems. This project also contributed to higher-rated workshop experiences within our Innovation Studio and the creation of a standard framework adopted across IBM designers.
‍
We also learned a few lessons that should transfer to future Human-AI Co-Creation design pursuits:

Lesson 1: Design for Collaborative Partnership

AI can be treated as an additional team member, adding divergent ideas that encourage deeper creative reflection.

Lesson 2: Leverage AI’s Bias

AI’s unique perspective and generative variability can be used challenge conventional thinking.

Lesson 3: Guide Users, but Allow for Exploration

Structured templates and automations help users quickly build a mental model of the capabilities and guide user interaction, and free-form prompting helps them go even further.

Lesson 4: Encourage Critical Reflection

Implement tools that facilitate users to question, disagree with, and debate AI generated content.

Lesson 5: Clearly Distinguish Between Human and AI Contributions

Implementing clear labeling of AI generated content to help users distinguish and divide.
‍

Conclusion

Humans remain the creative drivers in AI interactions.

While Generative AI can synthesize insights from its training data, it does not inherently understand human desires, needs, or goals. AI provides a perspective for humans to observe, reflect on, and use to make informed creative decisions. By implementing these enhanced co-creative interactions within a collaborative canvas environment, we've built a new kind of Meaning Machine interface that has proven useful for ideation, reflection, and action setting.