Social Tinkering: Why Collaborative Curiosity Beats Vibe-Coding

How can we design AI tools that protect the messy, communal discovery at the heart of learning?

MIT Media Lab artist-educator Caitlin Morris argues that social tinkering—learning through shared experimentation and productive struggle—cultivates the curiosity and resilience that individualized AI tutoring too often overlooks.

Caitlin is a Cosmos Grant recipient in our second cohort whose work focuses on fostering social connection in learning.

---

I was an unlikely software developer. Sitting alone in my university dorm room as a sophomore, staring at a laptop screen and textbooks filled with code I barely understood, I found myself in tears of both frustration and panic. Nothing was working and I didn’t know why. As someone who had typically excelled academically, this feeling of inadequacy was new and terrifying. Basic computer science concepts like sorting algorithms remained frustratingly abstract, no matter how many hours I spent. My interest in programming hadn't stemmed from any inherent love for computer science, but as a means to an end - I needed it for data collection in cognitive science experiments and computational architectural design projects.

Coding finally clicked when I discovered Processing, a procedural art tool where inputs and outputs are clearly connected through immediate visual feedback. What caught my eye initially was purely aesthetic: the beautiful, interactive visualizations that others had created. On platforms like openProcessing, I could see the finished products and the code behind them, with little barrier to the wonder of experiencing real-time feedback from code. For the first time, programming felt tangible, accessible. I could make small changes to someone else's code and immediately see the results. That tiny spark of agency ignited genuine curiosity. Perhaps most importantly, I felt part of a community. I could remix others' sketches, ask questions in forums without fear of judgment, and eventually build the confidence to contribute back. Connection to that community transformed my relationship with learning entirely.

A sketch, with editable code, on openProcessing: “noisy circles” by michelleinspace

Through this process, I grew to see coding not merely as a means to an end, but as something fascinating and joyful in its own right. The social tinkering process shaped my mind in fundamental ways. It made me comfortable with uncertainty, gave me confidence about sharing my work with others, taught me to embrace failure as part of discovery, and developed my capacity for creative problem solving. Soon, I was spending my free time sketching diagrams, translating algorithms into graphics and metaphors that made sense to me: Zeno's paradox as an animation technique, loop structures as spiraling patterns.

This journey ultimately led me to work as a media artist and technologist. I've coded 400 motors to operate in millisecond-precise synchronization, developed custom physiological sensors for research, and built platforms that integrate cutting-edge hardware and software. Each project required not just technical knowledge, but the confidence to experiment, fail, and persist that I'd developed through tinkering.

Tinkering scenes

Now, with AI-powered coding tools, students risk losing exactly what transformed my learning experience. AI offers unprecedented opportunities for personalized learning, adaptive education that meets students where they are, and dramatically increased access to powerful tools regardless of background or resources. These benefits are genuine and important. But what we gain in efficiency, we may lose in the resilience that comes from struggling and discovering together. Can we prioritize building AI systems that protect the spirit of tinkering - alone and together - to build genuine learning capacity?

Tinkering vs. "Vibe Coding"

The recent rapid evolution of AI coding tools gives me both excitement and concern. Tools like Cursor, Windsurf, and other AI coding assistants represent an incredible leap forward. They can generate entire code blocks based on simple prompts, creating functioning systems in minutes rather than hours or days. This creates massive opportunities for rapid prototyping, but is the absolute opposite of tinkering.

But these tools also fundamentally change how learning happens. Consider what happens when you want to create a simple interactive visualization. With traditional tinkering in Processing, you might start with a basic circle:

ellipse(width/2, height/2, 50,50);

You'd run the program, see a circle, then wonder: "What if I made it move? What if I wanted it smaller? What if it cycled through rainbow colors?" You'd add variables, adjust parameters, encounter errors, fix them, and through this process, develop an intuitive understanding of how the system works.

With "vibe coding" (a term coined by Andrej Karpathy), the process is fundamentally different. You might prompt: "Create an interactive visualization with circles that respond to mouse movement." In seconds, the AI generates 30 lines of functioning code. It works immediately - but you haven't built a mental model of how or why; maybe more importantly, you haven’t built the comfort with not-knowing that comes from building something step by step.

The vibe-coding approach is output-centric; it prioritises getting things to work but misses the how or why

While modern AI coding tools like Cursor now support fine-grain editing and tinkering-style manipulation (a significant improvement from earlier versions), they still are built around individual work. Students may rapidly iterate, but they do so with AI as their feedback loop, rather than being inspired by others, learning through teaching and showing and participating in critical feedback; developing the social learning skills that make tinkering powerful.

The Social Dimension of Learning

Most AI coding tools today are fundamentally single-player experiences. Yet as an educator, I see daily how learning in groups creates entirely different - more resilient - learning experiences. This happens organically in project-based classrooms: a student makes a small discovery while modifying code, expresses delight, and suddenly three peers are leaning over their shoulder, asking "How did you do that?" Social interaction transforms individual struggle into collective resilience-building. The “aha” moments in a student are peak experiences for a teacher - and they create the opportunity for students to develop genuine interest and motivation.

In my research studying pairs of learners, I've found that some of the most powerful moments of learning resilience come from interactions that might initially seem counterproductive: disagreements, challenges, and questions that create productive tension. When one student challenges another's approach or expresses confusion about a solution, it forces both learners to articulate their thinking more clearly and navigate uncertainty together. This collaborative struggle builds capacities that solitary work, even with AI assistance, cannot emulate.

Social tinkering creates specific conditions that build learning resilience:

• Psychological safety for exploration. When a student tries an unexpected approach and receives encouragement rather than judgment, it validates risk-taking for the entire group. As an educator, some of my most successful moments have been when I’ve thrown ego to the wind and put my own (good or bad) ideas forth to demonstrate this safety net, often breaking the ice for students to follow suit.

• Articulation builds understanding and confidence, transforming vague interest into structured questions. I've seen students who couldn't articulate what they wanted to build suddenly find clarity when a peer simply asked, "What are you trying to do?"

• Groups naturally distribute cognitive load. While one student's attention might flag after 20 minutes of troubleshooting, another can pick up the thread, allowing the collective to sustain investigation far longer. During hackathons, I've watched teams maintain momentum for hours through this natural cognitive relay.

• Productive conflict navigation. When students disagree about approaches or encounter unexpected behaviors together, they learn to navigate uncertainty collaboratively. They develop emotional regulation skills, learn to ask for help effectively, and discover that confusion is a normal part of learning, not a personal failing.

Diversive curiosity expands, epistemic curiosity deepens and focuses. (Made with ImageFX)

Social ideation doesn’t just add curiosity; it multiplies and transforms it. Ideas don't merely transfer between people: they evolve and recombine, creating entirely new directions for exploration. This explains why vibrant learning communities consistently produce outcomes that exceed what their most talented individual members could achieve in isolation. What emerges is more than just technical knowledge, but rather the fundamental capacities needed for lifelong learning.

How do we preserve these dynamics as coding and education become increasingly AI-supported? What if Cursor created shared workspaces where novice programmers could peek over a peer’s virtual shoulders? How could we design AI coding tools with the social richness of Processing forums - where you can see others’ thinking processes, build on their discoveries, and learn through explaining your approach to peers? Imagine if Khan Academy's Khanmigo, instead of providing individualized tutoring, prompted students to brainstorm together: “Three other students are working on this same physics problem - want to share your approach?” If we can understand which collaborative mechanisms spark genuine interest and amplify learning, we can design AI systems that create connection rather than isolation.

The Missing Element in AI Learning Tools

While the wave of current AI-powered learning platforms excel at delivering instructional content, improving accessibility of information, and creating personalized interfaces, they often lack support for social tinkering. Many actively discourage it, whether intentional or not. AI platforms are built primarily around an instructionist model, prioritizing content delivery rather than creating spaces for collaborative exploration - the critical steps of “slow learning” that social interaction supports.

What would it look like to design AI-enhanced learning environments that support social tinkering? We need platforms that:

• Make thinking visible: Allow learners to see each other's exploration processes and stumbling blocks, not just final results

• Capture and share moments of discovery: Create mechanisms for learners to flag and share interesting findings with peers

• Facilitate question-asking: Build tools that encourage learners to pose questions to each other, not just to the AI. Students should have opportunities to play both sides of the learning relationship - sometimes as the expert, sometimes as the novice.

Toward this goal, I’m building an assessment tool that combines AI-powered analysis with human expert evaluation to capture the complex interplay of social learning.

As a teacher with experience collaboratively building hard things, and seeing the increasingly isolated and tech-centric learning habits of my students, I wondered: what if we could understand the mechanisms of how collaborative learning builds genuine interest and learning resilience, and use that knowledge to design AI systems that enhance rather than replace those dynamics?

This question drives my research into capturing the social dynamics of curiosity and motivation in learners. I’m building a tool called MoSaIC that combines AI-powered conversation analysis with human expert evaluation to understand how social interactions influence different aspects of learning. This tool analyzes interaction patterns - like questions, challenges and disagreements, idea development, uncertainty, and exploration - across different learning settings, while also integrating participants' subjective experiences. Through this analysis, we can discover patterns like productive disagreements, collaborative problem-solving sequences, moments where students teach each other, and expressions of shared discovery.

Mapping social interactions that support idea transitions and deeper interest in learning

In my studies with pairs of learners working on technical projects, I’m beginning to see fascinating patterns in a variety of social interactions. When one student expresses uncertainty or poses a challenging question, it can trigger what researchers call “curiosity contagion” - their partner becomes more engaged and willing to delve into uncertainty as well, opening up to reward through risk. Powerful learning moments often follow sequences where students challenge each others’ assumptions or admit confusion together. This reveals something crucial: the messiness of social learning - the disagreements, the need to explain your thinking to someone else - isn’t a bug to be solved by more efficient AI for learning at warp-speed. It’s a feature that develops deeper qualities - curiosity, personal interest, social skills - that are essential for thriving in a collaborative world.

What's particularly powerful in the social “aha” moments is how different types of curiosity interweave. What begins as an individual's uncertainty transforms into collective epistemic curiosity, a shared drive to understand the underlying principles. Throughout this process, students aren't just learning technical information; they're developing collaborative problem-solving skills, emotional regulation in the face of unexpected outcomes, and the ability to articulate half-formed ideas—competencies that extend far beyond technical knowledge. They’re engaging in a slower but deeper learning than an AI tutor would provide.

Rather than viewing AI as a replacement for human-to-human learning, I'm working to understand how AI can help us enhance and understand the social dimensions that make learning deeply engaging. My assessment framework maps the "curiosity contagion" that occurs in effective learning environments, revealing how ideas spread, evolve, and deepen through social interaction. My goal is that this work will progress into the development of tools to support the power of spontaneous collaborative discussions, while taking advantage of all the learning opportunities AI systems can provide.

The Future of Learning Together

As we rush to integrate AI into education, we need to be thoughtful about what we preserve and what we might be leaving behind. We're living in an extraordinarily exciting time for rapid prototyping and creative expression. AI coding and education tools have democratized the ability to turn ideas into functioning software in minutes, opening up possibilities that would have taken weeks or months just a few years ago. These efficiency gains and accessibility improvements could transform who gets to participate in creating technology—but we shouldn't sacrifice the messy, social, deeply human process of tinkering together.

Alongside this excitement about what AI can build, I'm equally excited about understanding how to design these tools to support human thriving, not just code optimization. The most capable creators of the future won't just be those who can prompt AI most effectively, but those who can collaborate meaningfully, persist through uncertainty, inspire others with their discoveries, and find joy in shared exploration. I’m convinced that the most genuine learning and innovation doesn't emerge from solitary work—or from one-on-one interaction with an AI—but from tinkering with ideas in the company of others. The challenge ahead is designing technologies that support this fundamentally human process rather than replacing it.

---

These are the questions driving Caitlin’s PhD research into social learning and AI. She’d love to continue the conversation—follow along on her new Substack for ongoing thoughts, or check out this recent MIT News piece about her work.