The Freedom to Create? Toward Openness and Pluralism in AI’s Future

In a world increasingly shaped by artificial intelligence, the question of creative freedom takes on new urgency. Camille François and Mark Surman examine both the promises and pitfalls of generative AI, highlighting its potential to foster or stifle diverse artistic expression. Emphasizing the importance of open-source principles, they advocate for approaches that empower creators, safeguard cultural richness, and democratize technology. Drawing from the work of artists and activists worldwide, they argue that AI’s trajectory is not predetermined; rather, it is shaped by our collective decisions to build a more inclusive and imaginative digital future.

“We believe that culture should be a two-way affair, about participation, not merely consumption. We will not be content to sit passively at the end of a one-way media tube. With the Internet and other advances, the technology exists for a new paradigm of creation, one where anyone can be an artist, and anyone can succeed, based not on their industry connections, but on their merit.”—The Free Culture Manifesto, 2004

NOTE: The original excerpt was published in Villa Albertine’s bookazine, States (available here).

Every creative knows that they stand on the shoulders of giants. Practices like studying, mimicking, and adapting the greats are stepping stones on the way to developing one’s own style. 

Writers are taught to learn the rules, then break them. Musicians sample, cover, and remix their favorite songs. Communities rally around their favorite fictional universes to create fan-art and fanfiction—extending and celebrating original works through grassroots creativity.

Yet remix culture feels different when a machine is doing the making. Generative AI tools seem to co-opt creation rather than enabling it, making art feel more like a slot machine—token in, picture out—than a labor of love. Some worry that products like DALL-E and Midjourney will flatten and homogenize visual culture toward the same video game aesthetics. Science fiction writer Ted Chiang argues that AI “art” abdicates the thousands of “choices” inherent in artistic agency, instead leaving the AI to average or copy the styles of its source data. And most proprietary AI products make it impossible for the user to understand what they are playing with, or to make their own choices about the model. What data is it trained on? What biases are ingrained? It’s no surprise, then, that so many creatives are wary of AI’s impacts on their livelihoods, and on culture as a whole.

Yet these qualities–blandness, bias, opacity—are not inherent to AI. These, too, are choices made by the companies developing generative AI products. In fact, AI could be different–it could lead to a more interesting, diverse, pluralistic creative world—if more people are involved in shaping it.

There are already examples of this happening. Peter-Lucas Jones and Keoni Mahelona, an indigenous couple in New Zealand, are using AI to preserve te reo, the Maori language: digitizing oral histories, building a language model, and maintaining data sovereignty to prevent non-consensual or exploitative use. In other cases, artists are curating their own datasets and training their own AI models from scratch. To take another example, Mat Dryhurst and Holly Herndon’s Holly+ project is an open source model, trained on Holly’s voice, that they encourage other musicians to create their own music with. The pair describe the project as a form of “identity play,” or a “fork” of Holly’s own artistic persona.

And then there’s French artist Pierre Zandrowicz, who trained his own AI model to produce a series of photographs, then deliberately destroyed the model afterwards. While he used open source tools, this wasn't to ensure reproducibility—quite the opposite. In an era where AI art is often criticized for eliminating scarcity since images can be generated infinitely, Zandrowicz took a radical approach. He generated exactly 200 pieces for his work ‘Ceremony’ and then terminated the model's existence, ensuring it would never create again. By using open source tools and then destroying the model completely, he gave the work closure and deliberately reintroduced rarity to AI-generated art. The process became a ceremony of creation and finitude—a conscious push against the infinite reproducibility typically associated with AI art.

Jones, Mahelona, Dryhurst and Herndon, Zandrowicz are especially resourceful early adopters. But training your own AI isn’t yet a trivial task—especially because most top models (e.g. GPT and Claude) are released as closed products. Because closed AI models don’t share their training data, model code, model weights, or other inputs, they can only be used within the narrow bounds of the paid API. And that means that at the end of the day, the “acceptable” uses of this technology depend on the whims of private corporations. Open source AI—which is designed to be remodeled and reshaped—could provide another path.

History shows us why openness matters. In the 1990s, the free and open-source software movement emerged as a radical alternative to the big commercial tech players of the day, asserting that software should be free–not just from cost but from centralized control. Developers gained the freedom to use, modify, and share code, catalyzing an era of unprecedented creativity and innovation. This ethos laid the groundwork for the creative diversity of the modern internet, creating a lego kit that has been used to build much of the digital world we live in today. This movement evolved into the “free culture” movement, which extended open source principles into the realms of art, music, science and culture. A new wave of computer-assisted creativity emerged, promoting the remix, reuse, and redistribution of cultural and scientific works. This shift recognized that creativity of all kinds can flourish when ideas can flow freely, building on the past to create something new.

The architects of the free and open source software movement were inspired by four freedoms: the freedom to use, study, modify, and share. And, as the free culture movement emerged, these same freedoms also inspired artists, scientists, educators, and all sorts of creators. Like the previous generation of software, AI has the potential to unlock new possibilities for human creativity and invention—or to lock things up and dumb things down. Thus, the four freedoms can be useful starting points for asking whether—and how—it might be possible to shape AI’s trajectory.

Freedom to Use: Expanding Control and Sovereignty

The freedom to use (and reuse) technology for any purpose is a fundamental principle of free and open source software. This principle emerged to prevent vendor lock-in, limit surveillance, and enable possibilities beyond what the original developers might have imagined.

At first glance, extending this freedom to AI seems straightforward: just download an open model and use it. Tools that help people run AI models locally and inexpensively on their own hardware (versus big cloud providers) can further lower the cost and compute requirements for building and using AI, and offer individuals more access, privacy, and control.

That said, openness in AI isn't as simple as declaring a model “open” or “closed,” nor is it even accurate to think of it as a single spectrum from “more open” to “less open.” Openness in AI can be defined across multiple dimensions, because different components of the stack—training methods, underlying data, model weights, documentation, and various licenses—can each have their own approach to sharing and access.

For instance, model weights are the internal values that determine how a model processes and responds to input data. You can think of them as adjustable settings or “dials” that the model tunes during training to make accurate predictions or decisions. A project might release its model weights while keeping training data private, or publish detailed documentation while being selective about licensing, or open source its testing procedures while keeping other components proprietary.

This poses interesting challenges for how we should approach the freedom to use in AI: Are AI systems released under these licenses free or open and free? How should we balance responsibility and safety with freedom and generativity? Concerns over the risks associated with powerful and adaptive AI systems have led many to create open-source-style licenses that limit what an otherwise “free” model can be used for. An example, the Responsible AI License (RAIL), puts constraints on disinformation and the provision of medical advice. The license for Meta’s Llama model prohibits using the model for military applications and the operation of critical infrastructure.

Therefore, we should recognize that openness has different meanings and implications across the multiple layers of an AI system. Understanding these dimensions helps us move beyond simplified open-versus-closed debates to appreciate the rich variety of ways that AI technology can be developed and shared.

Freedom to Study: Transparency for Accountability

Open source is defined by the right for anyone to see a program’s source code and understand how it works. 

But AI systems are often described as “black boxes,” with users unable to understand or predict their behavior. Open licenses have the potential to make it easier to study and audit AI systems, enabling scrutiny at multiple levels–from the datasets that train them to the algorithms that shape their outputs. This is very much what happened with software, where open source made it easier for developers to spot and fix security bugs.

In AI, this kind of transparency could invite even more wide-ranging collective examination. It could help researchers uncover biases, track errors, and improve safety. It could be valuable for citizens and policymakers seeking accountability. And it could offer artists and philosophers a way to peer inside AI to muse about how it ‘thinks’. By opening up AI, we enable diverse perspectives to inform how these systems operate and evolve.

However, “open source AI” is a contested concept. It is often only the final, “baked” AI models that are released as open source—with the training data and even the software used to create the model left locked up. These models are still incredibly opaque, making it much harder than it seems to inspect and understand the underlying properties and behaviors of the model. If we want to empower a broader community to understand the AI systems around us, more transparency is necessary.

Freedom To Modify: Enabling Pluralism

At the heart of both open software and free culture lies the freedom to remix and reshape—to build on others’ work and reinterpret it in your own way. This freedom levels the playing field: anyone can modify software to suit their community’s needs, or iterate on existing work rather than starting from scratch. 

Likewise, the most interesting artistic uses of AI aren’t just prompt engineering but instead heavily customizing a model to create more interesting results. For example, artist Jake Elwes retrained a facial recognition model on 1,000 images of drag queens to question gender, surveillance, and diversity in AI. Several existing open source generative AI models have offered some freedom on this front, allowing users to fine-tune systems and change model weights to meet specific needs or align them with local languages, values, and norms. This freedom fosters pluralism by enabling smaller communities and organizations to develop tools that reflect their unique realities, instead of being limited by the priorities of global tech companies.

However, most trained AI models remain constrained by their creators’ choices. While proprietary models can be fine-tuned within their owners’ bounds, truly forking an AI model–creating something fundamentally new and difference–requires complete access to model weights, training software, and pre-training data. It also includes the freedom to access and reuse elements like the architecture, training data, testing systems, and other foundational tools. Few models or systems offer this level of end-to-end openness today, limiting the diversity of new directions that researchers and developers can explore.

Freedom to Share: Community-Driven Innovation

Finally, sharing is the engine that powers both technological innovation and creative expression. In software, sharing allows developers to build on each other’s work, accelerating progress. In free culture, it enables new forms of collaboration, where artists and creators contribute to a living tapestry of ideas. In AI, the freedom to share plays a similarly critical role.

When AI models and datasets are openly shared–when they invite others to use, studyand modify–they have the potential to foster ecosystems that challenge centralized control. These ecosystems encourage innovation not by monopolizing tools but by distributing them widely, enabling people to tailor AI to their needs. Community-driven innovation leverages combined collective intelligence to build better technologies than any single actor. Free collaboration can lead to a world of AI that reflects many perspectives, not just those of a few big tech players in Seattle and San Francisco. AI can be a diverse ecosystem, not a monoculture. 

The Critical Juncture

AI’s future is still unwritten. Will it become a black box controlled by a few, or a canvas open to all? Will it empower more communities to preserve their cultures and amplify their visions, or will it impose homogeneity on a diverse world? 

We are still at an early moment in the development of generative AI. Corporations and governments are still trying to understand how the technology should work and toward what ends; the possibilities are still undefined. This is a moment for experimentation–for all of us, creatives, developers, activists, and more, to think beyond the products and uses we see today and toward the future that we want. The choices we make today will shape what AI looks like.

The lessons of the open source software and free culture movements teach us that innovation flourishes when control is shared; when all can participate in making technology, not just using it. Our collective imaginations are broader than any one individual. The future of AI, like the internet before it, can belong to everyone—if we choose to make it so.

About the Authors

Camille François is an IGP Affiliated Faculty member at Columbia SIPA and a researcher in digital disinformation and cybersecurity. Mark Surman is the president and executive director of the Mozilla Foundation.

Sources

1. Ted Chiang, “Why A.I. Isn’t Going to Make Art”, New Yorker, August 31, 2024
2. Karen Hao, “A new vision of artificial intelligence for the people”, MIT Technology Review, April 22, 2022
3. Anna Wiener, “Holly Herndon’s Infinite Art”, New Yorker, November 13, 2023
4. “Zizi—Queering the Dataset” (2019), multi-channel digital video without audio, 135-minute loop https://www.jakeelwes.com/project-zizi-2019.html