Cooperative Paradigms for Artificial Intelligence

An examination of the development of artificial intelligence (AI) reveals a growing concentration of corporate power and wealth at various layers of the AI stack. Today, nearly every stage in AI model development — from compute infrastructure to training data — is controlled by a handful of large technology companies. Very little public oversight exists when it comes to the development and governance of these systems.

The risks associated with concentrating AI development in monopolistic powers are generally well understood and have been widely discussed in existing literature.¹ There is a well-established need to explore alternative ownership and governance structures that would better serve the public interest as it pertains to AI development. In this essay, we will examine one such alternative structure: the cooperative. Cooperatives are worker-owned and operated firms with a rich global history of community-oriented business practices that distribute control and capital. Many well-known firms, such as REI, Ocean Spray, Dairy Farmers of America, and the Associated Press, operate as cooperative businesses. Today, it is estimated that cooperatives employ nearly 10% of the world’s population.² While they aren’t without challenges, cooperative models are structured to provide more equitable ownership and governance than investor-owned corporations.³

Below, we will discuss how such a structure can benefit the development and governance of AI while easing the growing concerns of various stakeholders: consumers wanting input into the systems dominating their lives, companies looking to set themselves apart by gaining consumer and regulator trust, and regulators aiming to prevent market monopolization.

Opportunities for Cooperative AI

Public Interest Infrastructure

Today, only a few companies, such as NVIDIA and TSMC, produce the graphics processing units (GPUs) required to train AI models at scale. These GPUs are primarily sold to a small number of large technology companies that have the resources to leverage them in compute offerings. Amazon AWS, Microsoft Azure, and Google Cloud are estimated to account for two-thirds of the entire cloud market,⁴ which has led to various antitrust investigations.⁵ Companies trying to develop their own AI models may spend millions of dollars to access the compute resources they need for training on these platforms.⁶ As a result, some of the emerging leaders in AI, namely OpenAI and Anthropic, have had to strike agreements with the large cloud providers,⁷ often in exchange for equity. This concentration of computational infrastructure by a few large technology companies has priced out many smaller global developers, created ecosystem lock-in, and made healthy competition in the cloud computing space incredibly difficult.

In an attempt to create an alternative to large corporate cloud infrastructure, projects such as the Co-op Cloud and the Commons Cloud have emerged,⁸ alongside burgeoning research into cooperative cloud computing.⁹ While these projects do not currently support the type and scale of compute infrastructure needed for the large-scale training of AI systems, they demonstrate a clear desire and need for this type of public infrastructure. Recent efforts to create a “public AI” option further validate the need for an alternative to existing corporate cloud providers.¹⁰ Bills such as the CREATE AI Act propose creating a National Artificial Intelligence Research Resource to provide compute infrastructure for AI researchers and students.¹¹

A recent paper, “An Antimonopoly Approach to Governing AI”,¹² recommends cooperative alternatives:

“At the cloud layer, the federal government could support the creation of a cooperative research-focused cloud, owned and operated by nonprofits, government, and universities to ensure sufficient compute and storage power for research into innovative, safe uses of AI — and without a shareholder profit motive. The federal government could also support the creation of a cooperative cloud for private companies, in which firms could train and operate models, and share in the ownership of the cloud, without fear that one of the big platforms will take their ideas or raise prices for the utility services they provide.”

While some progress has been made toward creating compute infrastructure in the public interest, considerable challenges remain due to the capital-intensive requirements needed to provide alternatives to the corporate cloud providers. Additional experiments in federated compute infrastructure and cooperative clouds, as well as continued exploration into federal funding models for public AI infrastructure, represent promising steps toward more cooperative AI development.

Data Cooperatives

AI systems require vast amounts of data for model training, much of which is sourced from the public internet, including user-generated media, forum discussions, and social media posts. Most individuals are unaware of how their data is being used and have little power to opt in or out of its collection. Recently, a wave of copyright lawsuits coming from artists, authors, and publications claim that these AI models and businesses are “built on the exploitation of copyrighted works.”¹³

On the other hand, companies training AI models need high-quality data to produce a high-quality model. After Google acquired Reddit data to train its large language model, Bard, many of the AI’s responses were shown to be inaccurate or otherwise flawed due to the poor-quality Reddit forum responses they referenced.¹⁴ This type of low-quality data can impact AI’s accuracy and relevance, leading to poor user experiences and ultimately negative business outcomes.

Data cooperatives offer an alternative to current extractive practices by aiming to shift the power from large corporations to the individual. They enable individuals to pool their data while retaining control over its use and collectively benefiting from its access. A data cooperative empowers end users to control who can access their data, what data they can access, and for what purpose. Furthermore, it allows end users to negotiate ownership of their data, including any financial benefits that might be distributed back to them. Hardjono and Pentland describe the need for data cooperatives in their piece on “Personal Data Management”¹⁵:

“Today we are in a situation where individual assets … people’s personal data … [are] being exploited without sufficient value being returned to the individual. This is analogous to the situation in the late 1800’s and early 1900’s that led to the creation of collective institutions such as credit unions and labor unions, and so the time seems ripe for the creation of collective institutions to represent the data rights of individuals. We have argued that data cooperatives with fiduciary obligations to members provide a promising direction for the empowerment of individuals through collective use of their own personal data.”

A leading example of this concept is a data trust called Superset,¹⁶ which is structured to allow members to contribute, govern, and be compensated for their data. Superset then negotiates on behalf of its members in a collaboration with Delphia, a customer of the data trust.¹⁷ Other collective data examples include Cohere’s Aya,¹⁸ which collected natural language data from thousands of people around the world to train a multilingual AI model. While Aya’s contributors aren’t financial stakeholders in the project, both the data set and AI model are fully open-source for anyone to download and use.

Explorations into cooperative data agreements also appear in academic research. A study titled “Worker-Owned Cooperative Models for Training Artificial Intelligence” explored having data workers (who often label and clean data for low wages) earn ownership of the AI systems they help create.¹⁹ Similarly, a recent paper, “An Economic Solution to Copyright Challenges of Generative AI,”²⁰ proposed “a simple framework that appropriately compensates copyright owners for using their copyrighted data in training generative AI systems based on the cooperative game theory …” Through recognizing the dual challenges of companies needing high-quality, often copyrighted data and AI disrupting creative industries, the authors present a cooperative solution to meet both parties’ needs.

Other examples where data cooperatives have proven successful in the technology sector include the Driver’s Seat Cooperative,²¹ supported by The Rockefeller Foundation,²² which enabled gig workers to boost their pay by crowdsourcing market information and optimizing their time. While not directly related to AI development, it showcases how data cooperatives can successfully scale and deliver benefits in other technology-dominated sectors.

Collective Governance

Lessons from the social media age demonstrate the tension that exists between the profit motives of large technology companies and their ability to self-govern to avoid diffuse harms to society. Given that AI will continue to have outsized impact on global stakeholders, it is critical to explore methods for greater collective community governance over the development and deployment of these systems. Many are rightfully skeptical of large technology companies’ claims to be committed to the public good. For example, in 2019, the Business Roundtable released a pledge of corporations to “commit to lead their companies for the benefit of all stakeholders — customers, employees, suppliers, communities and shareholders.”²³ However, a follow-up investigation demonstrated that there was no material change in how companies treated these stakeholders.²⁴

A variety of different corporate governance models are being tested in the AI space today. Anthropic is structured as a public benefit corporation and introduced a Long-Term Benefit Trust,²⁵ which appoints independent trustees to oversee the Anthropic board. OpenAI is a nonprofit that runs a for-profit entity but has experienced highly publicized governance turmoil and debate in the past year.²⁶ Meta launched the Oversight Board,²⁷ an independent entity that makes decisions on content moderation cases, which Meta is required to follow. While these are steps in the right direction, there should be further exploration of operating models that empower end users as stakeholders with decision-making power over AI systems.

There are increasing civil society efforts that advocate for collectively governed AI. The Collective Intelligence Project is leading alignment assemblies,²⁸ which aim to incorporate public input into the development of AI systems. Other methods, such as citizens’ assemblies,²⁹ are also being tested. For example, in Belgium, 60 citizens were randomly selected to discuss AI and inform the larger EU policy strategy.³⁰ Other advocacy groups like the Algorithmic Justice League are working toward “equitable and accountable AI.”³¹ Another emerging alternative is the concept of “Exit to Community,”³² where instead of an IPO or acquisition, a company instead transitions to cooperative community control. Efforts toward this end include a bid for X (formerly Twitter)³³ and a recent proposal to purchase and decentralize TikTok.³⁴

While there have been some shifts in corporate governance, there must be a greater movement toward empowering people to collectively and cooperatively govern the AI systems that are impacting their lives. As the public’s trust in institutions and large technology companies continues to decline, shifting power to the public might garner greater trust and buy-in.

Recommendations

Cooperative methods for AI development and governance present an alternative approach to the reigning paradigm of today’s corporate technology power. To support this future, a few key recommendations follow:

1. Efforts to promote greater cooperative methods within the AI development life cycle are most likely to gain traction in outlined areas such as cloud, data, and governance, where there is ongoing momentum and antitrust discussions.
2. While there are increasing experiments in corporate governance structures and public input mechanisms in AI, it is critical to ensure that these structures give actual power, enforceable oversight, and an economic stake for community stakeholders.
3. The pace of AI development raises the urgency for greater research and experimentation with cooperative AI paradigms, as well as policy interventions to ensure a healthy competitive market and innovation centered around the public interest. Immediate action is needed to prevent continued consolidation of power and wealth of large technology companies.

Exploring more cooperative paradigms stands to benefit all stakeholders: consumers gain more influence and an economic stake in the technology systems shaping their lives, businesses build trust with consumers and regulators while staying connected to public needs, and regulators facilitate a competitive marketplace with the possibility of improving AI’s societal impact. The rising tide of AI should lift all boats.