The emerging field of Public AI holds tremendous promise to democratize access to artificial intelligence and ensure it benefits society as a whole. As outlined in a recent document from Mozilla, Public AI aims to promote public goods, public participation, and public use cases throughout the AI development lifecycle. However, for Public AI to succeed, we must proactively acknowledge and address the significant challenges and risks involved. Only by identifying potential pitfalls early can the Public AI ecosystem chart a path to fulfill its lofty ambitions.
The Risk of Replicating Existing Harms
One major concern is that Public AI, despite good intentions, may inadvertently perpetuate or exacerbate some of the same issues plaguing the current commercial AI landscape. Open-source datasets and models, while more accessible, are not immune to encoding societal biases that can lead to unfair discrimination against marginalized groups. And in the absence of robust ethical guidelines and oversight mechanisms, public AI tools could still be misused or exploited by bad actors.
As Mozilla notes, “Public AI initiatives can still perpetuate harms like accelerating bias and surveillance or generating inaccurate and harmful content.”[1] Preventing this will require proactively integrating principles of responsible AI development, such as fairness, transparency, and accountability, into the Public AI ecosystem from the ground up. Importantly, the community must establish clear norms and expectations around the acceptable use of public AI resources.
Sustainable Funding and Revenue Models
Another key challenge is how to sustainably fund the development and maintenance of public AI infrastructure and tools. Training cutting-edge AI models is extremely resource-intensive, often requiring millions of dollars worth of computing power. Public initiatives may struggle to keep up with the rapid pace of progress driven by heavily-funded tech giants and startups.
“Funding is a challenge in physical infrastructure as well, as seen in many U.S. transit systems, which have been plagued by mismanagement, operational failures, unprofitability, and resulting service cuts,” the Mozilla document points out. For Public AI to thrive long-term, it will need to piece together a viable mix of government support, philanthropic donations, private partnerships, and novel business models. Importantly, this funding strategy must align with Public AI’s mission and values.
Balancing Government Involvement and Independence
Public sector investment and policies will undoubtedly play a critical role in catalyzing a Public AI ecosystem. Yet over-reliance on government control also presents risks. Across the world, we’ve seen the vulnerability of public institutions to illiberal leaders, corruption, and volatile political climates.
“Though having governments control Public AI development might help with imposing some safety measures, the past several years have seen widespread election interference, regressing civil rights and liberties, and the ascent of illiberal leaders around the world,” cautions Mozilla.
Shielding Public AI from being co-opted will likely require a resilient multi-stakeholder approach. A diversity of voices and players, including nonprofits, universities, open-source communities, and ethical private sector actors, can provide checks and balances. Governance structures should be designed to weather political changes.
Technical and Infrastructure Challenges
At a technical level, Public AI initiatives must overcome significant barriers to access key inputs like large-scale computing resources, sizable training datasets, testing suites, and tooling. Major cloud platforms and chip manufacturers are dominated by a few players. High-quality data for AI is often siloed in private companies or restricted by privacy regulations.
Computing for AI can be prohibitively expensive for resource-constrained universities and nonprofits, especially for intensive tasks like training large language models. Data privacy and security remain paramount concerns and Public AI solutions will need robust practices. Finally, the Public AI ecosystem should embrace open technical standards to ensure interoperability.
Workforce and Talent Challenges
Public AI will also need to build up a specialized workforce with the skills to develop and deploy AI responsibly. This may prove difficult when competing for scarce talent against big tech salaries and perks. Relatedly, it’s crucial that the Public AI workforce represents the diversity of the communities it seeks to benefit.
“The makeup of these roles tends to reflect the inequities in the STEM workforce broadly,” notes Mozilla. Intentional strategies around recruitment, training, retention, and inclusion of underrepresented groups will be key. So is recognizing and fairly compensating oft-hidden labor, like data labeling, that powers AI.
Learning from Historical Analogies
Luckily, the Public AI movement doesn’t have to start from scratch, but can learn from prior initiatives to establish public alternatives in other domains. For example, the BBC demonstrates how a public broadcaster can remain independent, trustworthy, and innovative in its public service mission, even amid a rapidly shifting media and political landscape.
The history of open-source software development also provides a powerful roadmap. Open-source operating systems, web servers, databases, and more have gained widespread adoption alongside proprietary offerings. They often lead on security and interoperability. Similarly, Public AI technologies and private AI products could ideally co-evolve and mutually raise the bar for each other.
Strategies for Mitigating Risks
Proactively identifying risks is just the starting point; the Public AI ecosystem must also take concrete steps to mitigate them. Mozilla’s call for democratic, multi-stakeholder governance of Public AI initiatives is spot on. Governance structures should be designed to align with Public AI’s core principles and represent the interests of diverse constituencies.
Fostering a pluralistic Public AI ecosystem, spanning many geographies and domains, can also help preserve its resilience and independence. Think public broadcasters in multiple countries, or the diversity of the open-source software ecosystem. Importantly, no single player should have the power to unilaterally decide the future of Public AI.
Ongoing public dialogue and scrutiny will also be essential to holding Public AI initiatives accountable. As Mozilla notes, “activists and advocates should also call attention to risks and harms from Public AI where they exist — helping Public AI live up to its highest ideals.” For accountability to be meaningful, Public AI efforts should embrace radical transparency- publicly sharing information on funding, governance, security practices, and known flaws and harms.
More broadly, realizing the promise of Public AI will require significant investment in public education and awareness. Impacted communities need avenues to not only understand these technologies, but to actively shape their development. Universities and nonprofits could play a major role in helping the public critically engage with Public AI.
Conclusion
The path to developing Public AI that is truly beneficial and trustworthy will not be easy. From replicating societal biases, to funding shortfalls, to technical roadblocks- the challenges and risks are significant and span many layers. Yet by proactively identifying these pitfalls, the Public AI community can start to put in place thoughtful strategies to overcome them.
What’s needed is a commitment to not just make AI public, but to make it meaningfully benefit and empower all users. Funders, policymakers, researchers, advocates, and the public must all remain vigilant in holding Public AI initiatives accountable to this higher standard. Only through sustained collaboration can we build a Public AI ecosystem that genuinely shifts power and fulfills its promise.
References: [1] Mozilla, “Public AI: Making AI work for everyone, by everyone” https://assets.mofoprod.net/network/documents/Public_AI_Mozilla.pdf, September 2024.
