AI needs a new recipe: Study calls for sustainable and collaborative development

The cake analogy, initially introduced by AI pioneer Yann LeCun, likened AI learning to a layered cake: unsupervised learning as the base, supervised learning as the icing, and reinforcement learning as the cherry on top. However, this new study expands the analogy to encompass the entire AI production cycle, from ingredient sourcing (data collection) to recipe creation (algorithm design), baking (training), tasting (evaluation), and distribution (deployment).

Artificial Intelligence (AI) is often described as a technological marvel, but how it is built, by whom, and who benefits from it remains a contentious debate. The field of AI has long been shaped by a small, dominant group of researchers and corporations, often sidelining crucial ethical, societal, and participatory considerations.

A new study, "The Cake that is Intelligence and Who Gets to Bake it: An AI Analogy and its Implications for Participation," by Martin Mundt, Anaelia Ovalle, Felix Friedrich, Pranav Agrawal, Subarnaduti Paul, Manuel Brack, Kristian Kersting, and William Agnew, takes a novel and insightful approach to understanding AI’s development. Using the analogy of baking a cake, the authors deconstruct AI’s lifecycle - from data sourcing to training, evaluation, and deployment - while highlighting the power dynamics at play. This study, submitted in arXiv, calls for a rethinking of AI’s participatory landscape and an urgent move towards inclusive and responsible AI development.

The AI cake: More than just an analogy

The cake analogy, initially introduced by AI pioneer Yann LeCun, likened AI learning to a layered cake: unsupervised learning as the base, supervised learning as the icing, and reinforcement learning as the cherry on top. However, this new study expands the analogy to encompass the entire AI production cycle, from ingredient sourcing (data collection) to recipe creation (algorithm design), baking (training), tasting (evaluation), and distribution (deployment).

The research critiques the lack of transparency in AI development, particularly in data sourcing. Just as the origins of cake ingredients - like cocoa or sugar - are often obscured by supply chains, AI datasets are frequently compiled without clear records of consent, ownership, or ethical considerations. This leads to bias, exploitation, and ethical concerns, as seen in the widespread use of labor from marginalized communities for AI moderation and annotation tasks.

Who controls the AI recipe? The problem of homogenization

Beyond ingredients, the study examines how AI recipes are written and who gets to modify them. Despite AI’s immense potential, the field remains dominated by a few large organizations that shape the dominant “recipes” (algorithms) used globally. These standardized AI models - often called foundation models - create a risk of homogenization, where a single approach to AI is applied universally, ignoring cultural, ethical, and contextual nuances.

This centralization of AI design limits participation from diverse stakeholders, particularly from regions outside of North America and Europe. The study calls for a shift from “one-size-fits-all” AI models to a more modular, transparent, and participatory approach, allowing communities to tailor AI systems to their specific needs.

The flaws in AI baking: Can the cake be fixed once baked?

One of the biggest challenges in AI is that, like a cake once it’s baked, an AI model cannot be easily altered. Once trained on biased or flawed data, correcting an AI system requires extensive re-training, often at enormous financial and environmental costs. This inflexibility contradicts the nature of intelligence, which should adapt, evolve, and learn continuously, rather than requiring expensive and energy-intensive retraining.

The study critiques current AI development processes that favor overhauling entire models instead of incremental, adaptive improvements. This results in wasteful computing practices that contribute to the excessive carbon footprint of large AI models. The researchers propose rethinking AI’s training process, making it more sustainable, adaptable, and reflective of human intelligence.

Who gets to eat the cake? AI’s unequal benefits

Finally, the study addresses the distribution of AI’s benefits and harms. Just as not everyone has access to gourmet cakes, AI’s advancements often benefit a privileged few while excluding marginalized communities. The widespread belief in AI as a universal solution (technosolutionism) has led to AI being oversold as a cure-all, even when it fails to address real-world problems.

For example, while AI is often marketed as a tool to improve healthcare and agriculture in developing nations, structural inequalities, lack of infrastructure, and data exploitation prevent these benefits from being equitably distributed. The researchers emphasize the need to challenge AI’s commercialization models and prioritize community-driven, ethical AI solutions rather than profit-driven applications.

Conclusion: Rethinking AI’s power structures

This study presents a powerful critique of AI’s current trajectory, urging researchers, policymakers, and developers to move beyond corporate-driven AI and create participatory, ethical, and sustainable AI systems. The cake analogy serves as an accessible yet profound lens to examine how AI is built, who controls it, and who benefits from it.

To truly democratize AI, the study calls for:

• Greater transparency in data sourcing and AI development
• Diverse participation in AI design and governance
• Sustainable, adaptable AI training processes
• Fairer distribution of AI’s benefits

AI is more than just a technological tool - it is a reflection of societal power structures. If we want AI to be inclusive, ethical, and beneficial for all, we must challenge the status quo and reimagine AI as a cake that everyone can help bake, taste, and share.