The paper "Participation in the Age of Foundation Models" by Harini Suresh, Emily Tseng, Meg Young, Mary L. Gray, Emma Pierson, and Karen Levy explores the challenges and opportunities of incorporating public participation into the development and governance of foundation models. Foundation models, such as GPT-4 and CLIP, are designed to be highly adaptable and generalizable across various tasks, but their scale and domain-agnostic nature pose significant risks, including disproportionate harm to marginalized communities. The authors argue that existing participatory approaches in AI and ML, which are typically application-specific, do not easily translate to the context of foundation models. To address this gap, the paper proposes a three-layer framework for participatory foundation models: the foundation layer, the subfloor layer, and the surface layer. The foundation layer is the base model, while the subfloor layer provides technical infrastructure, norms, and governance for a specific domain, and the surface layer involves specific downstream use cases. This framework aims to create more meaningful opportunities for participation by focusing on local, application-oriented issues and ensuring that affected communities have a say in how the foundation model is used. The authors analyze existing attempts at incorporating participation into foundation models, highlighting the tension between participation and scale. They argue that the current participatory ceiling limits the ability of impacted communities to meaningfully shape a universally applicable foundation model. To overcome this, they propose a multi-layer approach that allows for more context-specific and actionable participation. Three case studies are presented to illustrate the potential of the proposed framework: 1. **Clinical Care**: A tool for transcribing medical interactions using LLMs, where patient advocacy organizations can pool data and identify key risks. 2. **Journalism**: Collective action by newsroom employees to assert copyright claims over their material used in training foundation models. 3. **Financial Services**: Addressing fairness and equity risks in fraud detection algorithms used by financial institutions. The paper concludes that the proposed framework can create more meaningful opportunities for participation, ensuring that foundation models are developed and used in ways that better serve justice and equitable governance.The paper "Participation in the Age of Foundation Models" by Harini Suresh, Emily Tseng, Meg Young, Mary L. Gray, Emma Pierson, and Karen Levy explores the challenges and opportunities of incorporating public participation into the development and governance of foundation models. Foundation models, such as GPT-4 and CLIP, are designed to be highly adaptable and generalizable across various tasks, but their scale and domain-agnostic nature pose significant risks, including disproportionate harm to marginalized communities. The authors argue that existing participatory approaches in AI and ML, which are typically application-specific, do not easily translate to the context of foundation models. To address this gap, the paper proposes a three-layer framework for participatory foundation models: the foundation layer, the subfloor layer, and the surface layer. The foundation layer is the base model, while the subfloor layer provides technical infrastructure, norms, and governance for a specific domain, and the surface layer involves specific downstream use cases. This framework aims to create more meaningful opportunities for participation by focusing on local, application-oriented issues and ensuring that affected communities have a say in how the foundation model is used. The authors analyze existing attempts at incorporating participation into foundation models, highlighting the tension between participation and scale. They argue that the current participatory ceiling limits the ability of impacted communities to meaningfully shape a universally applicable foundation model. To overcome this, they propose a multi-layer approach that allows for more context-specific and actionable participation. Three case studies are presented to illustrate the potential of the proposed framework: 1. **Clinical Care**: A tool for transcribing medical interactions using LLMs, where patient advocacy organizations can pool data and identify key risks. 2. **Journalism**: Collective action by newsroom employees to assert copyright claims over their material used in training foundation models. 3. **Financial Services**: Addressing fairness and equity risks in fraud detection algorithms used by financial institutions. The paper concludes that the proposed framework can create more meaningful opportunities for participation, ensuring that foundation models are developed and used in ways that better serve justice and equitable governance.