世界经济论坛-数据公平：生成人工智能的基础概念（英）-2023.10-19页_4mb

Data Equity: Foundational Concepts for Generative AI

Introduction

Generative AI (genAI) tools, such as ChatGPT and Stable Diffusion, have gained widespread attention due to their transformative potential. The core concept of data equity ensures fair representation and accountability in AI systems, integrating principles of justice, non-discrimination, transparency, and inclusive participation. Addressing data equity is urgent due to the rapid adoption of genAI, which may exacerbate societal inequalities.

Key Definitions

• Artificial Intelligence (AI): Emulates human intelligence for diverse tasks.
• Machine Learning (ML): Uses algorithms to identify patterns in datasets.
• Generative AI: Produces new text, images, or other media based on training data.
• Foundation Models: Large-scale ML models trained on diverse datasets to adapt to various tasks.
• Large Language Models (LLMs): Subset of foundation models specializing in human language processing.

Classes of Data Equity

Four distinct classes are identified:

1. Representation Equity: Ensures marginalized groups are fairly represented in datasets.
2. Feature Equity: Includes attributes like race, gender, and income to address biases.
3. Access Equity: Promotes equitable access to AI tools and data, addressing digital divides.
4. Outcome Equity: Ensures fairness in AI outcomes and mitigates unintended consequences.
   Additional types (procedural/decision-making equity, temporal/relational equity) are noted beyond the report’s scope.

Data Equity Across the Data Lifecycle

Data equity intersects with all stages of the AI lifecycle:

1. Input Data Equity: Requires diverse and representative datasets.
2. Algorithmic Data Equity: Ensures algorithms are fair and transparent.
3. Output Data Equity: Advocates for equitable benefits from AI-generated outputs.

Challenges in Foundation Models

• Biased training data reinforces societal inequities.
• Large datasets are difficult to audit and may lack consent.
• Foundation models’ generality risks amplifying biases.
• Opacity in foundation models complicates fairness assessments.

Focus Areas for Key Stakeholders

AI-Creating Organizations

• Prioritize transparency, fairness, and inclusive data collection.

AI-Using Organizations

• Conduct audits, monitor outputs, and ensure accountability.

Policy-Makers and Regulators

• Develop ethical guidelines, regulations, and frameworks.

Civil Society

• Raise awareness and serve as advocates for marginalized communities.

Public and Communities

• Participate in AI governance and understand their rights.

Discussion and Conclusion

Data equity must be integrated into the development and use of genAI from the outset. Proactive measures by all stakeholders—industry, governments, academia, and civil society—are necessary to avoid perpetuating inequities. Addressing challenges early can shape genAI toward inclusive innovation. The report serves as an initial framework for broader discussions on equitable AI development.