Hugging Face Ethics and Society Newsletter #1

Hugging Face has launched the Ethics and Society newsletter to share the company's ongoing efforts to operationalize AI ethics. The initiative is led by the "Ethics and Society regulars," a cross-functional group of employees who maintain that ethical responsibility should be a shared commitment across the entire company rather than the domain of a single dedicated team.

Core Ethical Values and Framework

Hugging Face grounds its approach to AI ethics in four primary values derived from its open-source origins and research community integration:

  • Collaboration: Engaging with the open-source community by providing tools for documentation, evaluation, and community discussion platforms like Discord.
  • Transparency: Sharing internal thinking and processes, including project-specific value statements and perspectives on AI policy, to solicit community feedback.
  • Responsibility: Prioritizing the long-term impacts of technology by creating systems that are auditable and understandable, even for non-ML experts.
  • Reproducibility: Integrated as a fundamental value as the research community began using the Hugging Face Hub to host models and data.

Operationalizing AI Ethics

Rather than adopting a global list of static principles, Hugging Face employs a context-specific approach to operationalizing values based on the foreseeable effects of individual projects. This methodology focuses on "democratizing good ML" by implementing practical tools and guidelines:

Technical Tools for Ethical ML

To increase auditability and transparency, Hugging Face has developed and deployed several technical artifacts:

  • Evaluation Libraries: Tools to enable developers to evaluate models more rigorously.
  • Data Analysis Tools: Code specifically designed to analyze datasets for skews and biases, including experimental tools for data analysis that do not require coding.
  • Carbon Tracking: Tools to track carbon emissions during model training.
  • Responsible Licensing: The development of open and responsible AI licensing (RAIL) to address potential harms created by AI systems.
  • Reporting Mechanisms: A "flagging" system for model and Spaces repositories to report ethical and legal issues.

Community and Governance

To ensure an inclusive environment and the democratization of ML, Hugging Face has implemented the following governance structures:

  • Direct Feedback: Users can provide direct feedback on models, data, and Spaces hosted on the Hugging Face Hub.
  • Documentation Requirements: Implementation of model cards and documentation requirements to improve understanding of model utility and limitations.
  • Code of Conduct and Content Guidelines: Established frameworks to guide community discussions and maintain an inclusive space.
  • Private Hub: A secure environment for ML development.

Future Outlook

Hugging Face continues to research the criteria for "good" ML and intends to publish further pieces on values, tensions, and the operationalization of ethics in the coming months.

Sources