Hugging Face Ethics and Society Update Summer 2023

TL;DR

Hugging Face spent Summer 2023 engaging with legislators in the U.S., E.U., and U.K. to advocate for open-source AI and provide a values-based framework for regulation. Parallel to policy work, the organization advanced technical efforts in bias testing, dataset transparency, and environmental impact research to promote a more accountable and democratized AI ecosystem.

AI Policy and Regulatory Advocacy

Hugging Face is leveraging a vocabulary of values—including transparency, safety, and justice—to help shape AI legislation where formal laws are not yet established. The organization has focused on providing expert testimony and guidance to global governing bodies to ensure that regulation does not stifle openness.

Key regulatory engagements include:

  • United States: CEO Clem Delangue provided testimony to the U.S. Congress and statements at the U.S. Senate AI Insight Forum. Chief Ethics Scientist Margaret Mitchell provided comments to the Democratic Caucus.
  • European Union: The organization provided specific advice regarding the E.U. AI Act and the role of open-source software.
  • National Telecommunications and Information Administration (NTIA): Hugging Face submitted comments on AI Accountability.

Public Advocacy and Technical Discourse

To support the democratization of AI, Hugging Face experts have contributed to public discourse across major media outlets to clarify complex technical and ethical issues. These contributions cover several critical domains:

Environmental Impact and Sustainability

Sasha Luccioni has provided analysis on AI's energy consumption and carbon emissions for publications including The Atlantic, The Guardian, The Wall Street Journal, and New Scientist.

Model Safety and Bias

  • Generative AI Bias: Experts provided insights into bias in generative AI for Bloomberg, NBC, and Vox.
  • Doomsday Risks: Analysis of AI "doomsday" scenarios was provided to Bloomberg, The Times, and Sky News.
  • Misinformation: Margaret Mitchell contributed to discussions on AI-generated election misinformation for CNN, al Jazeera, and The New York Times.
  • Sentience and Singularity: Giada autobiography addressed concepts of AI "sentience" and the "singularity" for RFI and Popular Mechanics.

Open Source and Intellectual Property

  • Open Releases: Nathan Lambert and Yacine Jernite provided commentary on the state of open language model releases and the impact of Meta's Llama models for WIRED, VentureBeat, and Fortune.
  • Copyright: Yacine Jernite addressed copyright issues in generative AI training data for VentureBeat.
  • Artist Rights: Margaret Mitchell discussed the necessity of just handling of artists' work in AI for the Washington Post.

Technical Contributions to AI Ethics

Hugging Face has integrated ethics into its core product development and research through several "work work" initiatives focused on approachability, fairness, and rigour.

Model Evaluation and Transparency

  • Bias Testing: Hugging Face applied bias and harm testing to its multimodal model, IDEFICS.
  • Dataset Rigour: The organization added descriptive statistics to datasets to improve the understanding of what AI learns and how it is evaluated.
  • Content Policy: The organization updated its Content Policy to operationalize transparency responsibly.

Educational Resources and Research

  • Audio AI: Launched a new course on AI audio.
  • Clinical AI: Developed resources for Open Access clinical language models.
  • Responsible Generative AI: Produced a tutorial on Responsible Generative AI.
  • Academic Research: Published FAccT papers on the gradient of generative AI release and the articulation of ethical charters and technical documentation in machine learning.

Community Resources

Hugging Face has consolidated its ethics-related resources into a centralized hub to assist in current AI ethics discourse: https://huggingface.co/society-ethics.

Sources