Hugging Face Took AI Ethics to Congress, the EU, and 40+ Newsrooms This Summer

Hugging Face spent summer 2023 doing something unusual for an AI company: showing its work. The open-source AI firm published a rundown of its ethics and policy activities, spanning testimony before U.S. Congress, submissions to the E.U. AI Act, comments to the NTIA, and more than 40 press engagements across outlets like The Atlantic, Bloomberg, and The New York Times. The thread connecting it all? A belief that openness and accountability matter more than hype when laws haven’t caught up to the technology.

CEO Clem Delangue testified before Congress and spoke at the Senate AI Insight Forum, while Chief Ethics Scientist Meg Mitchell addressed the Democratic Caucus. Their core argument to lawmakers: open AI development can be a feature, not a bug, when it comes to safety and accountability. The company published a collection of its answers to the recurring question of why openness in AI is beneficial — a direct counterweight to the closed-model narrative pushed by some larger labs.

The team’s media blitz was unusually broad. Sasha Luccioni talked to The Guardian, The Wall Street Journal, and others about AI’s energy consumption and carbon emissions, even co-writing a Wall Street Journal op-ed on the topic. She also fielded questions on doomsday risk, bias in generative AI, and the often-invisible labor of marginalized workers who create training data. Meg Mitchell explained the statistical concept of perplexity to Ars Technica readers, discussed misinformation with CNN and al Jazeera, and addressed how journalists can shape AI coverage with CJR. Giada Pistilli tackled AI “sentience” and “singularity” claims with French media and gave a TED talk asking whether ethical generative AI is even possible.

Beyond the public-facing work, the team shipped concrete resources: a new AI audio course, a resource on open-access clinical language models, a tutorial on responsible generative AI, and two FAccT conference papers. They also applied bias and harm testing to IDEFICS, Hugging Face’s own multimodal model, and updated their content policy. That last part is notable — many companies talk about responsible AI while quietly shipping models that would fail their own audits. Hugging Face is at least attempting to run the tests on itself. Whether that rigor scales as the open-source AI ecosystem grows remains the open question.