Comments on U.S. National AI Research Resource Interim Report

Back in June 2022, Hugging Face filed a response to a joint request from the White House Office of Science and Technology Policy and the National Science Foundation. The topic: how to actually build out the National Artificial Intelligence Research Resource, or NAIRR, a proposed shared infrastructure meant to give researchers across the country access to compute, data, and tools for AI work. Hugging Face, which has built its entire platform around opening up machine learning to as many people as possible, used the comment period to push a handful of concrete recommendations.

The first ask was about who sits at the table. Hugging Face wants NAIRR to bring in technical experts who also have real track records on the ethics side, not just engineers who can optimize a model. They pointed to their own Chief Ethics Scientist, Margaret Mitchell, as the kind of person who could advise the Task Force on where AI systems might go wrong before they go wrong at scale.

Documentation was another big theme. Hugging Face argued that NAIRR should mandate standardized templates for describing both models and datasets, something like the Model Cards system that's already caught on widely in the ML community. The logic is simple: if every dataset and model comes with a readable, consistent writeup, researchers from different backgrounds can actually evaluate what they're working with instead of guessing.

Accessibility came up repeatedly, and not just in the abstract sense. Hugging Face wants NAIRR to fund low-code and no-code tools so people without deep technical training can still train or evaluate a model. They cited their own AutoTrain product as a working example of this in practice. They also flagged language as a barrier, arguing that any national resource needs to support multiple languages from the start, using the most spoken languages in the U.S. as a floor rather than an afterthought, and pointed to the multilingual BigScience project as proof that broad international collaboration produces genuinely useful open models.

The last recommendation dealt with the uncomfortable part: misuse. Hugging Face wants NAIRR to explicitly define harm, covering things like biased outputs, disinformation, and hate speech, and to keep updating that definition as new problems surface. They also suggested NAIRR invest in legal expertise to write Responsible AI Licenses, giving the organization actual tools to respond when someone abuses open resources rather than just hoping good behavior prevails.