Hugging Face has told the U.S. government that AI accountability can’t wait until a model is already live. In a June 12th filing with the NTIA, the company argued that scrutinizing only the deployment stage invites ‘surface-level compliance’ — the kind of box-checking that lets deeper problems fester until they blow up. Their position is blunt: if you want to understand what an AI system will do in the wild, you need to examine every choice made from data collection onward.
The company’s prescription leans heavily on documentation and external access. Internal practices like model cards and dataset documentation shape better development, Hugging Face argues, but they’re not enough on their own. Outside researchers, journalists, and advocacy groups need the ability to verify those internal claims. Without that external check, documentation becomes marketing. Hugging Face points to its own BigScience and BigCode projects as proof this model works — large, multidisciplinary efforts where governance and transparency were baked in from the start rather than bolted on after the fact.
This filing lands at a moment when Washington is still figuring out how to regulate AI without strangling it. Hugging Face’s framing — that transparency in both the ‘artifacts’ and the assessment outcomes is the linchpin — echoes what many researchers have been saying for years, but it’s notable coming from a company that hosts a massive share of the open-source models regulators are now eyeing. The company is essentially volunteering its community norms as a regulatory template.
What’s absent from the filing is any concrete enforcement mechanism. Hugging Face names the stakeholders who should participate — developers, researchers, advocates, policymakers, journalists — but doesn’t specify who has the authority to compel participation or penalize noncompliance. That’s the gap every AI accountability proposal eventually has to confront: good documentation practices are easy to endorse and hard to mandate. The NTIA’s request for comment closes soon, and the agency’s eventual recommendations will signal how seriously Washington takes that problem.
