Hugging Face’s multimodal learning group has published an ethical charter for their research project, making a rare move in the machine learning world: formalizing their values before the work is done, not after. The document, current as of May 2022, was shaped by discussions among the group’s researchers and engineers, with input from outside experts in ethics operationalization, data governance, and personal privacy.
What’s in it? The charter lays out a content policy that draws a hard line against promoting violence, harassment, hate, and discrimination, generating personally identifiable information, creating false information designed to harm, and using models carelessly in high-stakes domains like medicine, law, finance, and immigration. The team also commits to five core values: transparency, open and reproducible research, fairness, self-criticism, and giving credit.
There’s a refreshing honesty here. The authors explicitly state the document is a work in progress and that there is no consensus or official definition of ‘ethical AI.’ They acknowledge the charter will likely change and promise to track updates through GitHub. That candor is notable in a field where ethics statements often read like polished press releases. The group frames this as a bet: even an imperfect ethics document, written early and revised openly, is better than the alternative of not thinking systematically about harms at all.
This is not the first time Hugging Face has pushed on responsible AI practices. The company has made a name for itself by championing open-source models and hosting public discussions about bias and misuse. What makes this charter different is its scope—it applies to a specific research project, not the company as a whole—and its insistence on operationalizing values throughout the lifecycle, from data curation to model outputs. If the team actually documents those operational decisions as promised, it could give other research groups a template for doing the same without waiting for industry-wide standards that may never arrive.
