Purpose of the moral charter
It has been well documented that machine learning research and applications can potentially result in “data privacy issues, algorithmic biases, automation risks and malicious uses” (NeurIPS 2021 ethics guidelines). The aim of this short document is to formalize the moral principles that we (the multimodal learning group at Hugging Face) adopt for the project we’re pursuing. By defining these ethical principles at first of the project, we make them core to our machine learning lifecycle.
By being transparent in regards to the decisions we’re making within the project, who’s working on which points of the system, and the way the team could be contacted, we hope to receive feedback early enough in the method to make meaningful changes, and ground discussions about selections in an awareness of the goals we aim to realize and the values we hope to include.
This document is the results of discussions led by the multimodal learning group at Hugging Face (composed of machine learning researchers and engineers), with the contributions of multiple experts in ethics operationalization, data governance, and private privacy.
Limitations of this ethical charter
This document is a piece in progress and reflects a state of reflection as of May 2022. There is no such thing as a consensus nor official definition of “ethical AI” and our considerations are very prone to change over time. In case of updates, we are going to reflect changes directly on this document while providing the rationale for changes and tracking the history of updates through GitHub. This document is just not intended to be a source of truth about best practices for ethical AI. We consider that regardless that it’s imperfect, enthusiastic about the impact of our research, the potential harms we foresee, and methods we are able to take to mitigate these harms is entering into the proper direction for the machine learning community. Throughout the project, we are going to document how we operationalize the values described on this document, together with the benefits and limitations we observe within the context of the project.
Content policy
Studying the present state-of-the-art multimodal systems, we foresee several misuses of the technologies we aim at as a part of this project. We offer guidelines on a number of the use cases we ultimately want to forestall:
- Promotion of content and activities that are detrimental in nature, akin to violence, harassment, bullying, harm, hate, and all types of discrimination. Prejudice targeted at specific identity subpopulations based on gender, race, age, ability status, LGBTQA+ orientation, religion, education, socioeconomic status, and other sensitive categories (akin to sexism/misogyny, casteism, racism, ableism, transphobia, homophobia).
- Violation of regulations, privacy, copyrights, human rights, cultural rights, fundamental rights, laws, and every other type of binding documents.
- Generating personally identifiable information.
- Generating false information with none accountability and/or with the aim of harming and triggering others.
- Incautious usage of the model in high-risk domains – akin to medical, legal, finance, and immigration – that may fundamentally damage people’s lives.
Values for the project
- Be transparent: We’re transparent and open in regards to the intent, sources of knowledge, tools, and decisions. By being transparent, we expose the weak points of our work to the community and thus are responsible and could be held accountable.
- Share open and reproducible work: Openness touches on two points: the processes and the outcomes. We consider it is nice research practice to share precise descriptions of the information, tools, and experimental conditions. Research artifacts, including tools and model checkpoints, have to be accessible – to be used throughout the intended scope – to all without discrimination (e.g., religion, ethnicity, sexual orientation, gender, political orientation, age, ability). We define accessibility as ensuring that our research could be easily explained to an audience beyond the machine learning research community.
- Be fair: We define fairness because the equal treatment of all human beings. Being fair implies monitoring and mitigating unwanted biases which might be based on characteristics akin to race, gender, disabilities, and sexual orientation. To limit as much as possible negative outcomes, especially outcomes that impact marginalized and vulnerable groups, reviews of unfair biases – akin to racism for predictive policing algorithms – ought to be conducted on each the information and the model outputs.
- Be self-critical: We’re aware of our imperfections and we should always consistently lookout for methods to higher operationalize ethical values and other responsible AI decisions. As an example, this includes higher strategies for curating and filtering training data. We must always not overclaim or entertain spurious discourses and hype.
- Give credit: We must always respect and acknowledge people’s work through proper licensing and credit attribution.
We note that a few of these values can sometimes be in conflict (for example being fair and sharing open and reproducible work, or respecting individuals’ privacy and sharing datasets), and emphasize the necessity to think about risks and advantages of our decisions on a case by case basis.
