Samy Bengio, Alina Beygelzimer, Kate Crawford, Jeanne Fromer, Iason Gabriel, Amanda Levendowski, Deborah Raji and Marc’Aurelio Ranzato
This year NeurIPS launched a new set of Ethics Guidelines that are intended to support authors and reviewers when it comes to considering the ethical implications of their research. In this blog post we outline the motivation for this initiative, provide details about the process by which the guidelines were drafted, and highlight key elements of the guidance provided.
Overview of the Ethics Guidelines
Careful considerations of the ethical implications of one’s work ought to play a central role in AI research. First, whenever research is applied to real-world settings it has the potential to affect people’s lives. Our hope, as a community, is that the net outcome of research in our field will be positive for society, particularly for the most vulnerable. With this goal in mind, it is important that authors put their work in context and articulate not only the potential benefits of their work but also the potential risks. Moreover, when risks are identified we should consider what mitigation measures can be put in place.
Second, there are aspects of AI research itself that require adherence to protocols grounded in ethical principles. For instance, whenever research uses human-derived data, researchers should explain how authors obtained such data and whether consent was granted, whether care was taken to mitigate biases, etc.
The NeurIPS Ethics guidelines were developed to support authors and reviewers in the process of reflecting and documenting these ethical considerations. The guidelines are divided in two parts. The first part identifies reflection points for the potential negative societal impact of AI research. For instance, does the method or application directly facilitate injury of living beings or raise safety concerns? Or might the approach cause serious environmental damage? Here the authors are expected to see the big picture, put their work in context, and comment on foreseeable applications of their work.
The second part of the guidelines focus more directly on research best practices, such as using a reproducible research methodology and taking care when dealing with data, in particular with human derived data. For instance, we ask authors to comment on whether the data they use contains any personal identifiable information, whether the data might encode, contain or exacerbate bias against people of any protected characteristics such as gender, race and sexuality.
Why do we need Ethics Guidelines?
Last year, NeurIPS 2020 program chairs asked authors to include a broader impact section in their paper to include reflections on the societal consequences of their work. The peer-review process was also augmented with a separate track to handle papers flagged for ethical reasons.
This year, the NeurIPS 2021 program chairs consolidated the effort started in 2020 by requiring authors to include a checklist in their submission. As outlined in the call for papers, “the paper checklist is intended to help authors reflect on a wide variety of issues relating to responsible machine learning research, including reproducibility, transparency, research ethics, and societal impact.” There was also a dedicated review process that handled papers flagged for ethical reasons.
While NeurIPS has pioneered the effort to include ethical considerations in the peer review process, there were previously no publicly available guidelines about what practices or outcomes would be considered ethically relevant to submissions. Last year’s effort can be seen as an experiment in learning for the field, which we used to gather feedback and refine our criteria. For instance, last year there was an entire NeurIPS workshop dedicated to the theme of how to think about the broader impact of machine learning research. Leveraging this discussion and in response to requests to provide more structure and transparency to our guidelines (Nanayakkara et al., 2021; Prunkl et al., 2021; PAI, 2021), this year our Ethics Guidelines provided more concrete guidance to both authors and reviewers, making explicit the evaluation criteria for the ethics review and helping authors reflect on these important topics prior to their submission.
The NeurIPS Board gave the General Chair the mandate to form a committee to draft ethics guidelines. The committee includes members who cover a large spectrum of views and expertises, from leading researchers in AI ethics Kate Crawford, Iason Gabriel and Deborah Raji to world renowned legal scholars Jeanne Fromer and Amanda Levendowski to machine learning researchers Alina Beygelzimer and Samy Bengio. Prior to the submission deadline the group worked hard to compile guidelines starting from those shared last year with the ethics reviewers. We wish to acknowledge also precious feedback and contributions from the NeurIPS 2021 program chairs as well as from the members of the NeurIPS Board.
The goal of the Ethics Guideline is to provide authors and reviewers (and also to the broader machine learning community) with material to facilitate ethical reflection. We are mindful that ethical considerations are often nuanced, and that people may sometimes disagree on the severity level of certain issues. In our view, this is a natural and potentially productive state of affairs. What matters is that we start a vibrant conversation among members of our community on these topics – and that we are able to develop appropriate precedents and standards that support responsible practices across the field as a whole.
As Charles Isbell articulated in one of last year’s keynote talks, AI research needs to be put in context. Researchers need to develop a deep appreciation of where data is coming from, how technology is going to be used, who benefits from it and who may not. In general, we aim to provide our community with a safe space to discuss these issues, and believe that with transparent discussion between authors, reviewers and readers, our community can make faster progress towards the responsible use of AI that is widely beneficial.
These guidelines are a living document which are subject to further refinement and iteration. They will evolve in tandem with our community’s wider understanding and sensibility towards AI ethics. If you have feedback for us in relation to these guidelines or simply have questions on how they apply, please reach out to us.