Impact Statements in ICML Submissions
This year at ICML, we are instituting an ethics review process for papers. Both ICLR and NeurIPS have such processes, so they should be largely familiar to members of the community. However, since the precise nature of ethics and impact statements in papers differs slightly between the different conferences, we want to highlight the precise language regarding this point in the ICML paper submission guidelines.
The following is the language included in the call for submissions (https://icml.cc/Conferences/2024/CallForPapers):
Impact Statements:
Authors are required to include a statement of the potential broader impact of their work, including its ethical aspects and future societal consequences. This statement should be in a separate section at the end of the paper (co-located with Acknowledgements, before References), and does not count toward the paper page limit. In many cases, where the ethical impacts and expected societal implications are those that are well established when advancing the field of Machine Learning, substantial discussion is not required, and a simple statement such as:
“This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.”
The above statement can be used verbatim in such cases, but we encourage authors to think about whether there is content which does warrant further discussion, as this statement will be apparent if the paper is later flagged for ethics review.
Please note that this statement should be included in submitted papers (not just added later for camera-ready versions of papers which are accepted).
As mentioned above, such statements are fairly common in machine learning conferences at this point, so we hope that authors will be familiar with the broad intentions of such a section. Put briefly, it is becoming clear that many machine learning methods and applications have the potential for substantial impact on the world, both positive and negative. Improved performance in computer vision, for example, raises the possibility of both positive and negative uses of technologies like facial recognition. Generative models that produce text and images have well-established potential to generate misinformation or other harmful content, in addition to potentially positive uses.
Many submissions to ICML will involve methods or applications that directly touch upon such impacts — e.g., methods that directly investigate technologies like facial recognition, legal and ethical aspects of generative AI, or which could be used directly to produce harmful content from models. In these cases, the impact statement offers an opportunity for authors to discuss these potential harms. There is no set format for the Impact Statement section, but authors are encouraged to highlight aspects of both the potential positive and negative impacts of their work, and explain why they believe the net impact of the paper warrants publication.
At the same time, many submissions do not directly engage with applications and impacts of machine learning that have such potential. A paper analyzing convergence rates of optimization methods, for example, doesn’t require a detailed analysis of the downstream impacts of the work. For this reason, we also provide a statement that can be used verbatim, if the authors believe that the work’s primary impact comes from a more generic improvement to the field of machine learning as a whole.
The distinction between these two situations is not always clear-cut, and we leave it up to authors to determine whether they believe an individualized impact statement is required or whether the generic statement will suffice. Reviewers will be able to flag papers for ethics review based upon the content of the paper and the impact statement. The goal here is not to disqualify papers based upon the exact wording of the impact statement, but rather to offer a chance for authors, reviewers, and the community as a whole to highlight the multiple impacts of work in machine learning. If you have any questions, please feel free to contact the ICML Ethics Chairs or PCs for additional information.
Update: Although as noted above we are following a different procedure for ICML, several people have pointed out that the NeurIPS Checklist provides a nice set of guidelines that could be helpful in determining what topics may be of relevance in your paper for the impact statement (particularly section 3 of the checklist, but many other sections are relevant as well). We are not adopting such a checklist for ICML this year, but authors are encouraged to consult it if they feel it would be helpful.
