Working with Nude Images in Research:
Guiding Questions & Recommendations
Using nude images in research poses significant ethical risks that should be carefully evaluated at all phases of the research lifecycle—project formulation, execution, review, and publication. No research that uses nonconsensually collected nude images will be completely without harm. However, when researchers have no alternatives and a compelling need to conduct work with nude images, there are several things researchers, reviewers, and publishers should consider in planning and evaluating the work. This page provides prompts to help researchers, reviewers, and publishers think through these challenges. There are no easy answers. Many of the questions will appear in multiple sections, as each phase of the research should evaluate the work.
This is a living document that we offer as a starting place. We welcome additions, suggestions, comments, and critiques, because ethical guidelines should be something all members of the community are invested in. Feel free to email us at intimate.datasets@gmail.com.
What do we mean by nude images?
Nudity is not inherently offensive, and what is considered nude or sexual is subjective. There is a long, fraught history of legal and policy efforts to define nudity, which researchers should familiarize themselves with before embarking on research leveraging these concepts.
Because these guiding questions are centered on minimizing harm to image subjects, we suggest that researchers consider any images that may be considered nude, intimate, sexual, or physically violating by image subjects to be in scope. For example, datasets of swimsuit photos, which are sometimes collected as “challenging images” in nudity detection, may be perceived as violating if they were taken and/or shared without consent.
Researchers
If you are a researcher considering conducting a study using a dataset of nude images, these questions will help you think through the ethical challenges of the work. The prompts are framed as questions because there are no “right” answers – researchers should use them to try to find the least harmful path forward.
Who will the project benefit?
Research should avoid treating individuals as a means to an end by risking their well-being in pursuit of a broader societal benefit [1]. Will this work benefit the individuals placed at risk? Will this work help mitigate the harms experienced by people whose intimate images are misused? Will this contribute to the marginalization of sex workers, whose images are likely being used?
Are you working with relevant stakeholders?
Research that claims to benefit society should ground this claim in evidence and look for solutions that contend with the real-world constraints of the problem. Are the project specifications and goals based on evidence or on value-laden assumptions about the problem? Are there opportunities to work with stakeholders (e.g., victim-survivors, industry, law enforcement) to ensure that the final system will be useful? Building partnerships may also enable researchers to leverage existing data that stakeholders may already have access to (e.g., previously flagged content).
Is it necessary to collect a new dataset?
Could the study be completed with proxy images rather than real nude images? Could synthetic images be used instead of real images? Note that synthetic images are likely generated with models trained on nonconsensual images, and that this approach still requires careful ethical evaluation.
How will images be protected during the study?
If there is a carefully assessed research need for nude images, the research team should create a concrete plan for data handling. This should account for secure data storage and a protocol for data access, especially if manual annotation or inspection of the images is required. In particular, data handling plans should include (1) a secure data storage and access scheme, (2) the policies that all annotators and researchers should adhere to when interacting with nude data, including what tools should be used, (3) procedures for how the team will handle any discovered nonconsensual or abuse content, including legal and mental health resources for researchers and workers who have to interact with such content, and (4) access removal after annotation or manual validation is complete.
How will images be protected after the study?
Datasets in AI/ML are normatively public to facilitate reproducibility, open science, and data democratization. However, the redistribution of nude datasets beyond the team escalates harm and constitutes image-based sexual abuse. By redistributing this data, researchers increase its visibility and create new avenues for dissemination. In some jurisdictions, such distribution could also be illegal.
In nearly all cases, researchers should delete nude datasets after the project is completed.
Does the study require illustrative examples?
Researchers should not include identifiable example nude images in their published papers without the consent of the subjects depicted, no matter the context. In many cases, relevant images could be described with text alone. If visual examples are absolutely necessary, consider using stylized art that will illustrate relevant image features, such as those in Hamilton et al. (2024).
There are few, if any, foreseeable reasons that would necessitate publishing real example images. If such a case were to arise, it should be extensively discussed with ethics boards, reviewers, and publishers. Researchers should further take all available steps to ensure the image is (a) not identifiable and (b) the image subject's consent has been obtained.
How are these ethics decisions being communicated in the paper?
Papers describing studies that use nude images should acknowledge the risks of using the images and what authors did to lessen that harm. Clear communication about ethics decision-making will also help set field norms. In particular, papers should disclose that the nude images were collected without the knowledge of the subjects depicted and discuss (1) why the study’s benefits surpass the study’s harms in using real nude images and (2) the concrete steps that the authors took to mitigate the harms that emerged from using nude images.
Reviewers
These questions will help reviewers evaluate the research practices of papers using datasets of nude images. When in doubt, seek guidance from ethics reviewers or conference chairs.
Problem Framing
Who does the research purport to benefit? Is there evidence to support this problem & this approach to addressing it?
How do the authors weigh the risks of using the data against the benefits?
What concrete steps have the authors taken to ensure that the benefits will be realized, particularly to the individuals who experience risk as a result of the study?
Data Sources
If the authors created their own dataset:
Did they document the data sources and their reasoning for choosing them?
Did the authors describe any measures taken to ensure that abuse material was not collected?
If the authors use a pre-existing dataset:
Did the authors review the sources originally used to create the dataset?
Dataset Access
If the authors plan to share the dataset:
Did the authors describe their methods for securely storing the dataset?
Did the authors describe how they will decide who to grant access to and for what purposes?
Example Images
If there are nude images included as examples:
What is the authors’ justification for publishing images?
Are they necessary or are stylized depictions/art sufficient?
Did the authors receive permission to publish the images? Did they take all possible steps to ensure the images are not identifying?
Dataset handling
During the course of the research, who do the authors state as having access to the dataset? Does this include people who are external to the team? (e.g., data annotators)? Are there any additional details provided?
What measures did the authors make to ensure that the data is securely stored?
AI-generated images
If the authors use AI-generated images, do the authors describe their process for obtaining the images (e.g., what model is used to create the images)?
Do the authors discuss the risks of using a generative model that was likely trained on nonconsensual nude images?
Publishers
Frontier venues and major publishers such as NeurIPS, ACM, and IEEE set standards for the entire computing community, and have the opportunity to take a strong stance on nonconsensual nude imagery in research publications.
We urge publishers to:
Create guidance downstream to journals, conferences, workshops, etc. against nonconsensually publishing identifiable nude images
Conduct an internal review of published works on topics related to nudity detection to identify papers that nonconsensually publish identifiable nude images and work with authors to either redact or remove the images
Ensure that victim-survivors who have had images published without their consent have a clear channel for reporting and receive priority responses