3 Strategies for Accountable, Ethical Online Behavior Research
Help CivilServant develop ways to inform people about their participation in online research and hold us accountable
In 2014, after researchers worked with Facebook to test the effect of newsfeed adjustments on the emotional tone of people’s future posts, academics took a closer look at the ethics of online behavioral research, in the midst of a wider public debate over the power of online platforms in society.
Two ideas were central to these conversations: consent and debriefing. In consent-based models of research, people are asked in advance if they are willing to participate in the study. Individual consent often works best under controlled, lab-style studies or surveys and interviews, where it’s easy to decide which people are part of a study and which people aren’t. Debriefing is a process where people are told after the study. Debriefing is also a way to identify any unexpected, harmful effects that the researchers weren’t looking out for, so the harms can be addressed.
In field research, which tests ideas out in the world, individual consent and debriefing can be hard to acquire. For example, consider this study that tested the effect of lawn signs on voter participation rates. It wouldn’t be possible to obtain the advance consent of every single driver who passed by the signs; it would be impossible to predict exactly who would drive by. Even if you could obtain consent, you wouldn’t be able to show or hide the sign for people who hadn’t consented to the study. Likewise with debriefing: a researcher might be able to place a camera next to every sign in order to figure out the license plate, identity, and address of everyone who passed by, but in the effort to contact everyone in the study about ethics, the ethics procedure might become more risky and intrusive than the original study.
How can we make progress on these research design challenges? While new ideas in law and ethics can clarify general principles, many of our debates rely on assumptions about design and about how people are likely respond to research.
when people talk about research ethics, they’re often talking about questions of power
In recent years, a growing number of scholars have started to prototype new ways to manage research ethics to see how well those approaches work. One of those people is Scott Desposato, who recently conducted a series of nationally-representative surveys asking Americans what they thought about the ethics of common approaches to public service announcement research (video). Desposato compared academics’ views on research ethics with residents’ views for studies about flossing, voter participation and DUI. Surprisingly, while everyone was less comfortable with research that didn’t ask for consent, more residents were comfortable with those studies than academics.
In Scott’s study and his other writings on solutions to ethical challenges, he points out that when people talk about research ethics, they’re often talking about questions of power–something that danah boyd has also argued.
Individual consent is one way for people to have some power around research; holding researchers accountable to local community review or elected representatives might be another. A researcher might not be able to reach every single person who’s part of a study, but the researcher might still be held accountable to the community that hosts a study. That’s what I do with CivilServant studies on reddit, which I do at the invitation of a community’s volunteer moderators, and which always include community-wide debriefings one the study is over.
Reading Scott’s work (and his edited collection on experiment ethics) I have been inspired to use the tools of design and research to improve on research ethics, finding creative solutions that hold our research accountable to the public while also continuing to grow public knowledge on important issues.
And that brings me to a practical question that I’m facing right now, together with my collaborators Merry Mou, Jonathon Penney, and Jonathan Zong.
Debriefing & Consent in Research Mitigating Negative Effects of AI Copyright Enforcement
Every day, AI-based law enforcement affects thousands of people when corporate-run machine learning systems guess if the things we say online have been copyrighted, and whether they should be taken down. These AI copyright enforcement regimes very likely have a chilling effect on people’s exercise of their speech rights when they withdraw from social media after receiving scary legal documents about their online postings.
In a study with Merry Mou and Jon Penney, we want to test ways to mitigate this chilling effect and help people flourish online even after an encounter with AI copyright enforcement. We believe that providing people with accurate information about the law could prevent them from becoming overly fearful and withdrawing from social media. We can reach people using information from the Lumen database, a public archive of every copyright takedown notice that publishes data on copyright enforcement in near-realtime. But before applying this idea everywhere, we want to be sure it’s actually helping. That’s why Jonathan Zong and I are planning to do an experiment to test our assumptions.
Of course we’re not going into this project blind. Our research is based on years of conversations and surveys by Jon Penney with people who have faced copyright takedowns. But there’s a catch: because we don’t know in advance who is going to receive copyright takedown notices, we can’t ask the people in this experiment for consent in advance. That leaves us with four options overall:
- Don’t try to help people who experience this problem
- Offer people support and information, but don’t ask if it’s actually helping them
- Test the idea without informing people they’re part of an experiment
- Design some novel kind of ethics process and accountability into this research, such as debriefing
The fourth option is the empirical ethics approach, one where we imagine, prototype, and test creative designs for ensuring the accountability of our research.
3 Strategies for Redesigning Consent and Accountability In Online Behavioral Research
Here are some of the strategies we are considering, and actually hope to test before settling on a process for this and future CivilServant studies on Twitter. I would love to see your feedback in the comments.
In online research, it’s sometimes possible to do things like “Community IRB,” where community representatives of participants review and approve a study before it runs. Community review is a basic requirement for CivilServant research on reddit, where communities have clear, if porous boundaries and leadership. Desposato writes about this idea in his research on local review of political science experiments. On Twitter, with a study that responds to people at their moment of need, we have no leaders to appeal to.
We may still be able to find a group of people to ensure that we balance the risks and benefits of similar research on Twitter. For our DMCA study, we could always reach out to people whose content was taken down in the past and ask them their views about our study design. We could even work with a high-quality, representative sample of those people, ensuring that if copyright enforcement continues in similar ways during our study, the people who participate in our experiment will be statistically similar to the people who we reached out to for feedback.
Once we find those people, we could potentially ask them to play any number of roles, including:
- Offering feedback on the risks of the study design, including where we do and do not intervene
- Discussing the study design with each other
- Voting on study designs
- Voting on whether they think someone like them should be included in or excluded from the experiment, especially if we have a way to automatically identify accounts like theirs
- Voting on whether we should halt or continue the experiment if we do experience any serious complaints
Oversight of this kind takes time and energy from participants, and involvement could become prohibitively time consuming. Furthermore, votes have numerous problems. Majority votes don’t always protect minorities, and higher thresholds can hold the process hostage to minorities. Creative design may be able to overcome these issues. For example, we could subdivide the sample and ask people to vote on whether their subgroup should be included in a study.
Because it takes time to participate in review of this kind, we would need to find ways to make the review process efficient and also ensure that people without the training or time will be able to hold our work accountable.
Imagine now that we decide to test the effects of 50 different messages for mitigating the chilling effect– we can’t reasonably expect people to vote on each message. Instead, we might ask people to review kinds of experiment and then provide audits of how well we kept within the agreed parameters.
Because we’re doing internet research, we have the ability to contact people essentially for free. While we can’t be sure that each person will receive or pay attention to the message (the same is true for any field experiment), we can technically carry out study debriefings at a much larger scale than ever before.
Imagine, for example, that we were to send a tweet to everyone who was part of the study telling them they were in the research and inviting them to click a link to learn more about it. We could require people to log in to the debriefing website and provide them with personalized information about exactly how the research worked, what we collected, and asking them to tell us about any harmful effects they might have experienced.
When I pose the idea of automated debriefing to researchers, they immediately start thinking about the problems with it: it won’t reach everyone, people won’t understand the debriefing materials, and maybe other people will eavesdrop on the message we send when debriefing them. Weirdly, people usually bring up these issues as an argument for keeping people in the dark about the research.
Fortunately, all of these objections rely on assumptions about people’s experience that can be tested empirically. How many people will learn about the research? How many will click through? How many people eavesdrop on our debriefing recruitment? Can we obfuscate the risk of eavesdropping by sending information about the research to a wider sample than those who participated, and only revealing the information privately once they click? These are all design questions that we can test.
Post-Study Opt In and Out
While debriefing processes inform people about the research, opt-out processes invite people to make a choice about how they want their information to be associated with the research going forward. There are many options:
- Choose how you wish to being mentioned by name in published research (CivilServant never names anyone unless they explicitly request to be named)
- Choose if your information will be anonymized or obfuscated in the data accessed by researchers
- Choose if your information will be anonymized or obfuscated in any datasets made public by researchers
- Choose if your information will be included in the aggregate results of the study in any way
- Provide the researchers with information that might lead your information to be automatically deleted, for legal or ethics compliance reasons (such as information that you are a minor, for studies that haven’t been approved to include minors)
Choices by participants can be hard for researchers to accept, especially if they affect the quality of the data: if enough people of the same kind choose to have their information removed, the researchers might come to an inaccurate conclusion. For example, imagine that people from a vulnerable group experience negative outcomes and decide to withdraw their data from the study. If the harms were limited to that group, the statistical models might conclude that the idea was beneficial. Without that information, decision-makers might take future actions that reinforce those harms, precisely because people decided to prevent the harms from being shared.
Designing understandable, ethical debriefing and opt-out interfaces is central need for publicly-accountable research
When offering people choices about how to participate, researchers have a duty to explain the full implications of those choices for the research and any decisions that might be made in response to the research. That itself can be a difficult challenge in cases where participants aren’t familiar with statistical methods. Designing understandable, ethical debriefing and opt-out interfaces is central for publicly-accountable research, since we need to know how well people understand the choices we offer them.
In opt-out processes, as with representative review, we might also imagine designing procedures that allow us to infer the preferences of people who were part of the study and apply them evenly across everyone who was in the study. For example, if 30% of people who clicked ask us to delete their information, should we do that only for them, or 30% of everyone in the study? There might be a good case for treating those who responded as a representative group, if they are actually statistically-representative.
Research ethics become more complicated for research that looks at communities in conflict, as Brian Keegan and I have discussed elsewhere. For example, if a study about online harassment offered every participant equal opportunity to opt out, then a dedicated group of harassers could skew or sabotage results by deciding to opt out en-masse. At the same time, I strongly believe that every person in social research, including alleged harassers, has common rights and deserves a meaningful voice in how that research is conducted. Getting the balance right is far from easy, but the path of transparency has never failed CivilServant so far. I am often moved by the thank-you letters I get from distrusted parties when I respect what I consider to be their fundamental rights to autonomy and privacy.
Testing Research Ethics Procedures
Throughout the academic year, Jonathan Zong and I will be testing some of these research ethics procedures with Twitter users. We’ll be asking people to imagine they were being contacted about a hypothetical study, and then we’ll ask them to work through different webpages and social process for accountable research. We’ll ask for their consent to record their ideas and responses, and report the outcomes of those different ideas.
We still have many, many questions, so we eagerly await your responses to this blog post and your ideas for collaboration.