|June 30th, 2014|
On the other hand, experiments like this are pretty standard for companies to do. When a chain wants to introduce a new sandwich they might divide their metropolitan areas into matched pairs, and then randomly pick an area in each matched pair to get the new sandwich. Or when a site is trying out a new design they might A/B test different ones to see if it does better than the old one. Or when Facebook gives 1% of people larger feed images and sees whether they use the site more or less than a 1% control group, in order to see whether this is a feature they should roll out to everyone.
When someone says "running experiments on people without their consent" we think of the Tuskegee syphilis experiment or the Nazi medical experiments. And those experiments were horrible. The unethical aspect, however, wasn't that they randomly divided populations into "experimental" and "control" groups and treated the two groups differently, it was the particular actions they did to those groups. Imagine they had done those things to all the people, and not collected any data. There would have been no 'experiment' but clearly this wouldn't be any better.
So here's a rule I think is about right for figuring out how ethical an experiment is: if A and B are two ways you could treat people, and instead of picking one to do for everyone you randomly choose whether each person should get A or B, this is no less ethical than the worse of A or B. So because the sandwich chain above could ethically roll out their new sandwich to everyone, or not roll it out, there's nothing wrong with them running a partial rollout experiment. But injecting dye into people's eyes without their informed consent isn't ethical, so an experiment where you randomly do this to prisoners is clearly unethical.
For examples of experiments I would consider to be completely ethical but didn't have any kind of informed consent, take cash transfer research (ex: Haushofer and Shapiro 2013). They gave people money, didn't say why, and then followed up to see what happened, comparing them to other people they didn't give money to. Or consider Zimmerman 2003 where Williams College randomly assigned students to roommates and then looked at whether having a roommate with a lower SAT score lowered academic performance. Or take an earlier Facebook study where users randomly saw a message encouraging them to vote, and (checking against voting records) found that they did. In all three of these cases the participants didn't give consent and had no option to withdraw, but because the two options (money or no money, lower or higher SAT-scoring roommate, vote promotion or none) were each acceptable options on their own, randomly choosing between them didn't make the options any less acceptable. Even though the change to people's lives was much larger than in the controversial Facebook study these are clearly ethical experiments to run.
So back to the new Facebook study. Clearly it's fine for Facebook to have some sort of algorithm where they choose what shows up in your feed. Users wouldn't like it if they got absolutely everything all of their friends posted, because there's far too much to read. And they wouldn't like a random selection, because they're more interested in the stuff some of their friends say than others. So after observing that you click "like" more on posts by X they start preferentially showing you posts by X. It's probably even fine for them to do this based on the content of your posts: if you post "got married" or "it's a girl!" that's a good clue that lots of people will want to know and they should show it widely. (Though the normal "more people see stuff with more likes" may handle this fine already.) This does give them a tremendous amount of power, however. For example, they could promote posts saying "voted for Q!" at the expense of posts saying "voted for R!," and this would be all sorts of wrong. So if some potential newsfeed selection algorithms are ok, and others aren't, where does this one fall?
If Facebook decided to slightly favor sad posts because they were worried that people were unwilling to 'like' negative but important messages, and that their existing like-centric algorithm was suppressing negative messages, that would probably be ok. On the other hand, favoring sad posts simply because they enjoy making people miserable would not be. This is unlike the sandwich chain example, because whether it's ok depends on their motivations, and puts it right on the edge of acceptable changes. In this case their reason seems to be to "investigate the common worry that seeing friends post positive content leads to people feeling negative or left out" (ex: Kross 2013). This seems good enough to me, though some people find manipulation of emotional status dangerous enough that they don't think any reason Facebook could give is sufficient. Regardless, this is the question for determining whether their actions were ethical, and while informed consent would have helped, many other experiments are completely ethical without it.
(I do think the study has problems, apart from its ethical status. It's trying to measure whether emotional states are being transferred, but all it's actually able to measure is whether a person who observes more messages with sentiment Z produce more messages with that sentiment. This could be emotional contagion, and probably is in part, but I think most of it is probably just people adjusting their words to their perceived surroundings. If you walk into a room where everyone is whispering, you whisper too, but that doesn't mean your emotions have changed. The closest they get to addressing this is with "This is not a simple case of mimicry, either; the cross-emotional encouragement effect (e.g., reducing negative posts led to an increase in positive posts) cannot be explained by mimicry alone, although mimicry may well have been part of the emotion-consistent effect." But if I notice everyone around me is happy, and respond by avoiding saying sad things, this could just be because I'm trying to fit in, not because I've been cheered up.)
(Prompted by the discussion on this post.)
Update 2014-07-04: Relevant xkcd: 1390.
Follow-up: Experiments and Consent