Professor of Economics and Finance, Frederic Esser Nemmers Chair, Co-Director of the Global Poverty Research Lab at Kellogg
Online platforms often feature users’ profile pictures to build trust and connection.
Fundraising sites like Donors Choose, for instance, commonly display photos of the teachers whose classrooms will benefit from the donations, while homesharing services like Airbnb often include photos of both guests and hosts in their respective profiles.
“If you want to make a reservation in a hotel, nobody would ask you to send a personal photo,” says Emil Palikot, a postdoctoral researcher at the Stanford Graduate School of Business. But “online platforms give big prominence to personal images.”
One major problem, though, is that these images can enable discrimination: users might not offer services to people based on their race, gender, or other characteristics. Yet removing photos—as some critics have advocated—might discourage many people from using the platform at all.
In a new study, Dean Karlan, a professor of finance at Kellogg and codirector of the school’s Global Poverty Research Lab, along with Palikot, Susan Athey at the Stanford Graduate School of Business, and Yuan Yuan at Carnegie Mellon University, investigated a different approach to reducing discrimination: offering guidelines for the photos that would benefit any type of user, but would be designed to disproportionately help disadvantaged groups.
The team focused on the microlending site Kiva, which connects borrowers to lenders. First they looked at a number of the stylistic features of the borrowers’ photos that were in a borrower’s control, like whether a photo was taken inside or outside. They found that some of these features were associated with raising money more quickly.
Then the researchers confirmed that different demographic groups used these features differently, on average. This meant that members of already-disadvantaged groups on the platform could actually be punished twice: once by discrimination, and then again by their tendency to prefer a certain style of photo.
By encouraging all users to take photos using the same, optimally effective features, the researchers suggest, online platforms can do two things at once. They can increase the efficiency of the platform overall—a particularly worthy goal for sites like Kiva, which aim to benefit people living in poverty—and they can reduce any inequities that are exacerbated by differences in the types of photos that different groups post.
“There are some features of profile photos—ones that people have control over—that could be contributing to inequality on these sites,” says Karlan, who is also chief economist at USAID. “Platforms that design policies based on images need to think about this.”
The team’s research provides recommendations of steps platforms might take to reduce disparities between user groups. “The size and reach of online platforms mean that inequities on these platforms affect outcomes for a lot of people,” Palikot says. “We hope that our research moves the ball forward, providing methods that quantify the extent of the problem and evaluate alternative approaches to mitigating it.”
This line of research began when another team reported in 2014 that Black Airbnb hosts in New York City charged 12 percent less than hosts of other races for similar properties. The discrepancy apparently arose because Black hosts, whose photos were visible on their profiles, had a harder time booking guests.
Similar patterns have been attributed to discrimination based on users’ names. For instance, Airbnb guests whose names sounded typical of Black people were 16 percent less likely to receive positive responses from hosts than guests with typically white names. And UberX drivers were about twice as likely to cancel a ride when they found out the passenger had a distinctively Black-sounding name, compared with passengers with white-sounding names.
The magnitude of these inequities is “striking,” Karlan says.
The researchers decided to investigate the effect of choices that users could control about their photos, which went beyond demographic characteristics. If the researchers could identify which styles of images were the most compelling to other users, perhaps they could recommend policies for profile photos that—in combination with antidiscrimination practices—would help even out inequities.
The team obtained data on more than half a million borrowing campaigns on the microfinance platform Kiva from April 2006 to May 2020. Borrowers, who typically lived in areas with little access to financial services, asked for small loans to support their businesses, education, or families. Individual lenders could give loans to borrowers based on the information on the borrowers’ profile pages.
To find out if lenders’ decisions were partly driven by borrowers’ profile photos, the team ran software that focused on 55 features in the images. These included characteristics that were largely fixed, which they called “type” features: gender, face shape, race/ethnicity, and so on. They also included changeable “style” features, such as wearing sunglasses or using a flash.
“The size and reach of online platforms mean that inequities on these platforms affect outcomes for a lot of people.”
Next, the team analyzed whether any style features were linked to receiving more cash per day, controlling for all other features identified by the software. They found that certain style choices, such as using an outdoor background, posing for the photo, or smiling, tended to boost funding; other features, such as using flash or having the person’s body occupy most of the frame, had a negative effect. For instance, taking the picture outside increased the average daily cash collected by about $9, and body-shot framing reduced cash per day by $10.
Importantly, such features were not linked with the probability of a borrower repaying the loan. This suggests that stylistic choices lead to inequities that are not explained by the lenders’ underlying risk of investment, Palikot explains.
The researchers then tackled a trickier question.
Consider the scenario in which one demographic group tends to get fewer benefits from a platform. It’s also possible that members of this group, on average, make similar photo-styling choices—choices that differ from ones made by other groups. For example, older users might take pictures in a different way than younger users.
If the disadvantaged group chose less-compelling style features, those choices could widen the discrimination gap. If they instead made choices that were more compelling, they could narrow the gap.
To look for these patterns, the researchers examined style differences linked to gender. On Kiva, male borrowers’ campaigns tended to be less successful; on average, women raised $36 more per day than did men. The team found that men were also less likely to smile and more likely to use body-shot framing.
Lenders may have preferred to give to female borrowers for other reasons. For example, women generally have less access to entrepreneurial finance. But one factor seems to be that “women are just better at making impactful images,” Palikot says.
To look deeper into this finding, the team ran a controlled experiment to evaluate the effect of gender, smiling, and body-shot framing. To do so, they used software that can create “deepfakes,” convincing alterations of photographs. For example, they could take a profile picture that resembled ones used by female Kiva borrower and make the person look male. Or they could take a photo of a person with a serious expression and make it look like they were smiling.
Then they recruited 400 people on the online research platform Prolific.co who had previously donated to charity. Each participant was shown six pairs of fabricated images and asked which picture they preferred.
The experiment confirmed the researchers’ earlier results. Study participants were 31 percent less likely to choose pictures of borrowers who looked male. Adding a smile to a person’s photo increased the chances of being selected by 34 percent, while using a body shot decreased the chances by 17 percent.
Finally, the team considered which steps could be taken to counteract inequities. Online platforms like Kiva often have to balance fairness and efficiency, the researchers say. By “efficiency,” they mean how much the platform is achieving its intended goal—for Kiva, that might be the amount of money given in loans.
A platform that prioritized fairness could boost a disadvantaged group by making their profiles more prominent. But if that group tended to choose less-attractive style features, on average, then efficiency on the platform would drop.
Alternatively, platforms could suggest that all users make more attractive style choices. If more people avoided body shots, for instance, efficiency would go up. Fairness could increase too, if the gap in photo styles between demographic groups would narrow.
For example, consider a scenario in which 75 percent of people with indoor photos responded to the platform’s new recommendations by swapping in outdoor photos. If most of those users were in the disadvantaged group—because that group was less likely to use compelling photo features to begin with—then the difference in the fraction of people with outdoor backgrounds between groups would shrink.
Platforms could also use a combination of policies. If they wanted to emphasize fairness and were willing to take a hit in efficiency, they could issue photo-styling recommendations to everyone while also bumping up the disadvantaged group’s profiles.
“It’s up to the platform to what extent they care about one versus the other,” Palikot says. The decision isn’t necessarily clear-cut; for example, higher efficiency on Kiva means more help overall for impoverished people.
It’s also debatable whether the gender gap in borrowing campaigns would be considered an inequity that should be corrected, given women’s lower access to financing and the gender-based discrimination they likely encounter in other parts of their lives. It’s up to Kiva to make the call, Palikot says. He says the team used the gender gap as an example, but their methods could be applied to other groups.
The study shouldn’t be taken to mean that the burden is on users to counteract discrimination by changing their profile pictures. Platforms need to keep pushing and enforcing antidiscrimination policies, Palikot says. For example, Airbnb has taken some steps in that direction, such as showing guest photos to hosts only after a booking is confirmed and kicking out users who won’t agree to commit to inclusion.
But photos may not disappear entirely, and the reality is that antidiscrimination policies haven’t completely eliminated gaps. Any small changes that continue to reduce inequities “are also a change in a good direction,” Palikot says.
Athey, Susan, Dean Karlan, Emil Palikot, and Yuan Yuan. “Smiles in Profiles: Improving Fairness and Efficiency Using Estimates of User Preferences in Online Marketplaces.” Working paper.