When massive open online courses, or MOOCs, exploded in popularity in the early 2010s, educators were particularly excited about the courses’ potential to give disadvantaged students equal access to a quality education.

But a bevy of recent research has shown that online learning has largely fallen short of that goal. The same factors that have held back low-income or minority students in physical classrooms also plague virtual ones. Studies have found that online-learning resources had trouble attracting low-income students—or, in the case of school-age children, their parents—and that those who did participate in online classes performed more poorly than their peers.

These findings come at a time when the “digital divide,” the gap between high- and low-income families when it comes to internet access, is narrower than ever: Nearly 90 percent of Americans now have internet access, according to census data, and among Americans between the ages of 18 and 29, that number is 99 percent. But researchers have found that it takes more than just access to an internet connection to benefit from online resources: It also takes basic technology skills, the ability to draw on social networks for help and guidance when needed, and a willingness to look to the internet for information and resources.

For a study released this week, Pew Research Center interviewed more than 2,750 people who have access to the internet. They devised a new measure of connectivity called “digital readiness,” and arranged internet users into five distinct categories based on their readiness. Three of those categories contained people who are “relatively hesitant” to use technology for learning, even though they have access to a reliable internet connection.

“The data show that there are real barriers to drawing people to use digital resources for learning,” said John Horrigan, the Pew study’s lead author. The main barriers that Horrigan and his team identified were a lack of digital skills and a lack of confidence in one’s own ability to find trustworthy information on the internet.

Horrigan named the least technologically savvy group “the unprepared.” His study found that the 14 percent of Americans who fell into this group are the least likely to use online learning tools, due to the barriers he identified.

A small group of Americans—just 5 percent—were categorized as “traditional learners.” Even though they’re interested in learning outside the classroom, they prefer to avoid using resources tools to do so.

At 33 percent of Americans, the largest single group of internet users is made up of “the reluctant.” They’ve got better computer skills than members of the “unprepared” group, but they’re very unfamiliar with educational resources online and where to find them.

Altogether, the study found that just over half of Americans who have access to the internet don’t have the tools to use it effectively, or at all, for learning. And when the Pew researchers examined the demographic groups that were most likely to show up in the three “hesitant” groups, they found a handful of traditionally disadvantaged groups: minorities, women, and lower-income households.

The same types of divisions appear to be affecting enrollment in online courses. Although free and open to all, MOOC enrollment is biased toward the already highly educated and the relatively wealthy, according to a pair of studies from researchers at Harvard and MIT.

One study published last year examined the average neighborhood income of enrollees in nine Harvard MOOCs by matching up students’ mailing addresses to census income data. The researchers found that the neighborhoods that MOOC students lived in had an average income of $14,000 more than the average neighborhood in the U.S.

In another analysis, which examined more than 150,000 students who took at least one of 68 online courses that Harvard and MIT offered between 2012 and 2014, the same authors estimated that for every $20,000 added to the average income of a person’s neighborhood, the odds that the person would enroll in an online course increased by 27 percent.

There are a lot of reasons why this might be the case, says John Hansen, a researcher at Harvard who worked on the studies with MIT’s Justin Reich. Higher-income college graduates might dive into a MOOC for fun, Hansen said, or they might be willing and able to take on a class that won’t earn them a formal degree. That’s less likely to be the case for lower-income students or those without the resources to learn for recreation.

Better-educated and higher-income students are also more likely to know where to look for the best online classes. That’s according to Di Xu, a professor of education at the University of California, Irvine, who studied online courses in the context of community colleges for an extensive research project at Columbia’s Teachers College. “Such awareness or information may come from personal knowledge, social network through friends, colleagues, family members, etc.,” Xu said. “Students from low-income background on average may have less access to such information.”

When they examined enrollment in the Harvard and MIT MOOCs, Hansen and Reich also found that students’ socioeconomic backgrounds helped predict their course performance. For adolescent students, parental education levels made a big difference: Students with at least one parent with a bachelor’s degree were 1.75 times more likely to complete a course and earn a certificate than students who don’t have a parent with a bachelor’s degree.

Parents’ outsize roles in their children’s learning patterns are apparent in even younger kids, too. Betsy DiSalvo, a professor and researcher at the Georgia Institute of Technology, studied families with elementary- and middle-school children. She found that parents in lower-income families were less likely to use the internet to find learning resources for their kids, or to find information about how to help them with school.

Even though they had access to the internet, parents in lower-income families didn’t have the digital skills to find the best online resources, DiSalvo found. Many who were particularly uncomfortable using technology were afraid to visit sites or open emails that would infect their computers with malware. “I no longer take emails that I don’t know anything about, because that’s like the only computer we have,” a mother of five school-age children, who’s working toward an online degree of her own, told researchers.

Parents who spend a lot of their time thinking about the basics of supporting a family may not have the resources to dive deep into supporting their children’s education. “I think it’s tied to this idea of Maslow’s hierarchy,” DiSalvo said. “Their first priority is getting their kids to do well in school. They’re so focused on that that they aren’t necessarily focused on looking at what’s fun engagement, or what’s going to spark their interest.”

If young students from lower-income families aren’t getting the online support they need to keep up in school, and are less likely later in life to sign up for online courses—or, if they do, to excel in them—it raises the question: Is online learning actually an equalizing force?

Hansen, the Harvard researcher, calls the answer to that question the “MOOC paradox”: Democratizing education could actually compound existing inequality.

“It's heartening that a bright student in a small town with under-resourced schools could take a great computer science or philosophy course for free,” he said. “But if it’s a great course, we should expect that students in good schools with well-educated parents will enroll, too, and probably at higher rates.”

The various researchers I spoke to agree that the solution would require online-course providers and policymakers to reach out to lower-income and disadvantaged students. Simply releasing a great, free resource into the online wild hasn’t been enough to overcome the systemic forces that have segregated outcomes in physical classrooms. Universities use affirmative action to balance out those forces—online learning, for the most part, hasn’t emulated that model.

Xu is part of a team at UC Irvine that received a five-year grant from the National Science Foundation last year to research teaching and learning techniques in online environments. She says the scholarship around virtual learning is still developing.

Hansen, for his part, praised course providers that have started taking steps in the right direction. EdX, an online-learning nonprofit co-founded by Harvard and MIT, announced in 2014 that it would offer free courses for high schoolers to help bridge the “readiness gap” facing them before college. And Khan Academy, a popular online-learning website, teamed up with the Boys & Girls Club of America last year to offer free SAT test-prep courses, with a focus on reaching underserved students.

“The people making these tools have the best of intentions in terms of serving a really broad demographic, and often times in terms of equalizing education,” says DiSalvo. “But they’re not accomplishing that yet.”