“Ghost work” is anthropologist Mary L. Gray’s term for the invisible labor that powers our technology platforms. When Gray, a senior researcher at Microsoft Research, first arrived at the company, she learned that building artificial intelligence requires people to manage and clean up data to feed to the training algorithms. “I basically started asking the engineers and computer scientists around me, ‘Who are the people you pay to do this task work of labeling images and classification tasks and cleaning up databases?’” says Gray. Some people said they didn’t know. Others said they didn’t want to know and were concerned that if they looked too closely they might find unsavory working conditions.

So Gray decided to find out for herself. Who are the people, often invisible, who pick up the tasks necessary for these platforms to run? Why do they do this work, and why do they leave? What are their working conditions?

Gray ended up collaborating with fellow MSR senior researcher Siddharth Suri to write Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Houghton Mifflin Harcourt).

The Verge spoke to Gray about her research findings and what they mean for the future of employment.

This interview has been lightly edited for clarity.

Labeling data to feed to algorithms is one obvious example of ghost work. Content moderation is another. What are other examples?

Filling out surveys, captioning and translation work, any sort of transcription service. Doing web research, verifying location addresses, beta testing, user testing for user designs. Anything you can think of as knowledge work, like content creation, writing editorial, doing design. You name it. The list is endless. All of those are tasks that can be distributed online. It’s all of the things we’re used to seeing in the office, and this is what it looks like to dismantle that as a full-time job and turn it into projects for myriad people.

Am I right in thinking that basically every tech company relies or has relied on ghost work?

I would be hard-pressed to find any business that sells itself as AI that either didn’t deeply rely on ghost work to generate their basic product or isn’t very much reliant on it today. There are so many startups and businesses out there, anything that calls itself “business insights” or “intelligence and analytics.” That’s using crowdsourcing or collective intelligence, and that’s relying on ghost work. There is no way around the need for people to sift through the piles of what’s called unstructured data.

Sometimes, people think that as technology gets better, we won’t need this type of ghost work anymore. But you write that “the great paradox of AI is that the desire to eliminate human work generates new tasks for humans.” So clearly you don’t subscribe to that belief. Why not?

What might change are the specific tasks. Believing that AI will never need humans labeling data means believing that language will never change, style will never change. Service industries, especially, are so difficult to fully automate because being able to listen to somebody’s voice and register that person’s silent anger is such a human capacity. So there are cases when AI will, I argue, always fall short.

Engineers are always wonderfully optimistic about opportunities. As an anthropologist, I know how complicated it is to think cross-culturally about these questions. Even if we fairly reliably get to 100 percent of spoken English with a flat Midwestern accent, what about when you move into vernacular and slang and folks who will splice together languages and code switch? Anytime you see an autotranslation of a talk, you see the places where language breaks down, often around somebody’s name.

Those are the kinds of computational problems that are intractably hard for AI to capture because there’s not enough data consistently available to model what’s going to be the next utterance that somebody is saying using Spanglish. We’ve already effectively automated all of the easy things.

One interesting thing you mention is that we don’t have good labor statistics for how many people are doing ghost work. Why is that?

The biggest challenge is that the ways we count jobs are often in relationship to professional identities, or really clearly defined capacities or skills, and no one is oriented to a world of work that is project-based. We don’t have the language to describe an image tagger or a captionist. One of the findings in our research is that people have really different mental models. They may or may not identify as self-employed. They may or may not identify as a journalist if they write for a content farm, and that might change whether they decide to answer a survey question to help us measure this workforce. Let alone the fact that ghost work is distributed around the globe, and there is no global bureau of labor statistics.

A key question in this book is: who are the people doing ghost work? So, who are they? It sounds like they could be almost anyone.

When we got our initial set of surveys back from the four different platforms we studied, it was clear that there were as many women as men, though they worked different hours. People had college educations, but that wasn’t surprising because that maps on to knowledge work and information services broadly.

They are all of us. These are the folks who, for reasons of social capital, don’t have access to a network that was going to boost them into the full-time job. That’s the pattern I see sociologically or anthropologically. They’re first-generation college-going more often than not. This is a group of people who don’t have strong social ties to elites.

What are people’s motivations for this work?

There’s not one type of person doing this work and not one single motivation. There is a core group of people who are turning to this work, often because of other constraints on their time. People would say that they don’t have time to commute and were going to be commuting for a comparably paid job at least two hours, and that was going to cut into the amount of money they could make. That’s the calculus they’re making here. So they’re deciding to turn to this work and effectively. Once they’ve figured out how to make enough money on enough platforms, they cobble together the equivalent of a full-time salary for them to meet their needs. We call those folks “always on,” and they’re turning this into full-time work by the number of income streams. But this group of people is a small percentage — 10 to 15 percent, depending on the platform. This is what the research tells us about all these platforms. The core group of people is doing the bulk of the work.

Then there are the “regulars,” a deep trench of people who can step in at any time. The regulars are the ones that enable the “always on” people because if the “always on” steps out, there are enough people in that pile of regulars who are going to be able to step in. They’re often caregivers, and they had other motivations; they were pursuing another passion project or they were going back to education and taking courses, and this gave them a means to be able to finance that.

Lastly, there is the long tail of experimentalists, which is the name we gave the people who try one or two projects, figure out that this is not for them, and leave. The most important part of doing anthropological work is we could meet the people who left and figure out why. And it had to do with never hooking into a community of peers to help lower their costs, feeling like they don’t have enough support, and that this was too difficult to figure out. And it was exhausting cognitively.

A feature of this kind of market is that anyone can work for anyone else. What happens in that kind of environment?

For anybody who becomes a regular or “always on,” they’re invested and bring the same framework they have to any job. It’s an amazing amount of self-policing because workers are invested in making sure that work comes back to the pool. They want to make sure their peers are doing well because, if not, that could work against their interests in getting the next job.

Businesses should be equally invested in this accountability in the supply chain. If they’re relying on lowering their costs of investing and what they need most is somebody who’s ready and willing and able to jump in for a project, the exchange is to create some mechanism that ensures that anybody who is entering is refreshed and has the opportunity to keep up. Otherwise, it’s not sustainable as a labor market.

But companies aren’t doing that. They’re not creating the accountability or trust or culture that would help the ghost workers.

If you talk to any of these companies, most of them believe that we’re going to get this automated and think, “I just need these people for a little while.” That’s precisely our problem and that’s historically been our problem since the Industrial Age: treating badly people who do the contingent work that can’t quite be automated. We stop paying attention to these people and their work conditions, we start treating them as something that can be replaced eventually, and we don’t value the fact that they’re doing something that a mechanical process or computational process can’t do.

I hate the parallel to horsepower. This is not like replacing horses with automobiles. People are not performing a mechanical task. They’re extending something distinct about humans — their creativity and their interpretation.

What should we do to address this? What are the policy suggestions?

At the very least, it means valuing everybody’s contribution. The first step is being able to identify the people who have contributed. In Bangladesh, it made a huge difference in textiles when companies selling products had to tell us who was involved in making the shirt on my back. There should be a clear record thanking anybody who contributes labor to an output or service. The consumer should always be able to trace back the supply chain of people who have had a hand in helping them achieve their goals.

This is about regulating a form of employment that does not fit in full-time employment or fully in part-time employment or even clearly in self-employment. I believe that this is the moment to say the classification of employment no longer functions. Anybody who’s working age should have a baseline of provisions that are supplied by companies.

If companies want to happily use contract work because they need to constantly churn through new ideas and new aptitudes, the only way to make that a good thing for both sides of that enterprise is for people to be able to jump into that pool. And people do that when they have health care and other provisions. This is the business case for universal health care, for universal education as a public good. It’s going to benefit all enterprise.

I want to get across to people that, in a lot of ways, we’re describing work conditions. We’re not describing a particular type of work. We’re describing today’s conditions for project-based task-driven work. This can happen to everybody’s jobs, and I hate that that might be the motivation because we should have cared all along, as this has been happening to plenty of people. For me, the message of this book is: let’s make this not just manageable, but sustainable and enjoyable. Stop making our lives wrap around work, and start making work serve our lives.

There is no lack of parenting advice available, and the sea of tips from well-meaning relatives and internet strangers alike can leave parents looking to scientific evidence for guidance.

The problem is that not all evidence is equal. There are plenty of studies out there, but a tiny study funded by a formula company, for example, should not be taken as seriously as a large randomized trial conducted by a government agency. “It is not super straightforward to evaluate all of this literature,” says Emily Oster, a health economist at Brown University and the author of Cribsheet: A Data-Driven Guide to Better, More Relaxed Parenting from Birth to Preschool, out now from Penguin Press. (Oster’s previous book, Expecting Better, was a data-driven guide to pregnancy.) To understand what the data really suggests when it comes to sleep training, toddler discipline, or language development, “it goes beyond just reading the papers,” Oster says. It requires evaluating which papers are good, which types of evidence are strong, and drawing conclusions — which are all skills that Oster uses as an academic economist.

The Verge spoke to Oster about the gaps in parenting literature, how to evaluate research, and the best advice she’s ever received.

This interview has been lightly edited for clarity.

How big is the academic literature on birth and child-raising relative to other areas of health? My sense has long been that topics related to women’s health are understudied. Was that the case for literature on parenting?

I think it’s hard to tell. There certainly is a lot of literature on pregnancy and moms and kids, and there’s also a lot of literature on a lot of other things in science. What I will certainly say is that there are ways in which the literature on this is frustrating, and there are some very important questions about which we don’t seem to know as much as I would hope.

One place that clearly comes up is screen time. A lot of people would like to understand the impact of screen time on little kids, particularly on the kind of questions that modern parents ask, like “Is it okay for my kid to play iPad games?” That’s something where we simply don’t have the kind of evidence we would need, partly because people differ in their choices about what to let their kids do on screen, but also because the modern version of screens is very different from the version even 15 years ago. So even if you wanted to ask, “What is the impact of this kind of exposure on kids who are graduating from high school now?” we aren’t yet ready to answer those questions. It’s an inevitable limitation, but it’s frustrating for those of us who are trying to make these choices.

On the other hand, what are some areas in which scientific literature has a lot of strong, good evidence? Vaccination seems to be one. What are some others?

Vaccination is one area. Although the literature is very new, the literature on allergies and food is very good. There aren’t a million studies, but the stuff that we have is very compelling and high-quality. A third place I would highlight is the sleep training stuff. Relative to a lot of these other hot-button first-year topics, the evidence there is more susceptible to randomized controlled trials, which are the gold standard of evidence in studies.

Can you tell me more about the types of data that are considered “high-quality” versus the studies that might not be as strong?

The baseline approach that most of the literature takes is to compare outcomes. Let’s take breastfeeding as a good example. You can compare IQ for kids who are breastfed and kids who are not breastfed and try to draw some conclusions. The issue is that, in a lot of settings, the kind of people who are breastfeeding are different from the kind of people who are not, and those differences may well be in part, or entirely, what is responsible for the difference in kids’ outcomes, not the breastfeeding itself. So when we look at these kinds of studies, we want to look for places where we avoid that problem as much as possible.

One way to do that is to try to really carefully control for differences across families. That’s pretty hard to do, though, in some cases, you see siblings who were treated differently, and that tends to be better because they’ll, for example, have the same mom so you’ll at least hold constant the mom. The other thing we can do is randomized trials where instead of just comparing kids who are treated differently based on the choices parents make, you can encourage families to make one choice or another and later compare the kids. Because you’re choosing randomly what you encourage, the differences are more plausibly associated with the treatment and not differences across families.

Since you mentioned breastfeeding, which is a famously hot-button topic, can you tell me a little about what your analysis of the data showed us about its benefits?

There’s a lot of emphasis on all of the many benefits of breastfeeding, and these fall into multiple categories: benefits to the mom, benefits to the baby in the short-term, benefits to the baby in the long-term. A lot of the evidence is biased or problematic because, as we said, the choice to breastfeed is not random. So we have to look at better kinds of evidence. There’s one very large randomized trial and a bunch of studies that compare siblings and do a better job adjusting, and when you look at those pieces of evidence, what comes out is that there are some real benefits for the baby early on — there may be some improvements in digestion, lower allergy risk, a reduction in ear infections — and some long-term benefits to the mom in terms of breast cancer reduction. But some of those claims about later health benefits for kids — like higher IQ, lower obesity, less asthma — don’t seem to be borne out in the best data. So the picture ends up a bit more mixed. It says breastfeeding is great, and we should help people do that if they want to, but we may want to dial back some of the more grandiose claims.

One thing I liked about the book is how you not only go over the data, but you also talk about economic concepts, like how to think marginally in the case of choosing child care and comparing cost. Can you unpack that a bit?

Thinking marginally is a core component of economics. Optimizing — doing the best thing — requires thinking about the opportunity cost and things on the margin. So in this example, when you’re trying to decide what the right thing is to do with child care and evaluating options that cost different amounts, it’s easy to think just about money as money. Just dollars.

But I would argue that, in fact, you need to conceptualize what you will buy with the money, and what is the value of that? And that’s where thinking about the marginal, or next-best use of the money is really important. Sometimes confronting the next-best use of the money can cause you to make a different choice. You realize that the next-best use would be a new car, but I don’t care that much about having a new car this year, and I actually would rather have that child care choice.

I also liked how you added nuance to some stories that we’ve heard, like how marital satisfaction declines after having children. That does seem to be true, but it doesn’t decline identically for all couples, right?

So it’s true that, on average, marital happiness declines after you have kids, but those declines are smaller if the kid is planned and if you’re happier beforehand. It’s not uniform that everybody hates their spouse after they have kids, although I think it is good to recognize that the first year of your kid’s life is likely to be a time when most marriages do feel some stress. That is something you may have to work through, and a lot of people, of course, do come out on the other side.

You write that, as kids get older, it gets more and more difficult to find studies because the data gets way more complicated. How so?

In general, even very good studies are typically only able to figure out an average effect. We’re able to say, on average, did the kid sleep better after we did this kind of sleep training? For little kids, that’s often a well-posed question, but as you get to older kids, there’s so much difference across kids in how they’re going to react to things.

A question a lot of people ask me is something like, is a Montessori school a good idea? On average, we don’t have that much data, and it could easily be the case that a Montessori school is really good for some kids and really not good for other kids. Even a very good study of Montessori could mask some really good effects for some kids and really bad effects for some and our data is even less well-suited to pick up that kind of heterogeneity. That problem only gets worse as kids get older and parents are choosing things that they think will work for their kid. There’s less data, and the reason is partly that it’s so hard to imagine doing these studies well.

The book ends with you talking about the most helpful parenting advice you’ve received. You were taking your daughter to France and were really worried about her getting stung by a bee and being allergic and something bad happening. And your doctor was like, “I’d probably just try not to think about that.” I liked that a lot. Why do you consider that to be the best advice after diving into so much data and papers?

There’s a sense in which, for me, that is very important to hear. I think about that advice all the time because it’s pretty broadly applicable to a lot of things in parenting. We can get caught up in every tiny decision and miss the enjoyment of parenting and the part of this that’s supposed to be fun. It just pushed against some of my worse instincts as a parent to just obsess over everything. Sometimes you just have to accept that you cannot control everything. That’s hard, but it’s part of the fun. Also, the kid was eventually stung by a bee, and it was totally fine.