How one intern’s research had real-world impact for Twitch moderators

Like physical communities, online gathering places can vary substantially in terms of how welcoming they feel. Amanda Cullen, a PhD candidate in informatics at the University of California, Irvine, is deeply interested in how to foster genuinely inclusive communities in virtual spaces.

“My dissertation, and my research for the past couple of years, has been about examining diversity and inclusion issues in games,” she said.

That interest brought Cullen to an internship at Twitch — a live streaming service that features gaming, esports, and other entertainment — which Amazon acquired in 2014. She worked with the Community Health team and science and analytics manager, Sanjay Kairam. She was curious about how her work might have an impact outside academia.

“I wanted to dip my toe in the waters and see what it was like to be a researcher on the industry side,” Cullen said. “I got paired up with Sanjay and was given this really incredible project thinking about Twitch moderators.”

From research to real world impact

Channel moderators weren’t an obvious focus for her at first. Cullen had come to Twitch with a dissertation looking at the experience of women on the platform as professional players and streamers, and how the challenges and opportunities of streaming might be different for them.

But Kairam believed that Twitch’s community moderators were the “special sauce that makes the whole service work”. And he also suspected that there were gaps in the company’s knowledge about the kinds of support that might benefit them.

As Cullen discussed the focus of her research internship with Kairam, she realized the work of community moderators in fostering welcoming online spaces had been a blind spot in her dissertation, and that — by better understanding their needs and how to support them — she could push forward the overall goals of her doctoral research.  She recognized the opportunity and modified her research project at Twitch to survey and assess moderators’ needs.

Cullen started with the basics: how moderators came to the role, and how streamers would find them (every Twitch channel has the opportunity to select its own moderators). Then she got into how they worked with the channel creators, how they helped develop norms and standards for behavior within their channels, and how they actively worked to guide the community around those ideas.

Cullen was also interested in what tools moderators wanted or needed, and how they responded to both positive and negative behavior within their channels.

“That led us to think about new resources that could be created to help moderators think about their roles and feel a greater sense of community — not just within their channel, but throughout Twitch as a whole,” says Cullen.

Cullen’s findings were put to direct use at Twitch: “Amanda's work shed a lot of really detailed light on moderator needs, which has propelled us to build better tools for our moderators,” Kairam said.

This support potentially had some very positive knock-on effects for moderators.

“A lot of what Amanda's work helped to unpack was that moderators are playing different roles within a channel. It really got the team thinking about how we could fine tune the experience for different types of roles that moderators play — so that they can work more effectively in each of those roles,” Kairam said.

Cullen has enjoyed observing the impact of her research. 

“After completing my internship, I watched things change and saw announcements about new resources or products happen and thought, ‘I wonder how much I influenced that?’ I'm looking forward to seeing that continue,” she said.

Twitch as a science internship destination

While Twitch might not seem like an obvious destination for science internships, the service actually offers a variety of potential science and engineering intern opportunities.

For example, Cullen’s research proved important in identifying and understanding moderator needs that could be addressed through machine learning (ML).

“We have an ML-powered tool for moderators called AutoMod,” Kairam noted, “that automatically flags messages which may be toxic and holds them for channel moderators to review. Community Health also has an ML group called Proactive Detection, which designs models to identify and prevent various types of bad behavior across Twitch.”

Kairam also noted Twitch is actively recruiting science interns across a variety of subjects.

“A lot of times, when students think about interning at tech companies, they're thinking specifically about technical projects, technical roles,” Kairam said.

There are, of course, many technical opportunities for those interested in pushing the state of the art in machine learning: by generating recommendations for live content, for example, or building natural language processing models to detect harmful chat messages. But Kairam said there’s plenty of room for science interns in a variety of disciplines, including on the Community Health team, which is focused on creator and viewer safety and support.

“There's a lot of value to bringing into product discussions social science researchers, who are able to think through a complex problem and develop a new way of thinking about it,” he said.

Looking at problems from new points of view is powerful and enables Kairam’s team to “introduce a new mental model or framework that sort of shifts the perspective and allows us to act on information that we may not have been able to act on before,” he added.

Community Health research opportunities

There’s plenty more to research within Community Health as well, including how to proactively detect bad behavior in a live context or understanding chat behavior on the giant scale of Twitch chat.

“Addressing these challenges requires a breadth of perspectives, from understanding how healthy communities form and function, to building quantitative models of chat and channel safety, to developing machine learning models that can proactively detect and remove harmful content,” Kairam said.

“The nature and scale of Twitch, in itself, makes a lot of questions and a lot of different research areas super cutting-edge,” he observed. The fact that real-world application of that research can directly impact the safety (and fun) of millions of people around the planet is valuable to academic researchers like Cullen, who want their work to help others.

Would Cullen recommend the internship to other PhDs?

“Definitely go for it,” she said. “The internship gave me more confidence in my ability to do scientific research in an industry context.”

And there’s plenty of support, no matter how long the pandemic continues to impact the office work environment. While Cullen’s internship was initially supposed to be at Twitch headquarters in San Francisco, she remained at UC Irvine. And though Cullen wasn’t meeting people face-to-face, Kairam made sure to facilitate connections with other parts of the team.

“He recognized that if I was interested in industry employment, I would want to make friends and meet people in all kinds of different areas of the company,” she said.

That helped make the experience incredibly meaningful, Cullen said. ““It ended up being a really great and valuable time for me, even from the first day.”

Doctoral candidates and others interested in pursuing an internship opportunity at Twitch, can find opportunities, and apply here.