Every day, billions of people document their lives online. They post on social media, write product reviews, leave comments about their workplaces, share health concerns in patient portals and discuss everything from politics to parenting in online forums. Collectively, the conversations offer a glimpse into how people navigate everyday life.

Leeds School of Business Professor of Information Systems Kai Larsen believes researchers are only beginning to understand how useful those insights could be.

Kai Larsen

For generations, surveys have been one of the primary tools researchers use to study human behavior. But people today generate far more social media posts, reviews, comments and other forms of text than survey responses.

"We're estimating that it's about a million times more of these texts available for analysis than there are surveys we can ever do," Larsen said. "So we're drastically opening up the space for doing psychological research."

To help researchers make sense of that vast body of language, Larsen developed PsyProxy, a new class of AI models designed to translate written language into measurable psychological concepts. Rather than simply identifying whether a piece of text is positive or negative, the tool measures psychological concepts embedded in language.

Donald Edmondson, a behavioral medicine researcher at Columbia University who collaborates with Larsen, said PsyProxy's potential lies in its ability to turn people's own words into measurable psychological variables.

"People are constantly expressing psychologically meaningful information in their own words," Edmondson said. "Those data are rich in detail, and they often contain information researchers would never have thought to ask about directly."

Traditionally, understanding concepts such as trust, stress, loneliness, burnout or life satisfaction required designing surveys, recruiting participants and convincing people to participate.

Larsen believes PsyProxy can help researchers learn from language that would otherwise be difficult to analyze at scale. The system can be applied to a wide range of text, including customer reviews, employee feedback, patient messages, online discussions and other forms of written communication. Researchers could potentially use it to study topics ranging from consumer behavior and employee retention to public responses during major events such as economic downturns or public health crises.

For example, researchers could examine specific patterns. Are people expressing anxiety? Anger? Stress? Social isolation? Are certain concerns becoming more common over time? What real outcomes do they predict?

"Whether you have a dependent variable or not, you can see what people care about, what they talk about," Larsen said. 

Not a replacement for surveys

Despite the promise of language-based research, Larsen does not see PsyProxy as a replacement for traditional surveys.

"Surveys have a unique place in the world for getting answers to highly specific questions," he said.

Instead, he sees PsyProxy as another tool researchers can use. Surveys can help answer specific questions, while PsyProxy can help researchers find patterns in open-ended responses and other large collections of text. Researchers might even use it to analyze the "anything else you'd like to add?" responses that often get overlooked.

But Larsen said researchers still need to be careful about what language can and cannot reveal.

A customer who posts an angry review may never stop buying from a company, for example. Someone who expresses frustration online may never act on those feelings.

PsyProxy measures patterns associated with concepts such as frustration, trust and social connection, Larsen said, not what people ultimately do.

Language-based approaches also come with their own biases. Just as survey respondents may tailor their answers to what they think researchers want to hear, online conversations show only part of people's lives. Social media posts, reviews and comments can reveal valuable information, but they don't always reflect the full picture of someone's life.

“Anyone who has noticed that social media makes it look like everyone else is more successful than them is already dialed into some of these biases,” Larsen said.

For Larsen, the project is the culmination of an idea that first took shape nearly three decades ago.

In 1997, he began exploring ways to turn interviews, conversations and other forms of qualitative data into something researchers could measure systematically. The vision remained largely out of reach until advances in artificial intelligence finally made it possible.

"I waited 20-plus years for language models to get good enough to use for the problem I wanted to solve," he said.

Giving researchers the 'why'

Larsen also sees the approach as an alternative to relying entirely on large language models such as ChatGPT.

While those systems can summarize text and generate responses, he argues that researchers often need a way to measure concepts across large collections of text and compare findings across studies.

As AI systems have become increasingly powerful, they have often become harder to interpret, Larsen said. Researchers may know a model is making accurate predictions without understanding why.

Larsen said PsyProxy is designed to help open that black box by identifying the psychological concepts most strongly associated with a particular outcome.

"PsyProxy is giving the power of why back to researchers," he said.

Edmondson compares the system to a translator.

"People describe their lives in ordinary words," Edmondson said. "PsyProxy helps translate those words into psychological concepts researchers can measure, test and compare."

While Larsen sees potential applications across many fields, he is particularly interested in the impact the technology could have on health and well-being.

"PsyProxy is designed to help in all social and behavioral disciplines, from sociology and psychology to marketing, but behavioral medicine is near to my heart because of the difference it can make in our lives,” he said.

The research and tool are scheduled to be presented at the American Psychological Association annual convention in August.