The One-Person Team

For decades, solo-authored research papers have grown steadily less common, but around the time ChatGPT was released in late 2022, they started coming back. Something similar may be starting inside companies, where AI is allowing one person to take on work that once required a team.

Akira Matsui, a researcher at Kobe University, found the change in an analysis of more than 300 million scholarly works across 26 fields. It included established researchers who had always published with other people; some had never written a solo paper.

The increase in solo papers was much larger in fields where most of the work happens on a computer. Business, mathematics, psychology, decision sciences, and computer science saw the strongest increases. Chemistry, physics, and astronomy barely changed. This makes sense, given that a language model can write code, prepare data, run statistical analysis, review literature, and help draft a paper, but it can’t operate a telescope or spend the afternoon in a chemistry lab.

The new solo authors also stayed close to subjects they had already pursued with coauthors. Their papers leaned more heavily toward computational work and covered less ground. Among researchers publishing after 2022, the work of solo authors was 23 percent narrower in scope than the work of researchers who continued collaborating.

Matsui argues that AI may be taking over some of the work previously divided among coauthors. A researcher can carry a familiar piece of work alone, with the model helping with the analysis, coding, research, and writing.

The study can’t see who used AI, so the timing doesn’t necessarily establish that ChatGPT caused the increase. OpenAlex, the database behind the analysis, also changed some of its coverage and author matching during the same period. Matsui adjusted for field, seniority, citations, and productivity, and repeated the analysis within journals OpenAlex tracked continuously. The increase remained, although it was smaller.

The 23 percent difference is the part I keep thinking about. These researchers could complete work alone that would once have involved other people, and the work stayed close to subjects they already knew while covering less ground.

I’ve worked in large companies long enough to know how much time can disappear into coordination. A developer waits for an analyst, who waits for a subject matter expert tied up on something else. By the time everyone is available, the change has been sitting for three weeks and nobody quite remembers why it was urgent.

AI can shorten that process considerably by letting the developer carry an application from requirements through testing, or the analyst move from a question to a working model without waiting for several other people. That can save a lot of time.

Some handoffs carry context, though. That’s often where a developer learns why a requirement exists or an analyst hears that the customer already rejected the proposed approach. A colleague may remember that a similar project failed five years ago for reasons that never made it into the documentation.

AI can generate alternatives and challenge assumptions, and I use it that way all the time. It still starts with the problem and the context provided. That context can be extensive and still leave out the one thing a colleague would have mentioned ten minutes into a conversation.

The missing conversation doesn’t appear in a productivity report. Its absence may show up months later in a decision nobody questioned or an idea the team never encountered.

There’s also the problem of how people learn the work. Junior employees learn by helping more senior team members gather information, work through analysis, and prepare early drafts. The work might be tedious, but it puts them close enough to see how an experienced person changes direction or recognizes a mistake.

An analyst building the first version of a presentation hears how an executive reasons through the recommendation. A developer writing tests learns where the system tends to fail. Those assignments are also among the easiest to give to AI.

Every time one of those assignments goes to AI, the senior person saves time, but the junior person misses one of the key ways people become senior.

Companies can use the time AI saves for better coaching and harder assignments, and I hope they do. The immediate incentive runs the other way. The tool is fast, needs little explanation, and won’t add another meeting to the calendar.

Matsui’s study doesn’t tell us whether the new solo papers were better or worse, and solo papers remain a small share of published research. I wouldn’t use it to predict the end of teams.

I’ve spent a fair amount of my career trying to reduce the drag created by large teams, and I still think that’s good work. Matsui’s paper left me wondering what else disappears when the handoff does.

Algorithm and Blues publishes Sundays.

Sources

• Akira Matsui, “Return of the Solo Author: The Changing Division of Labor in Science in the Age of Generative AI”, July 2026.
• Lingfei Wu, Dashun Wang, and James A. Evans, “Large Teams Develop and Small Teams Disrupt Science and Technology”, Nature, 2019.
• Michael Andalón, Catherine de Fontenay, Donna K. Ginther, and Kwanghui Lim, “The Rise of Teamwork and Career Prospects in Academic Science”, Nature Biotechnology, 2024.