One Anonymous Upvote on Stack Overflow Boosts Contributions for Weeks
A pre-registered experiment on 22,856 Stack Overflow users shows one anonymous upvote boosts question-asking by 6.3% and answering by 12.9%.
ProTakeaways
- Pre-registered Stack Overflow experiment with 22,856 users randomly assigned zero, one, or two anonymous upvotes on new questions.
- One upvote raised subsequent question-asking by 6.3% and answering by 12.9% within four weeks; a second upvote added nothing.
- The answering effect persisted at 12 weeks; the asking effect faded, suggesting different mechanisms drive the two outcomes.
- Roughly 80% of the asking effect comes from the direct social signal, while roughly 65% of the answering effect flows through algorithmic amplification.
- Mere views without an answer did not predict future participation; receiving substantive human help did.
- Anonymized data and replication code are available on Zenodo; full paper on arXiv.
A single anonymous upvote, delivered within hours of posting a question, measurably changes how people behave on Stack Overflow for weeks afterward. That is the headline finding from a large pre-registered field experiment that quietly ran on the platform for four months in 2022, before ChatGPT reshaped the economics of programming Q&A. The study offers something rare in the debate about declining online participation: clean causal evidence about what small social signals actually do.
The stakes reach beyond academia. Prior work by one of the same authors found a 16% relative decrease in posting activity on Stack Overflow, with a larger effect in later months after ChatGPT launched, and the Stack Exchange network reportedly contributes roughly 3.4 times as much text by weight to The Pile pre-training corpus as Wikipedia does. When contributors stop showing up, both the human community and the models trained on its output lose their supply of fresh knowledge.
Sneaking upvotes into the live feed
The team ran a randomized controlled trial on live Stack Overflow traffic. Four times a day, an automated script scanned the questions feed and randomly assigned newly posted questions to one of three arms: no upvote, one anonymous upvote, or two upvotes staggered by two hours to mimic organic feedback. The first vote arrived within six hours of posting. Users could not see who upvoted them, and after the experiment ended the voting accounts were deleted, removing all experimental votes from the site.
A few design choices matter for interpreting the results:
- The study targeted fewer than 2% of new questions per day to avoid distorting the platform.
- Each of the 22,856 observations maps to a distinct user, so treatment effects are user-level rather than post-level.