At Microsoft, a wider circle of colleagues predicted who first used AI coding tools better than immediate teammates did — influence, or just similar people clustering?

This explores whether new tools spread because similar people cluster together and adopt for the same reasons (homophily), or because one person's use actually changes what their peers do (social influence). The corpus doesn't settle this cleanly for tool adoption. It does have one strong case study and a sharp idea borrowed from recommender research that helps separate the two.

The case study is Microsoft's rollout of Claude Code and Copilot CLI Do social networks drive adoption of new coding tools?. Engineers' social ties predicted first use better than career stage or tenure. The most predictive ties were the broader skip-level connections, not just immediate teammates. That detail matters. Your closest teammates are most likely to resemble you in role, codebase and habits, so if similarity were driving adoption, that is where the signal should be strongest. A stronger signal from wider, less similar ties points toward exposure and influence: hearing about a tool from someone outside your bubble. The study also found that retention tracked what engineers did with the tool, not who they were. Influence may get people to try a tool, but whether they keep using it depends on fit. Still, network studies like this rarely prove causation. People with wide networks may simply be the curious type.

Recommender research makes the same split, from a different direction Can friends with different tastes improve recommendations?. Methods that assume friends share tastes, and pull their profiles together, did worse than a model that uses friends' *different* tastes. Networks added the most value when they nudged people toward choices outside their usual pattern. Applied to tools: homophily explains why your cluster already uses what it uses, while influence explains how something unfamiliar gets in. A related finding shows that the channel through which people encounter something shapes who shows up and how they judge it Do different recommender types shape opinion convergence differently?.

Two other notes complicate the picture. Indian writers accepted more AI suggestions than American writers, and the authors read this as a cultural pattern of trust and collectivist technology adoption, not noise to control away Is higher AI use by Indian writers a confound to control?. A group-level norm like that is a third force: neither 'people like me' nor 'my friend showed me', but shared expectations about how much to rely on a tool. Partner-choice experiments show a fourth: people who started out biased against AI partners came to prefer them after repeated interactions showed the bots were reliable Do humans learn to prefer AI partners over time?. Direct experience can overturn whatever the network or culture first suggested.

Here's the takeaway you might not have expected. Homophily and influence probably act at different stages. Influence, especially through looser and wider ties, drives first contact. Fit, experience and cultural norms decide who stays. If you want to know whether a tool will really spread, look at who tried it, not just how many did, and at whether people kept using it after trying it.