With the current proliferation of educational games, MOOCs, and with the pervasive use of virtual identities such as avatars in systems ranging from online forums to virtual reality simulations, it is increasingly important to understand the impacts of avatars. Over two years, I led an initiative in MIT’s Imagination, Computation, and Expression (ICE) Laboratory conducting experiments involving > 10,000 participants to understand the impacts of virtual identities on users in virtual environments. Using a computer science learning platform and game of our own creation as an experimental setting, we have been studying the impacts of avatar use on users’ performance and engagement in computer science learning environments. This is a topic of increasing importance in human-computer interaction [69, 130, 132, 310, 452, 549]. While a great deal of work focuses on procedural thinking and problem solving, we argue that attending to learners’ identities and their engagement to be equally important. We systematically explored the impacts of different avatar types on users, beginning with distinctions between anthropomorphic vs. non-anthropomorphic avatars, user likeness vs. non-likeness avatars, and other conditions informed by insights from the learning sciences and sociology. Our studies have revealed that avatars can support, or harm, performance and engagement. Several notable trends are: 1) simple abstract avatars (such as geometric shapes) are especially effective when the player is experiencing failure, e.g., while debugging, 2) likeness avatars (avatars in a user’s likeness) are not always effective, 3) role model avatars (in particular scientist avatars) are often effective, and 4) successful likeness avatars that are a user’s likeness when doing well and otherwise abstract are effective. We describe our studies leading to these findings and end with a follow-up study.