Contemporary game production aims to change gender and racial stereotypes, but these efforts often yield minimal results. This study employs the “warmth-competence” stereotype model and computational text analysis to conduct a “encoding-decoding” analysis of character stereotypes in the game “League of Legends.” The findings reveal that at the encoding level, game creators strive to break the strong association between character types and gender/race. However, when Chinese players perceive characters, they still project gender and racial stereotypes onto them, although some stereotypes dissipate as character types diversify. This research constructs a Chinese stereotype content dictionary and conducts a computational analysis of the bullet comments from videos interpreting “League of Legends” characters by Chinese audiences. The results show that audiences rate male characters lower in warmth but higher in competence than female characters. Women portrayed in traditional roles, such as heroes, are perceived as having lower abilities, whereas those in non-traditional roles, such as anti-heroes, are rated similarly to their male counterparts. This suggests that non-traditional roles can help dismantle gender stereotypes. Additionally, Chinese audiences rate characters representing Asian and White individuals higher in competence than those representing other people of color. However, the interaction effect of race and character type on stereotype perception is not significant. This indicates that Chinese game audiences have a persistent view of racial stereotypes, and the strategy of diversifying character types has little impact on changing these racial stereotypes. This study illustrates the need for game encoding to shift away from traditional stereotypes, promoting diverse perspectives among audiences through multimodal approaches.