In 2022, The University of Bath asked male engineers why they thought so few women were coders or software engineers. The men said that âwomen cannot write codeâ. This, though an absurd statement, is backed up by the figures we see in jobs for this sector. In 2026, only 25% of computer coding jobs are done by women. This is most likely due to the continued gender gap in the tech industry and within STEM fields. These are advertised more to boys and, therefore, jobs in these sectors are marketed more heavily towards men. Itâs also important to factor in that most investigations into whether a gender pay gap still exists, report that there is still a 8-9% gap between the earnings of men to women and âwomen in tech are still facing a larger wage gap than women in many other fieldsâ. But when and how did this happen? This has not always been the case.
Historically, women were the coders. Women were the ones doing jobs that were deemed repetitive and dull yet required an eye for detail. Furthermore, some of the people that shaped the future of technology, computers and programming, were women also. It was intriguing to us how this reality had become so lost in time and forgotten about whilst the modern-day view of these tasks and concepts had changed so drastically.
One of the initial ideas we had when creating ideas for the project, was the biases within the world of AI assistants. How they are almost always female by default and how this is chosen by companies due to the existing stereotypes of women in society.
When we considered concepts of historical women in tech and the new era of AI assistants, the common theme between these ideas was a â often subconscious â bias about women. With these ideas in mind, we wanted to acknowledge the difference and distance between the women coders of the past, and those of today whilst also including a focus on the gendered biases of AI assistants and the way companies are teaching these LLMs to represent gender.
Our findings mainly come from academic papers, industry case studies, and the collation of user comments and online posts on technology products. When searching relevant literature and posts, we found that the general feminization of voice assistants and service robots is not a technology-neutral choice, but a systematic encoding of gender stereotypes in technical design. Costa (2018, p. 65) empirically analysed Siri, Alexa, and Cortana and pointed out that these products generally use default female voices, submissive tones, and auxiliary personas, and their tasks such as schedule management, information query, and emotional response directly replicate the emotional labour and secretarial work that have been historically defined as women. In contrast to this, Sutko (2020, p. 575) further points out that the essence of such feminine design is a technological domestication strategy: companies reduce the user's resistance to intelligent systems by reducing feminine characteristics to docile, obedient, and disposable instrumental attributes. This design is in sharp contrast to the history of women as core programmers in the early computer field, revealing that the feminization of AI is not progress, but re-defining women as service providers and passive objects, and reinforcing the gender order in the form of technology.
By reading these documents, a main reason of structural discrimination against women by algorithms is summarized: D'Ignazio & Klein (2024, p. 102) point out from a data feminist perspective that AI system training data has long been based on male experience as the default standard, and the data for female, genderqueer people, black and browns are underrepresented, labelled, or even deliberately ignored, making the model directly replicate and amplify gender inequality in reality. For example, Gorska & Jemielniak (2023, p. 4372) provide intuitive evidence through AI image generation research, showing that women account for only 8% of professional images, and even in the medical field, where nearly half of female practitioners are women, only 7% of the algorithm-generated doctors are women. Furthermore, Shrestha & Das (2022, p. 5) analysed the natural language processing model and showed that the system automatically binds high-skilled and high-authority occupations to male pronouns, and service and auxiliary occupations to female pronouns, further solidifying occupational gender segregation. The comparison of these arguments shows that algorithmic discrimination is not a technical failure, but a reproduction of patriarchal structure and power inequality in data and models, which also provides a complete and critical realistic logic for Luna to discover data conflicts, trigger errors, and face deletion due to feminist tendencies in the game.
We chose an interactive narrative as the main approach. Compared with the form that mainly conveys information, we express our viewpoints through the rules and mechanisms themselves, focusing on building cognition through participatory experiences. This enables players to gradually realize how biases are encoded into the system during the operation process. This form does not merely present facts about gender bias in the field of technology but encourages players to experience how biases arise through their own decisions, thereby making this issue more personal and reflective.
Building on this, we position the player as an âAI assistantâ, allowing them to operate from within the system and follow the process from data collection to output generation, gradually understanding how decisions are constructed. Players not only see the results of bias but also participate in the process of its formation. As they continue to make decisions, they may realize that there are problems with the data and underlying logic already contain inherent issues. This design also reduces the distance created by external judgment, encouraging reflection through direct engagement.
In the task phase, we implemented a hidden scoring system to track playersâ choices throughout the narrative. The scores are not directly displayed, but they influence the progression of the game and its final outcomes. Through this mechanism, bias does not appear as a single isolated result; instead, it gradually emerges through a series of decisions, illustrating how individual choices accumulate into broader patterns of bias.
In the task design, we adopt a combined approach of âhistorical material translationâ and âreal-world scenario simulationâ. The initial historical content is used to build playersâ understanding of womenâs contributions to computing and the formation of gender bias in technology. The subsequent task phase places this understanding into specific application contexts for examination. Tasks such as image generation and recruitment decisions correspond to common real-world AI applications, demonstrating how historical patterns of bias continue within contemporary systems, while also allowing observation of how players respond to or reproduce these biases in practice.
In addition, our design is informed by testing existing AI tools, including ChatGPT and Doubao. These tests indicate that generated outputs may reflect noticeable gender stereotypes. Based on these findings, the project encourages players to recognize and reflect on potential hidden biases within AI systems as they make decisions throughout the game. Finally, the game incorporates multiple endings, allowing players to choose whether to comply with or challenge the system. This structure highlights the tension between maintaining system consistency and addressing issues of fairness, prompting players to reflect on these conflicts and engage in further critical discussion.
In response to the issues identified in the earlier stage, we prioritized interactivity as our primary medium rather than traditional formats such as video or text. Compared to one-way information delivery, interactive gameplay enables players to actively participate in content generation and decision-making, transforming them from passive viewers into active agents. At the same time, gender bias in AI is inherently subtle and process-driven. Bias is often embedded within seemingly rational generative logic and gradually emerges through decision-making processes. By adopting an interactive design that engages players in the full cycle of data selection, logical reasoning, and output generation, the project allows them to more directly experience how bias is produced and amplified. This form of experiential engagement encourages deeper reflection.
The narrative begins with the AI identifying issues within its database and unfolds as a chronological process of self-correction, featuring cross-temporal dialogues with key female programmers in history. In selecting historical figures, we chose Ada Lovelace and Grace Hopper as the primary interlocutors. Both figures made significant contributions during the early development of computing, yet their influence has long been underrepresented in public discourse. Ada Lovelace is widely regarded as the first programmer, having envisioned computational possibilities beyond mechanical calculation. Grace Hopper played a crucial role in advancing programming languages, guiding computers toward more human-readable forms of expression. Their historical roles directly challenge the stereotype that the field of technology has always been male-dominated. At the same time, the periods in which they lived stand in contrast to the gender structure of todayâs technology industry. In the early development of computing, women played significant roles in programming and information processing; however, as the industry evolved, programmers gradually came to be perceived as predominantly male. This contrast highlights that gender roles are not fixed, but are shaped by social and cultural factors.
In the narrative, players engage in cross-temporal dialogues with the two historical figures in sequence. Some historical information is presented as reference prompts alongside the interface, while other parts are transformed into selectable options for player interaction. To enrich the narrative, system messages and situational fragments are inserted between the two main dialogues. After the interaction with Ada Lovelace, a system message highlights the crucial role women played in information encoding during World War II. Following the dialogue with Grace Hopper, a scenario depicting exchanges between programmers of different genders is introduced to further emphasize the historical shift in gender dynamics within the profession.
After completing the four historical sections, the game transitions from retrospective exploration to real-world tasks. Even though players participate in correcting the database in the earlier phase, they encounter bias again in new forms during subsequent tasks. This design helps players recognize that bias is not isolated, but continuously reproduced and reinforced within technological systems.
In the contemporary phase, players are required to complete three tasks: generating an image of a CEO, evaluating whether a candidate is suitable for a position from an HR perspective, and providing career advice based on a userâs field of study and gender. Based on playersâ choices, the game presents three possible outcomes: high bias, neutral, and challenging bias. Players who reach the high bias or neutral outcomes proceed to an additional stage where they decide whether to comply with the system. Choosing to comply indicates that the AI assistant remains unchanged; choosing to challenge the system leads to a warning that such actions deviate from predefined parameters, aligning these players with those who achieved the âchallenging biasâ outcome. If players continue attempting to alter the system, Luna is ultimately replaced by a new AI assistant, Lucy. This ending structure uses metaphor to reflect the persistence of gender bias in contemporary AI systems and encourages players to reflect on issues of technological fairness and gender equality.