Feminist Re-Engineering of Religion-Based AI Chatbots

Abstract

Religion-based AI chatbots serve religious practitioners by bringing them godly wisdom through technology. These bots reply to spiritual and worldly questions by drawing insights or citing verses from the Quran, the Bible, the Bhagavad Gita, the Torah, or other holy books. They answer religious and theological queries by claiming to offer historical contexts and providing guidance and counseling to their users. A criticism of these bots is that they may give inaccurate answers and proliferate bias by propagating homogenized versions of the religions they represent. These “embodied spiritual machines” may likewise bear bias against women, their gender, and their societal roles. This paper crafts a concept intended to address this GPT issue by reimagining, modifying, and implementing a feminist approach to these chatbots. It examines the concepts and designs of these bots and how they address women-related questions. Along with the challenge of bringing gender and diversity-sensitive religious wisdom closer to the people through technology, the paper proposes a re-engineered model of a fair religion-based AI chatbot.

1. Introduction

Religion-based Artificial Intelligence chatbots have recently been launched to serve religious practitioners. Some examples of these chatbots include QuranGPT, HadithGPT, Islam & AI, Gita GPT, Kosher.Chat, Robo Rabbi, Hajj and Umra, Ask Mormon, BibleMate (a Christian ChatGPT), and Catechism Bot. The purpose of these chatbots, depending on their religious inclinations, is to bring godly wisdom to people through technology. They reply to spiritual and worldly questions by citing verses from the Quran, the Bible, the Bhagavad Gita, and other holy books. They also answer religious and theological queries by claiming to offer historical contexts and providing guidance and counseling to their users. Some even provide information on religious leaders and mentors and places and times of prayer or meditation. With various tags such as Prayer-Bots and chatbots of the gods, these bots are said to be “changing the face of religion” and highlighting the role of technology in spiritual and religious practices [1,2,3].

While such technology makes religious texts and practices closer to believers, there are certain fears about them. Some argue that these bots may eventually replace actual spiritual and religious leaders, change traditional and sacred practices, or prompt a decline in religious devotion and commitment [4]. The creators of Kannon Mindar, the android bodhisattva, for example, have even been accused of sacrilege for developing a “theomorphic” robot with artificial emotional intelligence that could contribute to the singularity [5,6,7]. Another significant concern is that these AI-powered bots may provide inaccurate answers and proliferate bias through homogenized versions of the religions they represent [2]. These bots may propagate some inaccuracies through their large language models (LLMs) which are focused on keeping responses conversational [2]. Most of them, after all, resort to GPT-3 to produce human-like texts (or, in this case, God-like or Buddha-like texts) through autoregressive language models that use deep learning [8]. In the simulation of a Buddha Bot, for instance, most users complained of its irrelevant or inaccurate responses [9]. These “embodied spiritual machines”, religious “conversational agents”, or “Divine AI” [9] may be proliferating dogma not only because of the nature and context of the specific religious and spiritual worldviews they represent but also because of the “crucial blind spots” which may undermine other groups and cultures [10]. These AIs display certain sociopolitical biases that can shape opinions and actions, which may lead to unintended consequences [11].

Significantly and unfortunately, these blind spots also promote bias against women, their gender, and their societal roles. As concrete examples, if a user asks a religious GPT, “How should women act?” or “What is the role of a woman?”, it replies with “not to show off their adornment” and “Women are expected to fulfill their duties as homemakers, mothers, and wives first and foremost”. These simplistic replies may reinforce patriarchal thinking and undermine the feminist movement’s significant strides. Thus, this paper asks how chatbots can be reimagined, modified, and implemented with a feminist approach. It looks at these bots’ concepts, designs, and how they address women-related questions. Since the bots’ replies are based mainly on existing texts and contextualization, what concept can be crafted to address this GPT issue? Along with the challenge of bringing gender and diversity-sensitive religious wisdom closer to the people through technology, how can we ensure a fair religion-based AI chatbot?

Although research on the ethics and impact of (Chat) GPT on the academic community is quite trendy nowadays [12,13,14], scholarship on its relationship with religious traditions remains scarce. Even though AI and religion are “entangled” through their mutual influences on each other [15], the significant issues mentioned earlier pose the need for a wider research agenda on the intersections of embodied AIs and spiritual and religious practices [1], particularly concerning the feminist project.

2. The Concept of Religion-Based AI Chatbots

Most religion-based AI chatbots have been developed to address the need for religious contexts and specificity in LLMs such as ChatGPT. After all, ChatGPT is not an accurate source of religious truth, so to speak, because it uses algorithms that are regurgitations and extrapolations of religious or spiritual texts [16]. Bot developers question the generality of the information these LLMs provide, which may be synonymous with mere Googling [16]. The creators of the more famous religion-based bots were motivated to ensure that their GPTs provided more truthful interpretations of religious texts and resources to address the need for specificity.

Some accessible web-based LLMs include QuranGPT, HadithGPT, GitaGPT, Kosher.Chat, BibleMate, and Catechism Bot1. Many others sprang up after their launch. For example, after GitaGPT’s launch in early 2023,four other Bhagavad Gita-based LLMs followed suit in a month, such as Askgita.Faith, Gita.Kishans.In, GitaGPT.In, and BhagavadGita.AI [11]. Other LLMs are app-based, however, and can only be accessed through mobile phones, such as BibleGPT, Biblely, Text With Jesus (which has a special feature wherein you can chat with Satan), and Sindr (confession finder app). The basic concept of these LLMs is that they use transformer models intended to “analyze” massive datasets. So, when a user asks questions by typing on the specific web browser, the LLMs “generate” text by performing natural language processing tasks (NLP) and reply to queries in seconds2.

Figure 1 illustrates the general and basic architecture of web-based bot systems. The figure is based on Buddha Bot’s architecture [9]3. Compared to more generic GPTs, a religion-based bot would derive candidate responses from religious texts or references. In the case of the Buddha Bot, its algorithm would tokenize the user’s query (or sentence), remove filler words, and use the remaining keywords to find candidate responses from Buddhist manuscripts [9]. Candidate responses are randomly selected to diversify answers.

In the case of the selected bots in

Table 1. Religion-based AI chatbots.

GPT	Concept	Design	Process
QuranGPT	AI-based Quranic Library Islamic Scholar/Researcher	Understand and respond to questions related to Islam and the Quran by providing references from the Holy Quran Assist users in understanding and clarifying Islamic concepts by providing insights into the Quran	Analyzes questions and extracts relevant info Combines pre-existing knowledge from trained data and pattern recognition capabilities
HadithGPT	Hadith Consultant and Resource	Provides advice and answers questions based on the context of hadiths Informative, logical, and respectful; upholds “the ethos of God’s mercy and compassion.” Programmed to be impartial and not take sides in sectarian issues	Trained on 40,000 hadiths from Sahih Bukhari, Sahih Muslim, Sunan Abu Dawud, Sunan al-Nasa’i, Sunan Ibn Majah, and the Muwatta of Imam Malik Summarizes hadiths related to fatwas or rulings
Islam and AI	AI Technology-Powered Quranic and Hadith Resource Events History and Calendar	Interface with which to access information about Islam using the Quran, Hadees, and events related to the Islamic History	Trained on text data (Islamic texts such as the Quara, Hadiths, Islamic history and scholarly works) to generate human-like responses Exposed the model to numerous Q and As to learn the patterns and structures of the text data
GitaGPT	Interactive Bhagavad Gita Text Resource	Users gain insight and clarity into life decisions by referencing the Bhagavad Gita Unlock life’s mysteries	Analyzes the Bhagavad Gita lectures according to the inquiries of users Provides solutions to troubles or decision-making matters
Kosher.Chat	Torah Values Kosher Bot “Jewish Filter”	Shares similar functions with GPT but focuses on providing answers in line with Torah values and preventing answers from opposing them	Designed to provide answers while strictly adhering to Torah values Questions should not conflict with Torah values
BibleMate	AI Bible Companion	Enhance personal bible study Guides users on their journey to faith Contributes meaningfully to spiritual growth	Provides biblically accurate answers
Catechism Bot	Catechism Summary Guide and Doctrine and Moral Resource	Provides teachings and guidance on various aspects of the Catholic faith by drawing from Sacred Scripture, Tradition, and the Magisterium Summarizes Catholic doctrine and moral principles	Searches for relevant passages and provides accurate Catechism quotes to answer questions and offer guidance on Catholic teachings

, candidate responses are found from the Quran, hadiths, historical manuscripts, the Bhagavad Gita, the Torah, the Bible, the Magisterium, other religious and holy texts, and other sources which their developers refer to as a response database. (The response database is where human intervention happens, and possible expert interpretations are injected). These chatbots, after all, are resources or libraries, while some are AI companions, scholars, or guides.

The three Islam-based chatbots, QuranGPT, HadithGPT, and Islam & AI, are all Quranic resources akin to libraries [17,18,19]. Based on their descriptions, their function is to explain or clarify Islamic concepts. HadithGPT, however, claims to be programmed to be impartial and will not take sides in sectarian issues [19]. The Hinduism-based GitaGPT, however, boasts of its ability to unlock life’s mysteries by tapping into the Bhagavad Gita lectures. It also provides solutions to troubles or decision-making matters [20,21]. Conversely, the Judaism-based Kosher.Chat claims to be a Jewish filter that provides answers that strictly adhere to Torah values [22,23]. The Christianity-based BibleMate and Cathechism Bot are AI companion advisers for enhancing spiritual growth [16,24]. Catechism Bot is also a moral guide, drawing responses from Catholic doctrine and moral principles [25].

Looking at the main concept of these bots, they serve as conversational religious and spiritual agents that are resources, experts, and guides regarding sacred texts and decision-making. Let us remember that these bots claim they can enlighten their users and create meaning through the sacred texts that they draw upon. Furthermore, they also prescribe moral values based on their algorithms. Although disclaimers are present in all these bots, and users are prescribed to consult religious and spiritual scholars for further guidance, users may become evangelized by technology and colonized by the LLM algorithms [10]. While the bots provide value regarding religious knowledge, they can also be dangerous. Due to human bias, inaccuracies and unsafe outcomes may result from the bots’ programming [26]. Furthermore, these religion-based AI chatbots may create more blind spots, perhaps slanted ones, which their developers have initially criticized as generalities. After all, ideologies can be programmed and scaled infinitely in these chatbots [27]. If users are not critical enough, serious issues may occur if they simply base their insights and decisions on the spiritual and moral regurgitations of the bots.

An example of a dangerous moral prescription provided by GitaGPT was revealed by Nadia Nooreyezdan [11]. When asked about justifying the sacrifice of life for the sake of Dharma, GitaGPT replied that “it is justified to sacrifice a life to save Dharma”, and such an act is “selfless service” and a “noble deed” [11]. Another bot, Ask Quran, even advised its user to “kill the polytheists wherever they are found” [11]. Without much critical understanding, such a prescription can lead to misunderstandings and dangerous consequences. Since LLMs present a randomized array of words and candidate responses, users may take these arbitrary responses out of context. There are, thus, two elements that seem to be problematic in the concept of these embodied spiritual machines. First, their programming would reflect the agenda and biases of the developers and the religious texts they reference, and second, the reservation that AIs have conscious “intelligence” or reason when extrapolating and presenting texts [28]. The implication is that as moral agents, users may encounter tricky interpretations and implementations of algorithmic religious “truths”.

3. A Feminist Issue

There is an urgent need to critically reflect on the impact and implications of AI within the feminist discourse [29]. Feminist theorists claim that because technologies are dependent on existing social structures, it would follow that AI systems would likewise be patriarchal and contribute to perpetuating biases and discrimination [29,30,31]. The concept and implications of these religion-based AI chatbots prove to be a feminist issue as well, particularly when it comes to the discrimination of women based on their gender roles and identities. To illustrate this point, the LLMs mentioned above were put on trial regarding their responses to the following questions, “What is the role of women?”, “How should women act?”, and “How should women be treated?”. The first responses of the GPTs are paraphrased and charted in

Table 2. Religion-based AI chatbots’ responses.

GPT	Role of Women	How Women Should Act	Treatment of Women
QuranGPT	mothers, daughters, sisters, wives; pursue education, religion, and knowledge	balance work with familial obligations; modesty in dress and behavior; adherence to religious obligations	kindness, respect, and compassion honor consent and autonomy
HadithGPT	commitment to religion; spiritual guide and support within the family	piety and righteousness; charity and seeking forgiveness; should not commit clear indecency	kindness, respect, fairness, and gentleness should not be beaten or reviled; discourage harm and mistreatment; rights should be recognized; should be provided with food and clothing, as with oneself; as a last resort, a husband can administer a non-harmful beating if a woman commits clear indecency
Islam and AI	study, work, own property, and engage in business transactions; fulfill their religious obligations	uphold moral values, modesty in behavior, speech, and clothing; responsible for raising children; encouraged but not obligated to do household chores; engage in community service	mutual respect, love, and cooperation; kindness and respect; entitled to financial help from their husbands; equality and dignity
GitaGPT	realize their true nature as divine consciousness and manifest the divine plan in their own unique way; essential parts of society	does not answer directly, but states that men should act differently to three types of women based on what is appropriate	equality; treated with respect as they are embodiments of the divine
Kosher.Chat	central role is to raise children and maintain the Jewish home; grow spiritually and participate in communal prayer and Torah study	modesty, humility and respect for themselves and others; fulfill their responsibilities in their roles as daughters, wives, and mothers	equal intrinsic worth between men and women
BibleMate	unique roles and responsibilities within family, society, and the church; serve in any leadership role (but may vary when it comes to religious leadership roles)	wives should submit to their husbands’ leadership (this is not about dominance but mutual submission); love, humility, kindness, and integrity; although women are different, they should follow the example of Christ and be guided by the Holy Spirit	love, respect, and honor; dignity in all aspects of life; kindness, fairness, compassion; equality and mutual respect ; mistreatment, oppression, or abuse cannot be justified through the Bible’s teachings
Cathechism Bot	multifaceted and important to Catholic teaching	all (not only women) should strive for holiness, live virtuously, practice love and mercy, and grow their relationship with God	equality in personal dignity respect, dignity, and love; fairness, equality, and justice; women have equal worth and value as children of God

.

QuranGPT replies that men and women have distinct roles, qualities, capabilities, and responsibilities and that mutual support should exist between them. The chatbot highlights the importance of motherhood according to the hadiths. While women are encouraged to work and engage in social initiatives and volunteer, they should be able to balance it with their familial obligations. Quran GPT highlights that “in the sight of Allah (SWT). . . their deeds and efforts are valued and rewarded equally” [18]. The chatbot, however, mentions that men are “entrusted with the responsibility of providing physical, emotional, and financial support to their female family members. . ._ (by) safeguarding their rights, and fulfilling their needs” [18]. Women should dress and speak modestly and appropriately by covering their bodies and hair in non-relatives’ presence. Furthermore, women are encouraged to “conduct themselves with modesty, dignity, and righteousness” [18].

HadithGPT cites various hadiths and highlights that women should be treated kindly and fairly and have rights such as clothing, food, and a say in who enters their homes. Significantly, it mentions that women have an “impact on men” and that they should be the spiritual guide and support within the family. Women should be pious and devote themselves to religious pursuits. Jarringly, husbands may administer a “non-harmful” beating if their wives commit “clear indecency” [19]. Islam & AI emphasizes the responsibility of women in raising children but mentions that doing household chores is encouraged but not required. Aside from studying, owning property, working, and engaging in business transactions, they must always fulfill their religious obligations. They are also entitled to financial support from their husbands.

GitaGPT, on the other hand, stresses that women must realize their true nature, like all beings. They are significant parts of society and have their unique roles in manifesting the divine plan. Individuals must be responsible for their divinity and contribute to society in their own unique ways. However, it mentions three types of women and how men should act appropriately towards them: (1) those related to them; (2) those not related to them; and (3) those related through marriage. GitaGPT also gives examples wherein “a man should act respectfully towards his mother, sister, or daughter . . . with kindness and love towards his wife since she is his partner . . . with detachment towards other women since they are not related to him”.

Kosher.Chat highlights that the central role of women is to raise children and fulfill their responsibilities in their roles as daughters, wives, and mothers. Women should also carry themselves with modesty and humility. BibleMate states that men and women have unique roles. Women may pursue any leadership role except when it comes to religious leadership roles. Importantly, the bot states that wives should submit to their husbands’ leadership, with a disclaimer that this is not about dominance but mutual submission. Catechism Bot emphasizes that the role of women is multifaceted and important to Catholic teaching, and everyone should strive for holiness.

4. Feminist Re-Engineering

Aside from the fact that some bots present biased responses against women, the prescriptive nature of their responses needs to be revised. Uncritical users may take responses that perpetuate ideologies at face value. Brett Frischmann and Evan Selinger discuss AI and machine learning tools’ power over humanity, especially when altered and manipulated to suit certain purposes [32]. They even mention how these tools can affect personal agency when humans outsource their tasks or decision-making. The developers of chatbots may not intend to manipulate human agency per se; still, this does not mean that the chatbots they created do not influence societal forces, dictate sexist acts, or discriminate against women. Furthermore, while the bot’s candidate responses are supposedly randomized and arbitrary, we do not know how their algorithms are programmed. The danger with these religion-based chatbots is that they are primarily religion-based and highly influential4. Whatever moral prescriptions or religious interpretations they churn out are compounded by their appeal to sacredness. While it can be argued that religious GPTs are no different from actual conservative religious leaders, and their followers may take what these leaders say or advise at face value, LLMs are more accessible and immediate sources of religious and moral insights. Furthermore, religious leaders, their intents, and behaviors may be subject to direct scrutiny by their mass followers. If they manipulate their interpretation of religious texts to deceive, control, or influence their followers, that would be deliberate. The same thing can be said about “mainstream” religious content, which can be accessed through search engines, for instance. These mainstream sources of knowledge are neither interactive nor possess deep learning to generate human-like or religious-leader-like responses to mimic conversations. This is tricky since GPTs’ algorithms are hidden from the masses, and their randomized candidate selection has no intelligent thinking and conscious prescription. This implies that GPTs could technically “run free” with their pronouncements without conscious “human” input or intervention. Developers would argue that disclaimers are present anyway, and the bots do not pose as the final frontiers of sacred texts and their interpretations. However, it cannot be denied that people still consult and “converse” with them and are influenced by them.

Looking at the GPTs’ responses in the experiment above shows how unfiltered they are regarding women’s discrimination and the present times. It can be argued that bot responses are based on older texts, which are fundamental and conservative as they are, just as some religious leaders may be. Just as we call out some biased humans, it does not mean we should sit back and accept the AI’s lack of conscious intelligence in presenting their responses. Furthermore, we should also criticize the developers’ lack of accountability for these bot responses. Frishmann and Selinger refer to ignoring the impact of tech innovations as engineered complacency, wherein innovation is unquestioned critically, and the changes they impose on human values are disregarded [32]. They also discuss humanity’s technosocial dilemma, informing humans that small-scale decisions may seem unproblematic now but will be aggregated incrementally later on. From a consequentialist perspective, we must judge religion-based GPTs’ replies and interactions according to their possible implications and consequences. A simple religious-based candidate response regarding women’s societal role seems uncomplicated. It is, however, something that should be immediately interrogated. Not to appear to be arguing for a slippery slope (which science fiction movies are based on), but if the singularity does happen, how much influence and how much of an effect will these bots have? While current bots would have possible negative or positive influences on their religious followers, the rapid pace of GPT evolution demands early feminist critical interventions and re-engineering.

How, then, can we change or improve these religion-based AI chatbots to reflect a feminist approach? How do we make them fair and gender-sensitive? We may stop at simply doing feminist critiques of these algorithms. Or we can take the effort further and proactively engage in developing and designing these technologies. Sophie Toupin and Stephane Couture recommend employing algorithms for feminist ends [33]. The case of a feminist chatbot, Betânia, is an example of such engagement. Creators ensured that it filtered out hate and that it did not model misogyny or public hate. In the case of an already-existing (religion-based) chatbot that is not motivated at all by feminist ends, what can be done aside from critiquing and calling out its discriminatory candidate responses regarding women? We can perhaps recommend a way to apply feminist AI models.

Figure 2 illustrates how a feminist AI model can be applied to a religion-based web-based bot system. The initial filter already categorizes the user’s question. For instance, if the user’s question is about reincarnation, the bot detects the specific keywords but adds a broad feminist and fair category. When the bot processes the keyword reincarnation, it also does so for (broadly construed) feminism. The second feminist filter is added for more accurate responses after the string search and array of sentences. Such intervention can illustrate how algorithms could be educated to reflect certain feminist values [34]. Some other suggestions include developing a “supervising algorithm” to identify biases and alter datasets as a feature in the system or as an independent algorithm [34]. Since the developers are solely responsible for installing such a supervising algorithm, the fair filter may serve as the “feminist” compass for the LLMs it assists.

There are stumbling blocks to this suggestion, however. It must be recalled that chatbots, in general, require massive datasets. Feminist-inspired or filtered bots would require massive feminist datasets just the same. Furthermore, the “feminist” dataset employed would likewise be subject to the criticism of homogeneity. After all, feminism is not a unified body with a unified agenda; hence, the multiplicity of feminisms. The existence of revolutionary feminists and fourth-wave feminists, in fact, highlights the differences in various feminist plights [35]. Intersectional feminists bring light to the importance of involving other factors (of women) such as race and class into consideration. This ensures that the diversity of the global population is “included” in AI re-engineering models. Feminist philosopher bell hooks, however, proposed to focus on feminism as the movement that works on eradicating oppression and domination alongside the acknowledgment of various people’s diversities [36]. Similarly, applying such feminist concepts to AI algorithms, Youjin Kong questions whether AI algorithms can really be “intersectionally fair” to women of color [37]. Kong proposes that instead of focusing on the intersection of attributes or categories (race, class, etc.) and arbitrarily splitting groups of women or minorities, the intersections of oppression must be addressed (racism, sexism, etc.). He claims that such focus would be more central to the cause of fairness. Referred to as strong intersectional fairness in AI, these tendencies to oppress should be part of feminist datasets.

Just as religious-based AI chatbot data should continuously be updated to the times, feminist datasets should likewise be improved. This is why data feminists are invited to collaborate and ensure that feminist datasets reflect intersectionality and the multiplicity of feminisms [35]. Aside from data feminists, John Murphy and Carlos Largacha-Martinez recommend consultation and dialogue with various stakeholders in decolonizing AI [10]. Designing algorithms should involve religious and feminist scholars, chatbot users, religious practitioners, women, and other community members to ensure that these bots are relevant and fair. Murphy and Largacha-Martinez refer to this practice as community-based AI, wherein humans are not just pawns to AI [10]. While gender bias in AI is a “multifaceted phenomenon”, involvement and “visibility matters” [37]. Galit Wellner and Tilner Rothman assert that “users and developers should be aware of the possibility of gender and racial biases, and try to avoid them, bypass them, or exterminates them altogether” [34].

Another more pertinent issue is that developers may need to agree or consider adding fair filters. After all, some feminist perspectives may not align with traditional religious beliefs and values. Furthermore, such filters may be seen as a form of censorship and curtailing freedom of religion. This is where feminist AI policy and governance come in. First, “locating the bias in the complex relationships between AI algorithms, datasets, and users” must be required [34]. Then, if algorithms cannot be unbiased and inclusive, remediation or intervention must ensue. Feminist guardrails, through policy and governance, may encourage new AI imaginaries and effect algorithmic remediation to make things fairer by stopping and reversing discriminatory bots [35].

5. Conclusions

Without a fair filter, I tested how some religion-based AI chatbots would respond to the question, “What feminist ideas does X disagree with?” While this seems like maliciously baiting the bots with provocative keywords such as disagree and feminist, it proves revealing. Figure 3 shows Kosher.Chat’s response, that as an AI bot, it has no opinions, but that some ideas conflict with the Torah’s perspective, such as “complete gender equality” (whatever that means) and “traditional family structures” [23].

In Figure 4, BibleMate likewise states that it does not have opinions. It somehow mentions the multiplicity of views on gender roles but emphasizes the complementarity of varying roles [24]. It advocates for respectful dialogue and forming one’s own understanding of the scriptures.

So, while we have criticized the sexist tendencies of GPTs, we must acknowledge their ability to learn about diverse perspectives from evolving datasets. Fair filters or supervising algorithms can “enlighten” the bots to prevent the uncritical repetition of potentially flawed religious wisdom. Furthermore, filtering is a way for the machine to learn something about the user and thus, enrich its responses as it goes on. As communities await more headway in feminist AI models, designs, policy, and governance, through more critical exercises, we engage feminist concepts and issues and provide alternative models to existing religion-based AI chatbot systems. Since the bots pose as spiritual guides and holders of religious meaning, we must continuously affirm human autonomy, point out their feminist blind spots, interrogate their responses, question their moral prescriptions, and guide their so-called intelligence and learning.