Sacred Meets Synthetic: A Multi-Method Study on the First AI Church Service

Literature Review

There has been limited empirical research thus far on the relatively novel phenomenon of AI in religious services, this project is exploratory. Therefore, literature attending to this specific phenomenon was not found. However, adjacent phenomena—such as AI in the religious sphere and religious community in digital environments or with digital tools—has been researched more widely and helps to provide context to this study. With ChatGPT making large language models (LLMs) accessible to a wide public, their use is also relevant within religious spheres. LLMs can act as a sounding board or streamline ideas and processes, which is why some religious professionals have embraced their use.

For example, the AI model Sermonly is an AI assistant for writing sermons. Even earlier, more rudimentary applications have used (smart) algorithms to facilitate their religious practices: a British convent of the Poor Clares uses a device called The Prayer Companion, which scans international news for potential events that might inspire the nuns' intercession prayers (Gaver et al. 2010). Additionally, smart technology can be worn privately in the form of the eRosary, which tracks daily prayer progress (Vatican News 2019). Apps like Evermore guide users through daily Christian meditations and also track individual progress. Yet, research on the impact of those generated outputs on user experiences is yet sparse (Chua 2024).

Research shows that Christian practices of digital religiosity are becoming increasingly accepted and widespread (Campbell 2020; Wilkins-Laflamme 2022). According to Hutchings (2017), digital religiosity typically includes three sets of experiences: sustained supportive social interaction, a shared feeling of belonging, and a practice of prayer. These practices hold significant value for their users, and these spaces should be considered as meaningful forms of religious community (Hutchings 2017). Additionally, with virtual reality environments being explored both as meeting spaces for congregations (Jun 2020) and as educational spaces (Markert 2023), avatars as representations of remote participants are enabling new forms of community in digital places.

Additionally relevant is Campbell’s (2012) concept of a networked religion: western religion’s shifting in form is part of a broader change in cultural practice that is becoming increasingly more global, digital, and interconnected. Online and offline practices of religion are not strictly separated or dichotomous but rather are integrative and mutually transformative. Therefore, an AI church service can be regarded as part of a networked religion, which integrates traditional religious practices within digitalized contexts. Campbell (2012) calls these “convergent practices,” blending technology with traditional liturgical elements. Additionally, Evolvi (2022) argues that digital religiosity is not immaterial, rather as embodied in that material technologies and embedded in everyday practices. An AI church service can be seen as a materialization of a digital religious medium (AI creating religious contents) into the material spaces that religious communities inhabit. In this way, AI is a religious agent, and thus barriers between online and offline, in-person and digital, are transcended into a digitally connected and AI-interfaced religious network.

Contemporary theology is exploring the idea of AI’s status as a religious agent. The discussion surrounds topics of human likeness as an image of God and religious properties usually reserved to humans (Herzfeld 2002). Furthermore, topics such as redemption (Strand 2021) and sin (Smith 2022) are discussed for these non-human actors. These considerations apply to both AI applications and robotics. Robots have also been used in various religious settings: from the Buddhist preaching robot Mindar (Hardingham-Gill 2019), or the educational companion monk Xian'er that leads meditations (Cheong 2020), to the infinite, interreligious prayer robot The Prayer by Diemut Strebe (Strebe 2020). Internationally, robots are increasingly populating the religious sphere (Puzio 2023; Simmerlein and Tretter 2024). Apart from some exceptions (Löffler, Hurtienne, and Nord 2019; Riek et al. 2010), only a few robotic applications used in religious practices have been empirically researched.

An outlook on the research field shows that the last years have seen a noticeable increase in research interest in religious practices with and around Artificial Intelligence (AI). Both robotics as well as language generation have been tried as means to add to or enhance religious practices. Yet, the novelty of the phenomenon means little research exists on how religious attendees experience these artificial agents, and whether any differences exist among attendees in terms of who is more receptive versus skeptical of AI pastors.

Project Aims and Design

The primary aim of the project was to explore the reactions of a religious community to attending a service with a distinctly non-human agent leading the congregation. In order to ensure authentic responses, an overriding requirement in both the creation and delivery of the service was to remove the human element as far as possible. Since a human agent conceived and realized the idea, there is an irreducible human element underlying the process, but the aim was to avoid human interference whenever possible. Another design feature was to aim for the process to be as intuitive and affordable as possible, to create a service that could theoretically be adopted by any congregation that chose to explore the use of AI without requiring extensive resources, inaccessible technology, or expert knowledge. The reason for this decision was to create something that was replicable, in that it could theoretically be implemented by any congregation and feasibly realistic to attendees as implementable.

Generating Sermon Text

Technically, the basic content of a worship service is text. Both situationally created text for a given Sunday (sermon, prayer) and traditional text recited at various points in the liturgical order (blessing, readings). Recent advances in text production based on large language models (LLMs) thus technically allow the integral parts of a service to be created by algorithms. The creation of the sermon began with training OpenAI's ChatGPT on some basic parameters: its role as an AI leading a service, its context in the German Protestant Kirchentag, the motto (Mark 1:15: “Jetzt ist die Zeit”/“Now is the time”) of this year's iteration of the event. With these basic instructions, ChatGPT was prompted to outline the concept of such a service.

Following this framework, subsequent prompts caused ChatGPT to generate the underlying text according to the structure it had previously outlined, while at the same time prompting it to consider existing text. This ensured the coherence of the liturgical texts. The developer’s only active intervention was to prompt the model to elaborate further (prompt: “elaborate further on this point”) to ensure that the overall text was of sufficient length. Since ChatGPT's outputs are of varying quality depending on the input language, the process was conducted in English and the text was then translated by the AI-driven translation software DeepL. While there was human oversight to prevent the model from formulating hateful or inappropriate content, no intervention was required, and the first draft was in each case used to maintain the principle of the least amount of human intervention in the process.

Accuracy of Liturgy

Apart from a few minor errors, which were deliberately left uncorrected to limit human intervention, the resulting service contained the usual elements of Lutheran liturgy. Typical elements of a Bavarian Lutheran service were presented by the AI in a way that corresponded to the formula for such events (ELKB 1994): Greeting, Introductory Prayer, Psalm, Kyrie and Gloria, Daily Prayer, Bible Readings, Hallelujah, Confession of Faith, Sermon, Intercessions, Lord's Prayer and Blessing. By adhering to the traditional outline of the Lutheran liturgy aspects of Lutheran theology which make the service meaningful to its participants like the justification of the sinner or the central status of the scripture (Meyer-Blanck 2009) were present throughout by mere consistency with the tradition. This was apparent in the reaction of the congregation, which participated in the usual manner of a Lutheran service: praying together and reading the Psalms.

One exception was for the sacrament of communion, which is optional in the Lutheran tradition and is celebrated only occasionally in many Bavarian Lutheran churches. Since communion consists of the distribution of bread and wine, which an AI without bodily characteristics cannot do, the sacrament could not be celebrated under the circumstances. Otherwise, the AI accurately included all typical elements of a Bavarian Lutheran service.

Video and Audio Creation

The AI video creation tool Pipio was utilized for a visual and audio representation of the text through animated actors. Several two-dimensional avatars presented the spoken parts of the service. Figure 1 displays an example avatar that was used to visualize an AI-generated human delivering the sermon. The tool also implemented speech support by reading the AI-generated sermon text aloud in pairing with the avatar visualizing the speech. Additionally, AIVA was trained on existing hymns to create AI-generated music, and thus the audible music accompanying the service was also composed with artificial intelligence. In summary, the AI mimicked five central features of religious services: religiously inspired sermon text, liturgical accuracy, visually captivating speakers, aligned auditory speech, and music.

OPEN IN VIEWER

Figure 1.

One of the avatars used as pastors in the service.

Despite these AI advancements, there were also technical limitations. One was the lack of typical Lutheran robes, and a second was the inability to direct the avatars with gestures such as raising hands during the blessing of the congregation. Third, the choice of which avatars to use may have influenced reactions. Fourth, attendees were not able to lead the singing or join in to traditional communally sung hymns. Fifth, a procession to communion could not be supported, and though this is an optional feature of Lutheran services it could have added to a limited experience as a religious community. In summary, the AI also lacked four to five central elements of religious services and could not readily mimic the communal experiences of many religious events.

Data Collection Instruments

The first tool used to assess participants’ reaction to the experience was a survey that could be completed online or afterwards in written form for 3 days after the service was conducted (Supplemental Appendix 1). Event attendees were informed before the service started that a qualitative questionnaire and a quantitative survey were open for voluntary participation, which would be used for empirical research by the creator of the sermon and the author of this research note. The addition of a physical response mode, in addition to the digital one, was intended to include those who were uncomfortable with mobile phone surveys or did not have or bring a mobile device to the event. The survey was in German, and the responses correspond to the suggested translation in the English UEQ+ or were translated accordingly where these were deemed inappropriate.

Survey Measures

The first questions in the survey asked about age, subjective religiosity (from “not religious” to “very religious”) and denominational affiliation. These three were hypothesized to be useful indicators for t-tests. Additional measures included emotional ratings.

Age. Half of the participants were aged between 46 and 65 years (n = 38; 48.7%), while a comparable group was aged between 16 and 45 years (n = 36; 46.2%). Those younger than 16 or older than 65 were too few to be considered separately (n = 4; 5.1%).

Subjective Religiosity. The subjective religiosity (see Figure 2) of the participants was rather high (µ: 4.975; median: 5 on a 7-point Likert scale), which can also be explained by the context of the event: although the service was generally open to everyone, it was part of a larger religious event lasting several days, for which a ticket had to be bought, so that it was almost exclusively attended by participants of the German Protestant Kirchentag.

OPEN IN VIEWER

Figure 2.

Subjective Religiosity from ‘not religious’ (1) to ‘very religious’ (7).

Denominational Affiliation. 83.5% of the participants identified themselves as Lutheran or Reformed, 7.6% as other Protestant denominations (Evangelical, Pentecostal, etc.) and 6.3% as Roman Catholic, one person as non-religious and another as another Christian denomination. One person did not indicate their religiosity. Given the context of worship at the German Protestant Kirchentag, the majority belonging to institutionalized Protestantism is hardly unusual. Given the homogeneity of denomination, the data set did not allow for robust conclusions based on this criterion.

Emotion Rating. Participants rated 16 different emotions on a 7-point Likert scale, with opposite emotions on the side, using the well-established UEQ+ (accessible at https://www.ueq-online.org/) developed by Schrepp and Thomaschweski, which has been used in various settings to assess the experiential value of technology. The UEQ+ consists of different scales, which are composed of items representing different emotions. For this research, the scales measuring attractiveness, stimulation, usefulness and value were considered relevant to this project. Scales that are relating to the personal interaction and customization (e.g., adaptability or perspicuity) or those that are measuring the experience of the developers (e.g., dependability) were excluded as well. Schrepp and Thomaschweski (2019) also advised which scales are best used in which context. The selected scales are both rather broad, which makes them suitable for various experiences attendees could have had with the church service, as well as focused on emotional reactions. Rather than rating the service as a product, these scales emphasize the subjective experience of the participants.

Practical-Theological Implications

There are wide discrepancies in the reception and mixed results on the question of the desirability of incorporating AI into the specific practice of worship. The results evidence a diversity of opinion, which ranges from anticipation like “I am looking forward to the future” to discontent like “Incredibly terrible.” To decide either for or against the usefulness of such technological advances would be to exclude a significant portion of the participants. Clearly, a relevant number of participants were invested in the experience, while others were irritated.

Whether a long-term exposure to such an AI-based worship service would change the experience cannot be answered here but would be interesting for understanding the overall impact of artificial intelligence in religious practice. Further, larger, and more precise research would be helpful in future studies. Using an approach based on uses and gratification theory (comparable to e.g. Ratcliff, McCarty, and Ritter 2017) could be informative. To do so would require a more conversational methodology to interviews investigating the specific needs and intentions of each participant. The data suggests an ambivalent stance of attendees toward the experience of AI in religious worship.

Expectations and Overestimations About AI in Religious Practices

The data indicate a wide array of expectations about what AI is and does in worship practices. While some were surprised about the general capabilities (“How much GPT is actually capable of”), the majority was underwhelmed or had expected more of the state of technology: “AI performance [is] surprisingly bad.” Amaras’ law also applies to religious settings: “We tend to overestimate the effect of a technology in the short term and underestimate the effect in the long term.” (Ratcliffe 2016). It is worth noting that the service was conducted with limited preparation time and resources, and a more impressive realization of the project could be imagined, especially due to the technological advances made in the meantime. Yet, there was still an obvious case of overestimation in many of the responses. Since most churchgoers are not avid technophiles, false assumptions and conclusions about the technology seem inevitable. It is unclear whether this technology will soon, or ever, reach a level of sophistication that participants will find acceptable on the whole.

While interest and openness to technological innovation in the religious sphere is conceivable, the gap between expectation and reality may be an ongoing problem in the acceptance of technology. This is all the more notable given the highly complex nature of a liturgy. Minute details, movements, and symbolism must be considered, which makes religious practices particularly challenging for technological advances, even more so when they are supposed to be largely autonomous and not corrected at every move by a human expert. The acceptance of AI in religious practices hence is likely to be related to the overall knowledge about the technology. At the moment, AI in religious practices is affected by a high degree of ignorance about their limits and possibilities.

Religious Experiences with Technology

Participants in the questionnaire report having made religious experiences with the technology. While the quantitative measure is not sufficient to separate positive experiences from specifically religious experiences, some of the written responses suggest that some experiences are interpreted as specifically religious (“Spiritual feeling was present!” or “God works in various ways”). Whether this is sufficient for that label may still be debatable. Religious or not, the seriousness of the experiences is hard to deny. While the number of people who are deeply moved by such practices may be small, they should still be taken seriously. Further research would be needed to explore the question of whether our society is becoming more receptive at the experiential level to technology in religious and spiritual matters. The data clearly suggests that such experiences are possible.

Anthropological Differences

Technology is a prominent part of the responses, yet some hurdles may not be overcome by innovation, since they are grounded in differences between AI and humans. Some answers pointed out genuine anthropological differences. Three of them, reflected in the reactions, are worth mentioning. First, there is an undeniable gap between humans and AI in their subjective nature (“AI can’t incorporate personal experiences”). As religion deals with mortality, guilt, fear, pain, and other emotive experiences, AI relates from an unobstructed distance. This, secondly, has to do with their lack of embodiment (“Embodiment and resonance are missing”). While algorithms can have a spatial representation, as robots do, they lack the flesh of people in the pews (Henry 2016).

Similar to the lack of subjective experiences, respondents clearly felt a disconnect with the AI due to them missing human flesh (“The Word became flesh… For a true experience of faith, I need a human experience”). What the reactions here suggest, thirdly, is a monumental difference between human-human and human-AI relationships. What the participants were missing was the interpersonal relationship that humans experience in religious communities. When the human pastor is replaced by an avatar, this basic human need for community and co-presence between congregation and pastor is no longer met (“No personal relationship to the speaker yet” or “Somewhat impersonal.”).

However, what this reveals is a narrow focus on the pastor. With the AI failing to evoke interhuman resonance, the peers in the pews became more present, and through interjections during the sermon and communication between churchgoers, different and unusual forms of interpersonal relationships became more prevalent (“Experienced a sense of community during the Creed and the Lord's Prayer.”). This underscores the empirical research on emotions during the sermon where participants repeatedly suggest, the intricacy of religious experiences in communal practices. While some proponents of the protestant liturgical tradition long held the belief that the pastor is the integral focus of the experience of the sacred (Josuttis 1997) more recently empirical research shows how the congregation itself creates and atmosphere of joined religious emotionality (Pohl-Patalong 2011).

How and whether these differences could be overcome remains highly controversial. Remaining in an empirical framework this difference could be explored by contrasting AI church services with human-led services. Additionally, it could be possible to have AI-human congregant interactions that are contrasted with human-human congregant interpersonal interactions. While there might remain a gap between those two, further empirical research focusing on comparisons could provide the background for such assessments.

Limitations

The study has several limitations that warrant caution. All participants had to be fluent in German both to attend the service and to complete the survey. The study only allows conclusions to be drawn about a specific group of religious people, notably mostly Lutheran attendees at a religious celebration in Germany. Research participants volunteered for both the service and the survey. In contrast to the physical survey, it was not possible to participate in the qualitative questionnaire without a smartphone. Due to the infrastructure of the Mentimeter app, responses were anonymous and did not allow for follow-up questions.

Inputs by experts might have influenced the responses by the participants to some degree. While the qualitative questionnaire was already concluded, when the panel of experts discussed their experience, most of the participants of the quantitative survey rated the experience after having heard experts reflecting on the event. This might have to some extent swayed participants in their reflection.

In addition, the measure of subjective religiosity was not well specified. The self-assessment used in the survey may be one reason why religiosity yielded inconclusive results. Important and more complex measures about religious beliefs or behavior could be included in future studies to yield a deeper understanding. While the used UEQ+ successfully measured different degrees of emotional reaction, models more finely tuned to the specific religious context might bring a more granular understanding of the subject.

Further research may be needed to determine how religiosity correlates with the reception of AI in religious practices as such. More generally, more and larger research samples could investigate receptivity among more religiously diverse groups, with attendees representing a broader range of denomination affiliations. It would also be of value to seek to overcome self-selection effects by hosting an AI service within an event that was not explicitly religious in order to gauge interest among unaffiliated or non-religious attendees.