Abstract
Communication scholars have approached heterogeneous experiences in romantic interactions online from the perspective of algorithmic beliefs. People with higher algorithmic beliefs trust more that algorithms can help them find compatible matches in online dating. Such algorithmic beliefs have been theorized to have their effect through the mechanism of self-fulfilling prophecy. The current study offers a more granular test of the underlying cognitive and emotional mechanisms using a scenario-based design. Undergraduate students (N = 101) who had online dating experiences were randomly assigned to report reactions to 4 of 24 unideal online dating scenarios, producing 404 observation points. Crossed random effects modeling found that participants with higher algorithmic beliefs had more positive interpretations and fewer negative interpretations in unideal online dating situations. However, algorithmic beliefs were not related to distress. Findings suggest that algorithmic beliefs may enhance online dating experiences through facilitating more adaptive appraisal processes and offer insights for potential interventions against online dating burnout.
Introduction
More U.S. couples now meet their partners online than through other means (e.g., friends and family), and over 30% of U.S. adults report having used online dating.1,2 Despite its growing prevalence and benefits, online dating has sparked concerns. Chief among these concerns are the frustrating online dating experiences,3–6 and in general, lower relationship success compared with relationships that begin offline.7,8
These findings show challenges in navigating the complex algorithmic online dating systems. 5 To explain, online daters are typically connected through computational procedures—commonly referred to as algorithms, which are critical in understanding online daters’ social experiences. 9 Communication scholars have therefore approached online dating experiences through an algorithm-related mindset, algorithmic beliefs.4,10,11 Algorithmic beliefs refer to the beliefs that dating algorithms are a legitimate means of matchmaking; for example, people with high algorithmic beliefs may agree with statements like “Dating algorithms will provide me with better quality partners.” 10 Traditionally, views have held that love and affection cannot be calculated or approached with stringent rationality. 12 As such, the idea of finding compatible romantic matches through algorithms is novel and countertradition.
Such novel beliefs can have positive effects, however, which, per previous conceptualization, can be explained by the self-fulfilling prophecy.4,10,11,13 In face-to-face interactions, it has been established that interpersonal beliefs and expectancies shape how people appraise social interactions and their subsequent response. 14 For example, individuals who are made to believe strangers dislike them prior to an interaction can interpret strangers’ behaviors as more hostile than actually are and respond in less likable ways, which can make strangers actually dislike them at the end of that interaction. 15 Sharabi in the first study of algorithmic beliefs, as well as Hu and colleagues in their follow-up studies, contended that people with higher algorithmic beliefs may have stronger faith in their online dating sessions and produce more positive appraisal of their online interaction, thus approaching their potential partners in more likable ways and perceiving better experiences online.4,10,11 Empirically, higher algorithmic beliefs have been linked to lower negative emotions online.4,10,11
Though the sociocognitive mechanism has been central to the theorizing of the self-fulfilling prophecy in the extant online dating literature, the connections between algorithmic beliefs and how users appraise a specific interaction in online dating have not been empirically examined.4,10,11 Lazarus contended that appraisals of an event are primarily concerned with the valence of the event’s impact. 16 As initial interactions online are often characterized by ambiguity and uncertainty, online daters can have both positive and negative interpretations about these interactions. 11 For example, the other person being emotionally reserved may be interpreted as “not being interested” or “being serious and taking things slow.” Both interpretation types can shape online daters’ subjective experiences and are therefore examined in the current study. 16 Aligning with prior work, general distress—a commonly assessed malignant emotional outcome—is also assessed in the current study.4,10,11 The rationale of the self-fulfilling prophecy has suggested that algorithmic beliefs can encourage more adaptive responses to online dating interactions—that is, lower negative interpretations, higher positive interpretations, and lower distress.4,10,11 We therefore hypothesized that algorithmic beliefs are negatively related to negative interpretation (H1), positively related to positive interpretation (H2), and negatively related to distress (H3).
We employed a scenario-based approach to test these hypotheses. We specifically focused on unideal scenarios because individuals’ reactions tend to be more maladaptive and concerning, and our intention was to observe how people respond differently to less-than-satisfactory situations. Compared with past studies, which employed generic retrospective reports or tracked ongoing relationships,4,10,11 the scenario-based approach allowed for experimental manipulation and scenario-level examinations. Examining many scenarios also helped produce more generalizable findings than approaches using fewer scenarios (e.g., a speed dating experiment that trains confederates to send certain messages). Taken together, the current study helps construct a theory of algorithmic beliefs in understanding online dating experiences, which can be useful in addressing current real-world concerns about online dating and hopefully other online social problems alike.
Methods
Procedures
Undergraduate participants (N = 130) were recruited from a large Midwestern university in February 2025, and removal of responses by participants with no online dating experiences led to a final sample of 101 participants. They were randomly assigned to 4 of the 24 hypothetical unideal online dating scenarios, presented in random order. Participants imagined each scenario happening to them and reported their reactions in each scenario. Next, they completed two distractor tasks (i.e., counting subway lines, simple calculations). Finally, they completed measures of their algorithmic beliefs in general as well as trait and demographic measures. All procedures were approved by the Institutional Review Board of the authors’ university.
Participants
The mean age in the final sample was around 20 years (SD = 1.66). Most participants identified as women (n = 65) or men (n = 35), and one identified with another term. The sample was predominantly White (n = 83) but also included participants who identified as Asian (n = 9), Black or African American (n = 6), Native American or American Indian (n = 1), or multiple races (n = 2). Some participants were Hispanic (n = 4), and others were non-Hispanic (n = 97). Some participants were actively using online dating (n = 23), and others were active users before but not currently (n = 78).
Scenario development
One author who is an experienced online dating researcher drafted candidate scenarios that are likely perceived as unideal in online dating, based on previous qualitative findings. 17 Each scenario is described in two to three sentences. All authors—who are regular online daters—discussed the face validity of the scenarios. That is, the scenarios should be common, represent uncertainty, and allow for varied interpretations (unlike clearly harmful situations like sexual harassment, where idealization would be maladaptive and thus unsuitable). After another round of scenario modification, we eventually arrived at 24 scenarios (Supplementary Appendix). For example, for “shortened message,” participants were asked to imagine the following scenario happening to them:
You met someone online. After an initial period of long, engaging conversations on dating apps, social media, and/or through text messages, their messages start getting noticeably shorter.
Two common criteria assessed for hypothetical scenarios are realism and functionality. 18 We measured how participants perceived realism of and relationship uncertainty in the scenarios to further check whether the scenarios we used were indeed qualified. We wanted to exclude scenarios that are perceived by participants as unrealistic or ideal (i.e., low uncertainty-inducing). To access that, we followed the convention in communication research that employed the scenario-based approach to set the scale midpoint (4 = “Neither Agree Nor Disagree”) as a meaningful baseline for comparison. 18 For a scenario to be included in the final analysis, it should not show significantly lower-than-midpoint realism ratings by participants—that is, not being perceived as unrealistic. A valid scenario should also not show significantly lower-than-midpoint uncertainty ratings by participants—that is, not being perceived as a certainty-inducing scenario. Assessment outcomes are reported in the section “Results.”
Measures
All measures used a seven-point Likert scale (1 = Strongly Disagree, 7 = Strongly Agree). The independent variable of interest, algorithmic beliefs, was measured with the seven-item scale developed by Sharabi. 10 An example item is “Dating algorithms can be better than I am at finding me a partner.” All seven items formed a reliable measure (M = 3.81, SD = 0.93; α = 0.78). For outcome measures, as each participant went through four scenarios, to reduce their burden, distress was briefly measured with two items from the negative affect scale in Positive and Negative Affect Schedule—Expanded Form, 19 including “Being in this situation, I would feel upset” and “… distressed.” (r = 0.68). Negative interpretation was briefly measured with two items, “I could find negative interpretations for this situation.” and “I would have some negative interpretations for this situation.” (r = 0.68). Likewise, positive interpretation was measured with the same two items as for negative interpretation, with the wording “negative” altered to “positive” (r = 0.80). We used such generic measures for our initial test, in line with Lazarus’s typology that suggests that positive and negative appraisals are two overarching relevant primary appraisals. 16 See Table 1 for descriptive statistics and zero-order correlations.
Descriptives and Zero-Order Correlations
**p < 0.01, ***p < 0.001.
SD, standard deviation.
We measured scenario-induced relationship uncertainty and realism to assess scenario quality. Scenario-induced relationship uncertainty was measured with four items from the Future subscale in the Relationship Uncertainty Scale (α = 0.93). 20 Realism was measured with three items assessing how realistic a scenario was, for example, “This scenario commonly occurs in online dating.” (α = 0.88). See the Supplementary Appendix for descriptives.
Finally, we included covariate measures, encompassing demographics such as age and gender, trait measures such as attachment anxiety (M = 4.10, SD = 1.24; α = 0.80) and avoidance (M = 3.14, SD = 1.24; α = 0.75) (The Experiences in Close Relationship Scale-Short Form, ECR-S), 21 trait positivity (The Trait Positivity Scale, TPS; M = 5.01, SD = 1.00; α = 0.88), 22 and mate value (The Mate Value Scale, MVS; M = 5.03, SD = 0.94; α = 0.76), 23 and technology use measures such as prior exposure to artificial intelligence (AI) (α = 0.65), 24 and relationship-seeking motivation in online dating (The Tinder Motives Scale, TMS; M = 3.24, SD = 1.23; α = 0.77). 25
Data analysis
As scenarios and individuals crossed each other, we applied crossed random effects models to our analyses using the lme4 package (version 1.1-36) in R, 26 which controlled random scenario and individual effects. 27 We also controlled for the effects of the aforementioned covariates because they may explain algorithmic beliefs and outcomes of interest simultaneously. We achieved an overall acceptable power of tests. Post hoc power analysis using the simr package (version 1.0.7) in R shows sufficient power in detecting effects of algorithmic beliefs on positive and negative appraisal (β ≥ 0.80), though the power was weaker for detecting effects on distress (β = 0.45). 28 No case was dropped for missing values.
Results
Scenario realism and uncertainty induction
One-sample t tests compared each scenario’s realism and relationship uncertainty ratings to the scale midpoint (4 = “Neither Agree Nor Disagree”). Results showed that most scenarios had higher-than-midpoint realism and uncertainty ratings. No scenario had lower-than-midpoint realism ratings (i.e., none was unrealistic) or uncertainty ratings (i.e., none was ideal and low uncertain). We therefore kept all scenarios for analysis completion. See Supplementary Appendix for t test results.
Hypothesis testing
We ran crossed random effects models and allowed random intercepts for both participant and scenario (Table 2). Results showed that, across the 24 tested scenarios, participants with higher algorithmic beliefs reported fewer negative interpretations of the situation (b = −0.17, SE = 0.08, p = 0.040) and more positive interpretations of the situation (b = 0.20, SE = 0.09, p = 0.030) than those with lower algorithmic beliefs. However, algorithmic beliefs were not related to distress (b = 0.14, SE = 0.10, p = 0.164). Overall, H1 and H2 were supported, while H3 was not supported. As for covariates, participants with higher attachment anxiety, exposure to AI technologies, and trait positivity also reported higher negative interpretation and/or distress. Participants with higher trait self-esteem also reported lower positive interpretation. See Table 2 for specific statistics.
Crossed Random Effects Models Regressing Interpretations and Distress on Algorithmic Beliefs and Covariates
Coefficient estimates are unstandardized.
*p < 0.05, **p < 0.01, ***p < 0.001.
AIC, Akaike information criterion; AI, artificial intelligence; SD, standard deviation; SE, standard error.
Discussion
In homing in on a theory of algorithmic beliefs, the current study used a scenario-based approach to test the relationships between algorithmic beliefs and cognitive (positive and negative interpretations) and emotional (distress) outcomes in online dating. We found that participants who had stronger positive beliefs in dating algorithms’ efficacy in matchmaking had fewer negative interpretations and more positive interpretations than those who had weaker such beliefs in unideal online dating scenarios. The relationship between algorithmic beliefs and distress was not significant. Findings suggest a potential sociocognitive mechanism by which algorithmic beliefs impact online dating outcomes.
Scholars have theorized that algorithmic beliefs can shape online dating experiences through the self-fulfilling prophecy.4,10,11 However, an empirical test of the mediating sociocognitive mechanism is lacking. A critical argument is that positive higher-level beliefs can guide situational appraisal, which then can shape social perceptions and behaviors.13–15 The current study demonstrated that algorithmic beliefs were associated with more adaptive general appraisal of the situation. Although this represents only an initial step in constructing a theory of algorithmic beliefs, it is a crucial step, as it highlights the importance of future experimental and longitudinal research designs to examine the mediating roles of relational construal and appraisals. If people with higher algorithmic beliefs had more benign appraisals, they are more likely to display prosocial and likable behaviors, which can positively shape their social experiences. In this line of future research, we see valuable opportunities to integrate insights from research examining relational mindsets (e.g., growth vs. destiny). 29
We do note that algorithmic beliefs did not lead to lower distress, which is inconsistent with prior work.4,11 A plausible explanation is that other mediating mechanisms work in the opposite direction than appraisal. One potential such mediator is expectancy violation. With greater expectation may come greater disappointment. It is plausible that higher algorithmic beliefs lead to poorer expectancy fulfillment, thus increasing distress, whereas the appraisal path reduces distress. This remains an empirical question for future research.
Another future direction is exploring communication behaviors. The next outcome down the line in the sociocognitive mechanism, after appraisal, is people’s actual behaviors. The actualization of beliefs into reality is essentially through social behaviors that are consistent with what people believe to be true. 15 Therefore, testing communication outcomes (e.g., anticipated future interaction, self-disclosure) in experimental settings can be beneficial. 10
This study is not without limitations. The hypothetical scenario, though offering cleaner manipulation than more realistic approaches, has lower external validity than the more realistic methods. Also, our initial test used generic measures of appraisal. Future research can address these limitations by using lab experiment designs and employing more granular measures.
Despite these limitations, our study is the first to test cognitive outcomes of algorithmic beliefs and offers valuable insights into constructing a theory of algorithmic beliefs. Such a theory can represent a unique contribution of communication scholars to grasping social behaviors in complex algorithm/AI-governed social systems, and is very needed at this point, given the prevalent mental health problems associated with online dating in recent years.3,4,6 Future research can further explore other types of appraisals and the emotional and communication ramifications of these appraisals. The impact of the examined covariates (personality traits and technology usage) on the effect of algorithmic beliefs also represents a valuable future direction. Specifically, it can be valuable to explore how choice of platforms with different affordances (thus serving different dating goals) may moderate the effect of algorithmic beliefs. Pragmatically, our findings suggest that to reduce online dating burnout, platforms should do a better job at explaining how dating algorithms work to enhance algorithmic beliefs and adaptive cognitions in online dating.
Footnotes
Acknowledgment
The authors would like to thank Joseph Nickalo for offering insights on the examined scenarios.
Authors' Contributions
J.M.H.: Conceptualization, investigation, methodology, Formal analysis, writing—original draft. Y.J.O.: Conceptualization, methodology, writing—review & edit.
Data Availability
The data used in the current study are available from the corresponding author upon reasonable request.
Author Disclosure Statement
The authors have no conflicts of interest to disclose.
Funding Information
The study did not receive any financial support.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
