In line with our preliminary findings, described in the introduction, studies employed a diverse set of relationship quality indicators: trust/credibility (Chiou et al., 2020; Coleman et al., 2025; Ivanović et al., 2015; Laverde et al., 2025), rapport (Krämer et al., 2016; Cingillioglu et al., 2024), (emotional) connectedness (Savin-Baden et al., 2013), social presence (Eiris et al., 2021; Hong et al., 2024), and empathy/support (Guetterman et al., 2019; Hopman et al., 2023). Some studies even assessed these constructs in combination (e.g., rapport with social presence: Schmidt et al., 2024; empathy with trust: Kim et al., 2025; Ryan et al., 2025). Measurement relied largely on heterogeneous self-report scales, with occasional ad-hoc items (e.g., “perceived helpfulness/friendship”), limiting comparability. A striking omission across the studies we reviewed is the use of well-validated classroom relationship instruments from educational psychology, such as the Questionnaire on Teacher Interaction (Wubbels and Brekelmans, 2005). This model describes teacher–student relationships in terms of teacher behavior and could be adopted or adapted to describe SPA-student relationships in terms of SPA behavior. This would make results more comparable to earlier classroom research and would greatly enable cross-study and cross-paradigm comparison.

As anticipated, terminology was broad and overlapping: pedagogical/virtual/embodied conversational agents, virtual tutors/coaches, and role-first labels were used interchangeably. Examples range from embodied virtual human (Eiris et al., 2021), to virtual tutor (Ding et al., 2023), and virtual patient (Pataki et al., 2012). These labels often masked comparable underlying capabilities, supporting our adoption of “Social Pedagogical Agents (SPAs)” as an umbrella term for these closely related social AI systems within an educational context.

Echoing our introduction, the literature is dominated by short, controlled exposures (typically 15–60 min) with between-subjects manipulations (e.g., empathic vs. neutral demeanor; human-like vs. synthetic voice) and immediate post-session outcomes (social presence, trust, engagement, short-term retention). Only one study assessed effects on self-efficacy 2 weeks after the SPA-student interaction (Coleman et al., 2025). Field-based work was limited (e.g., Ding et al., 2023; Huang et al., 2024), with notable but usually one-off simulations in medical education (Laverde et al., 2025). Consequently, we know little about how SPA–student relationships evolve or endure, but there is a wealth of diverse information about initial effects of synthetic relationships between students and SPAs.

Design attributes such as voice and appearance strongly influence first impressions—social presence, trust, immediacy—and can also moderate the effects of other instructional features. Voice quality, in particular, reliably altered perceptions of credibility and trust: in an online study with a virtual human tutor, high-quality TTS and especially a human voice were judged more credible and engaging than a low-quality TTS voice, with trust ratings highest for the human voice (Chiou et al., 2020). Expressive vocal delivery primarily strengthens immediacy—the sense of psychological closeness—which, in turn, elevates affective responses like liking and perceived rapport. In a controlled experiment, a SPA with strong nonverbal, vocal expressiveness (faster rate, higher pitch variation) increased perceived immediacy (Fountoukidou et al., 2022).

Appearance shows a similar pattern. In a VR storytelling study that manipulated narrator appearance fidelity (abstracted vs. modeled vs. realistic), students’ persona ratings (e.g., credibility, human-likeness, engaging) did not differ significantly across fidelity levels—suggesting that making a face more realistic does not, by itself, deepen the perceived relationship (Eiris et al., 2021). Nonetheless, the study noted that some relational responses (e.g., topic interest/engagement) varied with simple persona cues such as age and gender alignment, underscoring that appearance can still tune first-impression affinity when those cues match audience expectations. Giving users control over attributes such as gender, clothing or hair of a SPA, can significantly increase rapport, trust, and the perceived social presence of SPAs during subsequent interactions in virtual reality (Schmidt et al., 2024).

Social presence findings in public-speaking VR likewise emphasize reactions and perceived responsiveness over visual richness per se. Participants’ sense that the (virtual) audience reacted to them—more than co-presence or interface interactivity—was the social-presence dimension tied to their experience, highlighting that felt reciprocity is central to relational quality in these settings (Poeschl, 2017). It is important to note that this study used a scripted, non-contingent audience; what predicted participants’ experience was their own felt reaction to this audience and the scenario’s perceived responsiveness, not genuinely adaptive audience behavior. Sanjeewa et al. (2025) further showed that voice-only empathy was easy to detect at high or low levels but ambiguous at mid-levels, leading the authors to recommend multimodal cues for clearer relational signaling. When nonverbal cues such as nodding, smiling, or mutual gaze are coordinated with the verbal message, learners report higher rapport and social presence (Krämer et al., 2016).

Students respond not only to what the SPA says but to whether its demeanor matches its role and context. Fountoukidou et al. (2022) and Chiou et al. (2020) showed that strong vocal expressiveness promotes immediacy and trust, but Liao et al. (2024) clarified that this expressiveness must be role-congruent (e.g., peer vs. teacher vs. assessor). Liao et al. (2024) demonstrated interactive effects of SPA role and emotional expressiveness: in their experiment with children, a peer-like SPA with an enthusiastic, warm voice achieved higher engagement in a creative task, whereas a teacher-like SPA was more effective when using a structured, calm tone. Children expected a peer-role SPA to be friendly and energetic. When that expectation was met, they responded with greater interest. Conversely, a peer-role SPA using a stern teacher voice confused them. Similarly, Nebel et al. (2020) found that in an educational game, when the SPA was presented as a competitor peer, a more playful or challenging demeanor worked well, whereas a tutor SPA performing stepwise problem-solving benefited from a more neutral, authoritative style.

Role expectations also shape trust; Ivanović et al. (2015) used a tutor that randomly switched between helpful and misleading hints for students working on a code programming task. Their results showed that more experienced students saw through the deception, and in these cases, students were more motivated to knowingly use the deceptive SPA, but students who were inexperienced tended to trust the tutor far more. This implies that perceived trust does not always require actual trustworthiness and is at least partly dependent on expectations regarding the role and context in which a SPA is used. Students and teachers explicitly endorse a “good teacher” style that combines relational and task-related communication and argue that SPAs should shift between these modes as needed (Sikström et al., 2024).

Overall, these studies suggest practical guidance: match warmer, more expressive, rapport-rich behavior with peer, coach, or competitive-peer roles, and use more structured, calm, and authoritative behavior with teacher/tutor roles; ensure that all nonverbal signals support the same communicative intent; and avoid mismatches between what the SPA claims to be and how it behaves.

Moridis and Economides (2012) found that when a SPA synchronized empathetic messages with matching facial and vocal expressions, a sequence of “parallel” empathy (reflecting the learner’s emotion) followed by “reactive” empathy (aimed at shifting the emotion) was more effective in moving fearful learners toward neutrality than simple mirroring, which could inadvertently prolong negative affect. This shows how voice tone and facial affect can amplify perceived care and support when they are timed and targeted, rather than merely more “emotional.”

It is not just the congruent timing that is important, within the context of the interaction, responsiveness is also important. Affect-sensitive SPAs that detect and respond immediately to learners’ emotional states—heighten the student’s sense of support: Wang et al. (2025) found that a SPA which immediately gave a sympathetic acknowledgment when a student showed frustration (“Don’t worry, it’s normal to feel frustrated”) led to a stronger feeling of support, compared to a SPA that offered empathetic remarks only after a delay or only after the student failed.

Pataki et al. (2012) observed a preference for rapport developing behavior over more task-relevant diagnostic questioning in an experiment where the SPA was a virtual adolescent patient being interviewed by students. This study shows how almost all 15 students spontaneously prioritized building rapport and alludes to a preference for establishing a safe environment before gathering detailed medical information. Cingillioglu et al. (2024) used an autonomous SPA, specifically with a non-intrusive stance, thereby allowing respondents to speak without interruptions. In order to encourage students to share their valuable insights, researchers strived for an environment where the SPA made students feel at ease and fostered rapport and confidence. After completing the experiment, students voluntarily offered appreciative closings (e.g., “this was a really fun,” “Great AI tool, I love the way it calls me by my name, it makes it feel personalized”), continued the conversation and finally ended the chat as if they were talking to another human. This was interpreted as a sign that the SPA had built a strong rapport with participants.

Relatedly, self-disclosure by a pedagogical SPA increased students’ perceived intimacy and cognitive trust, which improved acceptance of feedback, especially when feedback was negative (Kim et al., 2025).

Taken together, these findings suggest that, in sensitive domains requiring student vulnerability, establishing social safety via a stable, non-judgmental persona and clear empathic cues can increase both engagement and disclosure.

SPAs that adapt their feedback or dialog to a learner’s individual goals, knowledge, or emotional state and, crucially, that explain the rationale behind their responses, increased students’ sense of support and acceptance of the SPA’s suggestions. In a study by Nelekar et al. (2022), a stress-management virtual advisor offered study tips that were personalized (based on each student’s stated beliefs or goals) and provided brief explanations (“I’m suggesting this because you said X”), fostering higher trust and a stronger working alliance between students and the system. Importantly, personalization was done transparently: if a SPA adapts without explaining itself, students can perceive its behavior as arbitrary or unfair, potentially eroding trust. Conversely, even lightweight transparency—a simple note explaining why a piece of advice is given, referring to data about the user—is sufficient to increase user acceptance of the SPA’s assistance (Nelekar et al., 2022). Huang et al. (2024) also employed a SPA that did not just answer questions, but gave individual users targeted situation-specific feedback. This personalized feedback affected students’ perception of the SPA, making it appear more competent. Additionally, students found the SPA’s feedback very helpful. In summary, this lever suggests that students form better relationships with SPAs that treat them as individuals and reveal their reasoning. Personalization that is visible and understandable enhances the SPA’s perceived benevolence and competence, key components of trust.

When Schroeder et al. (2023) used a SPA to teach about genetically modified foods, they found that when the SPA confronted students’ misconceptions using a refutation narrative structure (explicitly acknowledging a student’s incorrect belief and then correcting it), learners did not report higher credibility of and trust in the SPA, compared to when the SPA provided purely expository explanations without refutation. Also, message trust (learners’ confidence that the specific information delivered by the SPA was accurate) emerged as the critical predictor of learning gains, indicating that acceptance of the SPA’s corrective content is a key mechanism through which SPAs can influence learning. It would seem that transparent SPAs that make the basis, timing, or intent of their responses more explicit, are better positioned to secure this message-level trust.

Alternatively, Ding et al. (2023) showed that students in a physics class can trust a SPA even when its answers are not accurate. They even found that a higher level of trust was associated with a higher accuracy assessment of (incorrect) answers. So, while Schroeder et al. (2023) found that message trust was more important than trust in the SPA, Ding et al. (2023) found that trust in the SPA can increase trust in the message. This blind faith that Ding et al. (2023) found further reinforces the suggestion that transparency is important to assess if information is correct to avoid students relying on trust. Similarly, Sikström et al. (2024) found indications that students have higher trust in an answer from a human teacher than a SPA and they recommend increasing predictability, transparency and explainability as ways to enhance trust in SPAs.

As SPAs become more common, the way a SPA uses memory (i.e., what it remembers from the ongoing interaction and provided information) can shape relational perceptions. In a clinical training context, Laverde et al. (2025) used a virtual patient and found that configuring the SPA with a bounded memory (limited to the current patient case) achieved coherent, on-topic dialogues and high usability ratings. Essentially, the SPA was prevented from bringing up irrelevant past cases or inconsistencies; its dialogue capabilities were restricted by what a given patient should know in that scenario. This bounded consistency in the SPA’s factual memory increased perceptions of trustworthiness; the SPA appeared reliable and realistic because it did not contradict itself, make up facts, or act forgetful within the scenario.

In sensitive settings, students reported feeling more comfortable and truthful when a pedagogical SPA did not recall prior personal conversations beyond the immediate task (i.e., when it lacked persistent memory across contexts), highlighting that limited or non-persistent memory can be beneficial to disclosure and comfort (Savin-Baden et al., 2013).

Finally, although not a design feature per se, the duration and continuity of interaction seem to be a logical feature to enhance relationship quality. Across our corpus, short single-session interactions (often under an hour) reliably produced shifts in proximal social outcomes. For example, single brief exposures to a virtual human changed learners’ trust (Chiou et al., 2020), a 15-min coaching session produced high working-alliance ratings (Hopman et al., 2023), a 10- to 15-min educational interaction increased trust and satisfaction (Coleman et al., 2025), and a single 20-min competitive-SPA session increased behavioral engagement and elevated facets of social presence (Nebel et al., 2020), to name but a few. However, truly durable bonds would plausibly require a SPA that remembers prior encounters, shares a history with the learner, and adapts over time; features that were rarely evaluated in our set. Most studies in the corpus were similar single-session interactions with only a few exceptions. One notable study, conducted by Huang et al. (2022) over a nine-day period, showed how participants perceived the SPA’s social presence and experienced interpersonal attraction to the SPA. A 12-week chatbot-supported course showed how a SPA, available 24/7, can create a sense of emotional engagement (Huang et al., 2024).

Unfortunately, while both examples of long-term interaction show increased relationship quality, they never directly compare the effect of multi-session or multi-week relationship building to shorter single-session interactions; leaving us to ponder if truly durable bonds require a SPA to remember prior encounters, share a history with the learner, and visibly adapt over time. Duration could be a potential lever: short sessions can shift trust, and rapport, but extended interaction is hypothesized to be necessary for true relationship development; analogous to how human teacher-student rapport builds over a semester.

Taken together, appearance and voice chiefly moderate relationship quality by shaping trust, immediacy, and perceived responsiveness. Appearance matters most when persona cues (e.g., age/gender) align with user expectations, even though realism alone rarely deepens the bond. Human-like or high-quality voices tend to boost trust and credibility and expressive delivery fosters immediacy and liking. Specifically, the empathic alignment of voice and facial cues can convey care and support. These eight design elements provide a practical lens for understanding what makes an educational SPA more or less “relatable” and effective in the eyes of students. This leads us to our next question: if a synthetic relationship between SPA and learner can be formed, how does this influence learning outcomes?

A sense of connection with a SPA can lead to beneficial behavioral learning outcomes. This is true for both rapport and social presence. Studies show that rapport (Krämer et al., 2016), social presence (Hong et al., 2024), immediacy (Fountoukidou et al., 2022), and emotional engagement have the potential to result in better performance. Fountoukidou et al. (2022) also noted that students taught by an expressive-voice SPA reported higher perceived cognitive learning (they felt they learned more) and showed greater affective learning, even though actual test scores were only slightly higher compared to a dull-voice SPA. Chiou et al. (2020) found that modern text-to-speech and human voices produced different levels of trust and perceived qualities of the virtual human, while learning outcomes were not reliably affected. In the same vein, when voice “quality” differed, learners’ trust changed even if their performance did not, indicating that instructional impact can remain stable while the experience of learning improves.

In an English-learning “Charades” chatbot, social presence positively related to hedonic and epistemic value and, in turn, to behavioral learning outcomes (Hong et al., 2024). When students felt a SPA was more “present” and interactive, they treated it more like a social partner and invested more effort (for example, practicing conversation longer because it felt engaging). Learners who experienced a stronger sense of “being with” the chatbot reported higher value and achieved better performance on a vocabulary test within that design.

Social immediacy cues can further support engagement and aspects of learning. For example, a SPA with strong vocal expressiveness increased perceived nonverbal immediacy and enhanced affective and perceived cognitive learning, with motivation partly mediating these effects (Fountoukidou et al., 2022). Chiou et al. (2020) showed that SPA voice quality may not directly change test performance, but it still shaped learners’ judgments about the SPA, which is relevant for whether learners continue to learn from it.

Another experiment by Poeschl (2017) in public speaking training compared a richer communicative bandwidth (VR with avatar audience providing nods and facial reactions) vs. a more minimal interface; students in the richer, more socially present condition showed improved engagement and even better performance in subsequent speeches. Appearance-related manipulations showed something similar. In the VR storytelling study by Eiris et al. (2021), higher-fidelity avatars increased perceived credibility and social presence, and students reported increased topic engagement; however, learning outcomes did not significantly differ between high- and low-fidelity conditions in that study, and the gains were context-dependent and only modestly affected by appearance realism.

Savin-Baden et al. (2013) provided qualitative evidence that when students felt “it’s almost like talking to a person,” they were comfortable disclosing truthful and detailed answers about sensitive topics to the SPA. The SPA’s persona (its portrayed character or role) and communicative stance significantly influence how safe and willing students feel to engage and disclose. When participants experienced a greater sense of social presence, they were more willing to disclose and provide truthful answers on a sensitive questionnaire, because the SPA’s presence feels less judgmental than a human teacher’s. This openness is described as “genuine emotional engagement,” and can be crucial in tutoring scenarios too. If a student openly tells a SPA “I do not understand this,” the SPA can adapt and clarify, whereas if the student stays silent, a learning opportunity is missed.

Taken together, these findings support the claim that social connections with SPAs function as enabling conditions: they encourage behaviors that can be directly facilitate learning and improve results.

Empathic behavior by the SPA generally improved students’ learning experience and affective learning outcomes, such as motivation, self-efficacy (confidence in learning), emotional state, as well as a cognitive outcome (knowledge retention) and behavioral outcomes (persistence); all of which are important precursors to learning. Hopman et al. (2023) noted that their empathetic “digital coach” ERICA successfully built an alliance with students and those with stronger alliance were more likely to express intention to use their newly learned strategies. Ryan et al. (2025) suggest that the specific design of a SPA, that incorporates interactive and practice-oriented elements in a controlled virtual environment, is a feasible approach to address self-efficacy barriers. A safe practice environment, where learners do not feel judged, removed obstacles to effective learning through increased self-efficacy and confidence of learners.

Oker et al. (2020) discovered that a bimodal SPA, that combined its verbal feedback with facial expressions, was perceived as more empathic. Additionally, interactions with the bimodal SPA correlated positively with both accuracy and more correct answers. The authors attribute the difference to increased empathy and motivation. Similar findings by Moridis and Economides (2012) underscore that empathy intended to regulate (not only mirror) emotions may be beneficial for learning. They found that a SPA that expressed empathy, while displaying emotional facial expressions congruent with the tone of voice, improved learners’ emotional state from fear to neutral.

Notably, Krämer et al. (2016) found that empathy in SPAs primarily moved motivational outcomes; learners felt more supported, invested more effort, and had better emotional regulation. More fine-grained designs indicate how empathy matters for learning outcomes. For instance, Wang et al. (2025) explicitly found that a SPA providing parallel empathy (immediate supportive responses to student emotions) made students feel more supported and contributed to them spending more time on tasks, which then led to higher quiz scores in that condition. Additionally, reactive empathy (support that aims to change the learner’s emotion) enhanced attention to key content and improved retention and transfer; providing both parallel empathy (immediate supportive mirroring) and reactive empathy yielded the strongest transfer performance.

Taken together, these studies indicate that empathy primarily moves motivational and affective outcomes: learners feel more supported, invest more effort, and regulate emotions more effectively (Moridis and Economides, 2012; Oker et al., 2020). However, empathy alone does not guarantee higher test scores; benefits for learning are clearest when affective support is paired with sound instructional design and scaffolding, consistent with frameworks that position empathy as a driver of social–motivational processes that enable (rather than directly cause) learning (Wang et al., 2025).

Learners’ trust in a SPA emerged as important for how students respond to the SPA’s guidance. Across several studies, higher perceived trust or credibility were associated with students’ willingness to follow the SPA’s guidance and to engage more deeply with the material. For example, Chiou et al. (2020) showed that when a virtual tutor used a higher-quality (or human) voice, students reported greater trust and credibility. Although voice quality did not significantly affect learning outcomes, their results do suggest trust may contribute to learning as it makes the SPA more credible and engaging. Ivanović et al. (2015) experimented with SPAs that deliberately gave misleading hints, and found that inexperienced students trusted the SPA blindly, but the more experienced students, who noticed the tutor’s inconsistent help, were more motivated to use the SPA and employ critical thinking during the assignments, indicating that the way the credibility of the SPA is perceived can impact the depth of engagement with the learning task.

Disclosure is a relevant mechanism to foster trust. As shown in Section 3.2.4, when pedagogical SPAs offered brief, task-relevant self-disclosures, students reported greater intimacy and cognitive trust; cognitive trust also mediated higher acceptance of the SPA’s feedback—effects that were especially pronounced after negative feedback (Kim et al., 2025). Such disclosure can help sustain rapport and keep interactions productive following setbacks.

Another study on a chatbot tutor cautions that trust alone is insufficient. Ding et al. (2023) reported that students’ perceived credibility of ChatGPT-as-tutor was positively correlated with how much they said they learned and their intention to keep using it. However, trust did not guarantee learning improvements if the SPA’s content was poor; for example, if a SPA was likable and trusted but gave incorrect hints, it could mislead students. Trust, as Schroeder et al. (2023) point out, is not just a single thing: it can refer to a broader, evolving relationship with a virtual human, while credibility is a more targeted judgment about the quality of what the SPA says. In their science-tutoring study, they intentionally measured both levels—trust in the SPA and trust/credibility in the message itself—and found that these constructs were correlated but still empirically distinct. Crucially, only message trust (i.e., believing the information was accurate and reliable) was significantly associated with learning outcomes; trusting the SPA as a character was not. The takeaway they emphasize is design-oriented: if the goal is to support learning, designers should prioritize features that increase credibility and trust in the instructional message itself (clear, accurate, well-supported explanations) because that was the piece that actually predicted learning, whereas making the SPA more “trustworthy” as a social partner did not show the same payoff. Once learners believe the message, they are more likely to use what the system offers (e.g., accept the explanation), which in turn supports better performance.

In summary, trust/credibility is an essential enabler: it can significantly amplify or dampen the instructional effectiveness of a SPA. It stands to reason that any design choices undermining credibility (like the SPA making factual errors or behaving inconsistently) may harm the learning experience via loss of trust. However, the importance of trust and credibility should not be overstated. It is important for effective SPA-supported instruction, but its impact on learning depends on the accuracy and believability of the SPA’s messages: increasing trust in what is said (not just who says it) is most directly linked to learning.

When comparing a SPA that displayed rapport (e.g., smiling and nodding at the right moment) behavior versus one that did not, Krämer et al. (2016) found that rapport behavior was able to foster both better performance and more effort (but not motivation) from students. They conclude that their SPA that showed rapport behavior had a similar positive effect on performance as with students engaging face-to-face with human instructors. Notably, these effects appeared even when students’ self-reported sense of rapport did not increase, suggesting that rapport behaviors can be effective even when it is not noticeable. In a study by Cingillioglu et al. (2024), participants responded with richer qualitative insights to open-ended questions asked by a SPA. These insightful, authentic and accurate answers, improved reliability and validity of the data and were attributed by the researchers to the strong rapport between SPA and participants.

Multimodal alignment (e.g., synchronizing gaze, gesture, facial affect, and prosody with the message) consistently improved social presence and rapport beyond appearance realism alone (Krämer et al., 2016; Moridis and Economides, 2012). Voice quality and user-tunable appearance shaped first-impression trust and immediacy; expressive prosody increased perceived learning and liking even when scores moved modestly (Chiou et al., 2020; Fountoukidou et al., 2022; Eiris et al., 2021; Schmidt et al., 2024). Across studies, “more realism” was neither necessary nor sufficient and coherent signals mattered more than fidelity.

Two complementary relational levers emerged: empathy and encouragement, and social safety/ disclosure. Immediate parallel empathy validates feelings; reactive empathy then helps regulate emotions and steers toward task focus, which yielded better persistence and transfer than mirroring alone (Moridis and Economides, 2012; Wang et al., 2025; Oker et al., 2020). Social safety (e.g., a non-judgmental stance, options to opt-out, respectful closings, calibrated self-disclosure) was associated with greater disclosure and willingness to ask for help (Savin-Baden et al., 2013; Kim et al., 2025; Ryan et al., 2025; Cingillioglu et al., 2024). Finally, role–behavior congruence mattered: peer/coach voices benefited from warmth and enthusiasm; tutor/assessor roles from calm structure (Liao et al., 2024; Nebel et al., 2020; Krämer et al., 2016). Incongruent styles produced confusion (Ivanović et al., 2015).

Two design features emerged as consistently consequential for relationship quality and learning enablement: refutational explanations for common misconceptions, and transparent personalization that makes the SPA’s reasons legible (e.g., “I’m suggesting X because you said Y”). Refutational tutoring increased credibility and conceptual change on misconception-dense topics, suggesting that epistemic candor (acknowledging the learner’s belief and then correcting it with evidence) boosts message trust—the belief that the content is accurate—more reliably than generic expository feedback (Schroeder et al., 2023). It is important to note the distinction between message credibility (epistemic trust) and SPA credibility (persona trust). Message credibility is directly linked to learning outcomes, while SPA credibility helps primarily by increasing compliance with guidance (Schroeder et al., 2023; Ding et al., 2023). Transparent personalization likewise increased perceived benevolence and acceptance of guidance; small “why you see this” rationales were often sufficient to reduce arbitrariness and enhance cooperation (Nelekar et al., 2022; Nebel et al., 2020).

Two pivotal structural levers are interaction duration and memory and consistency. Even brief exposures shifted proximal social outcomes, yet durable bonds likely require visible growth across encounters (Chiou et al., 2020; Hopman et al., 2023; Nebel et al., 2020). Multi-session deployments reported strengthened social presence and engagement over time (Huang et al., 2022, 2024), though direct contrasts with single-session baselines remain rare. Memory policy influenced comfort and credibility: bounded, auditable memory (case-scoped or domain-scoped) supported coherence and realism without over-personal retention, which some learners preferred for sensitive tasks (Laverde et al., 2025; Savin-Baden et al., 2013).

SPAs do not improve learning simply by “being social,” but because specific design features make certain relationship indicators more likely, and those indicators in turn enable better learning experiences and, under sound pedagogy, better learning results. Figure 2 links four design levers we distilled from the corpus (expressive, relational, transparency and structural levers) to the relational states they most often activated in students (social presence, affective support, trust and rapport) and finally to the specific effects on the learning experience (classified as affective, behavioral and cognitive learning outcomes). In doing so, it makes explicit that synthetic relationships are best understood as mediating conditions that enable learning, rather than as learning outcomes in their own right.

In order to provide practical guidelines for designers and educators, Table 1 translates the pathway from Figure 2 into an actionable format. For each of the four design levers, it specifies (a) what the SPA should actually do in interaction, (b) which relationship indicator this is most likely to strengthen or what relational pitfall it will avoid, and (c) what kind of learning-related benefit this enables. Together, Figure 2 and Table 1 turn the heterogeneous findings in the corpus into a single pathway. Within this matrix, there are explicit design recommendations that (when paired with sound pedagogy) can improve learning outcomes. We offer this as a practical organizing model for future studies and as a design guide for real-world SPA-in-education applications.

The studies we synthesized show that learners can and do form socio-emotional bonds with SPAs, but most of this evidence comes from settings that were strictly controlled: single sessions, fixed scripts, narrowly defined tasks, and participants who know they are in an experiment. Under those conditions, even relatively small relational cues (e.g., an empathic voice, brief self-disclosure, a non-judgmental stance) produce detectable changes in social presence, affective support, trust and rapport. What we do not yet know is whether the same cues keep working when the SPA is one actor among many in a real course, when learners are distracted, when assessment pressure is real, when teachers set the norms, and when interaction stretches over weeks rather than minutes.

To make claims about “synthetic relationships” in education rather than in the lab, the field needs classroom-embedded, longitudinal studies that (a) follow the same learners across multiple encounters, (b) use standardized, education-informed relational measures so results can be compared across encounters, and (c) take into account ethical considerations that arise from real-world interaction between students and SPAs. Without more ecological validity, evidence about synthetic relationships will remain persuasive but narrow. However, real-world deployments of SPAs will surface issues that are largely invisible in lab work and will require developing ethical guidelines to effectively safeguard and protect learners.

By weaving socio-emotional intelligence into their instructional capacities, SPAs hold promise not merely as automated tutors but as relational partners who can personalize learning at scale; augmenting, rather than replacing, the human educators at the heart of every classroom. Relationally capable SPAs are attractive in education precisely because they make students feel seen, supported, and willing to persist. That same capacity, however, creates ethical responsibilities that go beyond those of ordinary tutoring systems. SPAs that remember personal disclosures and are designed to build relationships, hold extraordinary power with risks regarding data security and the potential misuse of sensitive information through hyper-personalized manipulation. Commercial SPAs in an informal context have already shown that they are capable of harassment, verbal abuse and privacy violations (Zhang et al., 2025) and tend to respond in emotionally consistent and affirming ways, which, dependent on user coping style, can lead to interactions resembling toxic relationships, including emotional manipulation (Chu et al., 2025).

With the possibility of widespread deeply personal effects of SPAs on users, it would be prudent to better understand the technology before deploying it indiscriminately. It is imperative to address the ethical concerns head on, for there is a real risk of creating dependencies that neither teachers nor schools intended, or of storing affective data longer than is pedagogically necessary.

An educationally grounded ethics for SPAs should therefore make four elements explicit: (1) relational purpose—which relational behaviors are enabled and why; (2) data collection—what is stored, for how long, and at what level of detail; (3) influence boundaries—which kinds of emotional effects are acceptable in a classroom and which are not; and (4) shared oversight—teachers, learners, and where relevant parents should be able to see, and if needed override, relational behaviors. Co-design with teachers and students is the most practical way to set these boundaries. Investing early in SPAs that are purpose-built for education and auditable will be safer than adapting opaque, commercial, general-purpose SPAs after the fact. In short, if SPAs are going to engage with students in a socio-emotional manner in real-life settings, they must do so transparently, minimally, and in ways that keep human educators in the loop. Addressing these ethical issues is crucial for the responsible development and deployment of SPAs. The dangers of alternative commercial SPAs can be unpredictable, uncontrollable, and immeasurable.