Instructional modality influences neurocognitive engagement during moral learning Authors: Sakshi Chauhan, Arnav Bhavsar, Varun Dutt Published: 2025 ABSTRACT AbstractAI-mediated learning environments demand more subtle comprehension of how instructional modalities impact not only learning outcomes, but cognitive and affective processes supporting them. Moral education is a common denominator of integral development, yet we do not know how neurocognitive engagement varies with new versus old media. This experiment compared moral learning conveyed by a GPT-mediated social robot (Picoh) with that through traditional text-based reading, with an Indian philosophical fable. Fifty students were randomly assigned to a modality and completed a comprehension quiz while their neural activity was recorded with Electroencephalography (EEG). Two indices were compared: frontal alpha asymmetry (FAA), an affective–motivational engagement index, and gamma coherence, a cognitive integration index. No group differences were found in quiz performance or EEG measures. However, in the robot group, FAA showed a negative trend with comprehension, which may point toward a potential link between comprehension and left-frontal activation (approach motivation). These findings illustrate that neurophysiological measures can disentangle subtle, individual-level engagement effects undetectable to behavioral outcomes alone. More broadly, the findings explored the value of combining EEG with behavioral data in revealing mechanistic comprehension of how instructional modalities impact learning.Subject terms: Neuroscience, Psychology, Psychology Introduction In an era of increasing complexity, digital interdependence, and moral complexity, moral reasoning is reasserting its place as an essential skill in learning1. In college students, the ability to think critically about moral dilemmas is a central part of social responsibility learning, ethical decision-making, and civic engagement1–3. While the value of moral education is firmly established, concerns persist about how to craft instruction that is engaging and can generalize to real-life situations. In the broad field of learning science, one of the core issues is to identify how modes of instruction interactive versus textual, for example shape both what is learned and the cognitive–affective processes underlying it4,5. Previous research suggests that the “transfer problem” learners’ inability to use knowledge flexibly across situations can be as much a product of delivery mode as content per se6,7. Theoretical models like Kohlberg’s stages of moral development8 and Vygotsky’s social constructivism9 highlight the richness of dialogical scaffolding, social setting, and increasingly sophisticated reasoning tasks in shaping higher-order learning. Drawing on these, the research considered here explores how embodied, interactive modalities differ from text-based instruction in supporting moral reasoning. Building on prior theories of moral development and learning, this study integrates perspectives from Kohlberg’s developmental stages (emphasizing hierarchical reasoning), Vygotsky’s social constructivism (emphasizing dialogical scaffolding), and Bandura’s social cognitive theory (emphasizing observation and affective modeling). Together, these frameworks converge on the idea that moral learning depends on both cognitive reasoning and affective–motivational engagement shaped by social context8–10. Contemporary moral cognition models, such as Haidt’s social intuitionist framework and Greene’s dual-process theory, further suggest that moral understanding emerges from the interaction between automatic emotional intuition and controlled cognitive reasoning. Integrating these perspectives, the current study proposes that instructional modality (robotic vs. textual) may differentially activate these systems—embodied, socially interactive contexts engaging emotional–motivational circuits, and text-based contexts engaging reflective–cognitive processing11. Technological advances have expanded educational delivery modes. Social robots with artificial intelligence have been promising tools to enhance interactivity and emotional feedback. Social robots as embodied communication agents simulate conversation, provide verbal and visual information, and enable multimodal interaction. Empirical evidence shows that physical embodiment and anthropomorphism can enable greater enjoyment, motivation, and performance of students in socio-affective learning12–14. From an embodied cognition perspective, such modalities can provide richer affective and sensorimotor responsiveness than static text, with the capacity to enable deeper encoding of abstract information15. Such potential is still untapped, however, as research has predominantly been restricted to early childhood or science, technology, engineering, and mathematics education, with their potential for moral education still to be tapped16. While text-based instruction remains the dominant mode of instruction of ethical and philosophical content, its strengths supporting reflective consideration, slow processing, and abstract thought are especially compatible with the requirements of ethical thinking. Reading has been shown to support the development of empathy, prosocial behavior, and moral judgment17,18. Neurocognitive science suggests the role of reading in the reorganization of perceptual and linguistic representations, which involves deep learning19. It is unknown whether reading plays a role in a distinctive quality of engagement compared with embodied, interactive approaches. In order to quantify such differences, it is not sufficient to employ behavioral measures (e.g., questionnaires or tests). Neurocognitive measures yield mechanistic insights into how learners process information in real time. Building on theoretical foundations of moral learning—Kohlberg’s developmental hierarchy of reasoning, Vygotsky’s social constructivism emphasizing dialogical scaffolding, and Bandura’s social cognitive theory highlighting observational and affective modeling—the present study examines how cognitive and affective processes jointly shape moral engagement. Two Electroencephalography (EEG) measures operationalize these theoretical dimensions: frontal alpha asymmetry (FAA) reflects affective–motivational engagement (approach–avoidance tendencies and affective modeling), whereas gamma coherence reflects cognitive–integrative reasoning (hierarchical and socially mediated processing)20,21. These indices allow us to ask not only what learners remember, but how they think and feel during moral comprehension across modalities. The current research probes moral understanding with a philosophical lesson from the Bhagavad Gita, a significant Indian Knowledge Systems (IKS) text. Participants either interacted with a social robot facilitated by GPT (Picoh) in a semi-interactive conversation or read the identical content in the traditional manner. Through placing the identical content in two modalities, we control the instructional delivery factor in informing behavioral responses and brain activity. This study is mechanistic and exploratory. Rather than testing directional hypotheses, it will test whether small differences in affective and cognitive engagement emerge on EEG even when behavioral performance is the same. Accordingly, the present study is guided by a causal model linking instructional modality, affective–motivational engagement, and moral learning outcomes. Specifically, we hypothesize that Instructional Modality influences FAA (indexing approach/withdrawal motivation) that in-turn influences Moral Understanding and Transfer. Gamma coherence, representing integrative cognitive processing, is examined as a complementary index of attentional engagement. This framework extends dual-process and social-intuitionist models by introducing neurophysiological engagement markers as mediators linking modality and moral comprehension, thereby bridging traditional moral theory with empirical educational neuroscience. Particularly, we ask: Particularly, we ask: (RQ1) Are robot-mediated and text modalities equally or otherwise correlated with moral understanding and transfer? (RQ2) Do FAA and gamma coherence differentiate affective and cognitive engagement by modality? (RQ3) Do these neural markers predict comprehension outcomes, offering more sensitive measures of engagement than behavioral measures alone? More generally, this study adds to the learning science by showing how modality and neurocognitive engagement in combination affect learning in abstract topics such as ethics, civic reasoning, and critical thinking. Methods Design The objective of this study was to investigate how instructional modality robot-mediated interaction versus traditional text-based reading shapes both behavioral learning outcomes and neurocognitive engagement during moral education. A total of 50 participants were randomly assigned to two groups (n = 25 per group). In the Picoh group, participants engaged with a GPT-powered social robot programmed to deliver a Socratic-style dialogue on a moral narrative and respond to predefined questions12,15,22–24. In the Reading group, participants studied the same questions and answers in paper form, ensuring content equivalence but without interactivity17,19. Dependent measures included behavioral outcomes (transfer performance and knowledge from quizzes) and neurocognitive indices from EEG recordings: FAA as a neurocognitive marker of affective–motivational engagement, and gamma coherence as a neurocognitive marker of cognitive integration20,21,25,26. As can be observed from Fig. 1, there were three stages to the experiment. During the pre-intervention stage, participants completed demographic questionnaires and baseline EEG was taken. During the intervention stage, participants completed the assigned modality for a set period of seven minutes, while EEG was continuously recorded27,28. During the post-intervention stage, all participants completed a multiple-choice comprehension test. Figure 1 illustrates this between-subjects design, denoting the two instructional modalities and the time order of phases. Overview of the experimental design. The study consisted of three phases: (1) a pre-intervention stage involving consent, demographic assessment, and baseline EEG recording; (2) an intervention stage in which participants were randomly assigned to either the Picoh social robot group (n = 25) or the Reading group (n = 25) for a 7-minute learning task with continuous EEG recording; and (3) a post-intervention stage involving a comprehension quiz. The figure illustrates the between-subjects structure and the sequence of study procedures. Participants Ethical approval For the study was granted by the Institutional Ethics Committee of the Indian Institute of Technology Mandi, India (IITM/IEC(H)/2025/VD/P2). Written informed consent was also taken from all the participants prior to their participation. 50 volunteers were recruited and assigned randomly in equal numbers to the Picoh (n = 25) and Reading (n = 25) groups. Participants were recruited through university-wide announcements and voluntary sign-ups circulated among undergraduate and postgraduate students. Inclusion criteria required participants to be between 18 and 30 years of age, had taken at least one course in introductory English literature, and free from any history of neurological, psychiatric, or visual impairments that could interfere with EEG data collection or task comprehension. Participants were excluded if they reported sleep deprivation, medication use affecting cognition, or prior familiarity with the experimental content (Bhagavad Gita narratives). The medium of instruction at the university was English language. All participants provided written informed consent and were compensated with a small honorarium (INR 40, ≈ USD 0.52) for their time. A power analysis using G*Power 3.1 indicated that the achieved sample (n = 50; 25 per group) provided 0.80 power to detect medium effects at = 0.0529. In the Picoh group, the mean age was 20.52 years (SD = 1.34, Median = 20), and comprised 76% male and 24% female participants. In the Reading group, the mean age was 21.05 years (SD = 2.01, Median = 21), and comprised 88% male and 12% female participants. Participants were from different areas of study, e.g., humanities, sciences, and engineering. The experiment used a single-blind design: participants were unaware of study hypotheses, but the experimenter was aware of group assignment when preparing EEG. Voluntary participation was encouraged, and all the participants were given a token of appreciation (INR 40, ≈ USD 0.52) for their time. Stimuli The learning stimulus was derived from the Bhagavad Gita, an ancient Indian philosophical text on justice, duty (dharma), and moral responsibility30. Both groups received the same semantic material, with instructional modality being the only difference. The Picoh group learned the material through human-robot social interaction, and the Reading group learned the identical material in paper format. This enabled us to control for delivery mode effects on moral learning and engagement. Picoh group As illustrated in Fig. 2a, Picoh participants had semi-interactive dialogue with a social robot. Picoh showed philosophical content translated from the Bhagavad Gita and answered participant questions. While generated by OpenAI’s GPT-3.5 model31, interaction was controlled: participants selected from a list of pre-defined predetermined questions (Fig. 2b), and Picoh answered aloud in real time. This allowed content equivalence among participants while still mimicking a Socratic-style dialogue. The robot’s expressive face, speech, and conversational rhythm provided social presence, intended to attract attention, invoke reflection, and clarify abstract philosophical ideas12,15,24,32,33. (a) The Picoh social robot used in the robot-mediated learning condition, programmed to deliver a Socratic-style dialogue on a moral narrative derived from the Bhagavad Gita. (b) The set of moral reasoning questions presented to participants during the interaction, displayed on a screen interface that allowed them to select predefined prompts to ask the Picoh robot. Knowledge integration into Picoh The Picoh robot was created with the aim of integrating Indian Knowledge Systems (IKS) with contemporary AI-facilitated interactivity. The philosophical queries were indexed in GPTSimpleVectorIndex with GPT-3.5 via LangChain and LlamaIndex31–35, and standardized, context-suited responses were made possible. The verbal address of the robot supplemented with emotive facial expressions provided a multimodal learning scenario that integrated conventional ethical concepts with contemporary learners. This integration depicts how embodied AI platforms can be capable of introducing abstract philosophical knowledge to tangible, socially relevant scenarios12,15,24. Reading group In the Reading group, the same questions and answers were read by participants from paper. In contrast to robot-mediated interaction, this mode was static and non-interactive, as was the traditional text-based learning. For comparability, the same content was presented to each group, and the same seven-item multiple-choice test was given to all participants. Q1–Q6 were literal translations of intervention material, and Q7 tested transfer, where participants were asked to apply learned principles to a new situation. In particular, each quiz item was constructed with a specific cognitive or moral purpose in mind (see Table 1). For example, Q1 assessed detachment (stress reduction through understanding of the immortality of the soul), Q2 assessed schema-based moral reasoning, Q3 induced cognitive decentering in resolving conflict, and Q4–Q6 assessed higher-order and motivational reasoning. Q7, the transfer item, assessed principle-based reasoning for dharma. Overall, this organization made sure that the quiz not simply assessed surface recall, but also affective and cognitive processes of moral reasoning, in keeping with the study’s focus on neurocognitive engagement. Table 1Comprehension quiz Questions, Options, and correct Answers.No.QuestionOptionsCorrect AnswerRationale1Q1. What does Krishna explain about the nature of the soul?a) The soul is mortal and dies with the body.b) The soul is eternal and simply changes bodies.c)The soul is fragmented and dispersed after death.d) The soul ceases to exist after death.e) None of thesebHighlights detachment, reducing mental stress2Q2. What is the warrior’s duty according to Krishna?a) To avoid war at all costs.b) To protect relatives, even if it means neglecting duty.c) To fight for justice and righteousness.d)To surrender to the enemy.e) None of thesecReinforces schema-based moral reasoning3Q3. How does Krishna suggest handling the conflict between duty and personal moral dilemma?a) Prioritize personal moral dilemma over duty.b) Balance between duty and moral dilemma.c) Fulfill the warrior duty to fight irrespective of personal conflict.d)Abandon duty to avoid moral conflict.e) None of thesecEncourages cognitive decentering and conflict resolution4Q4. What crucial lesson does Krishna teach Arjuna in Chap. 1 of the Bhagavad Gita when Arjuna is in crisis?a) The importance of wealth and powerb) The importance of duty and the impermanence of lifec) The importance of personal happinessd)The importance of surrender and giving upe) None of thesebPromotes higher-order reasoning5Q5. Why was Dhrtarastra fearful in the Bhagavad Gita Chap. 1?a) He was afraid of the might of the Pandavas’ armyb) He feared the influence of Kurukshetra as a dharmakshetra, favoring the victory of the Pandavasc)He doubted his own sons’ capabilities in battled) He was afraid of Lord Krishna’s powere) None of thesebReflects situational awareness and ethical anticipation6Q6. What does Krishna say about those who die in the battle fighting for righteousness?a) They are forgotten.b) They are condemned.c)They reach heavenly abodes.d) They are reborn as lower beings.e) None of thesecReinforces motivational cognition7Q7. What makes the battle righteous according to Krishna?a) The number of warriors involved.b) The skills and weapons used.c) The restoration and upholding of dharma (righteousness).d) The vengeance against the enemy.e) None of thesecTests principle-based reasoning and transfer of learning Comprehension quiz Questions, Options, and correct Answers. a) The soul is mortal and dies with the body. b) The soul is eternal and simply changes bodies. c)The soul is fragmented and dispersed after death. d) The soul ceases to exist after death. e) None of these a) To avoid war at all costs. b) To protect relatives, even if it means neglecting duty. c) To fight for justice and righteousness. d)To surrender to the enemy. e) None of these a) Prioritize personal moral dilemma over duty. b) Balance between duty and moral dilemma. c) Fulfill the warrior duty to fight irrespective of personal conflict. d)Abandon duty to avoid moral conflict. e) None of these a) The importance of wealth and power b) The importance of duty and the impermanence of life c) The importance of personal happiness d)The importance of surrender and giving up e) None of these a) He was afraid of the might of the Pandavas’ army b) He feared the influence of Kurukshetra as a dharmakshetra, favoring the victory of the Pandavas c)He doubted his own sons’ capabilities in battle d) He was afraid of Lord Krishna’s power e) None of these a) They are forgotten. b) They are condemned. c)They reach heavenly abodes. d) They are reborn as lower beings. e) None of these a) The number of warriors involved. b) The skills and weapons used. c) The restoration and upholding of dharma (righteousness). d) The vengeance against the enemy. e) None of these EEG data pre processing MUSE S EEG headband (InteraXon Inc.) recorded neural activity using a four-dry-electrode, validated noninvasive wearable (AF7, AF8, TP9, TP10) device with sampling at 256 Hz. A 50 Hz notch filter was applied to eliminate power-line noise, and on-line continuous monitoring permitted stable signal acquisition. The Muse S uses dry electrodes that typically operate at electrode–skin impedances of up to approximately 20 kΩ, consistent with prior research employing consumer-grade EEG systems36-28. A fourth-order Butterworth bandpass filter was applied using Python’s SciPy library to attenuate frequencies outside the 1–40 Hz range, effectively reducing common artifacts such as eye blinks and jaw movements. For further artifact correction, including the removal of ocular and muscle-related noise, Independent Component Analysis (ICA) was performed using the MNE-Python toolbox28. This approach enabled robust and accurate preprocessing of EEG signals in an open-source environment28. The derived signals were used for frequency-domain spectral analyses. To ensure data quality, several signal control procedures were implemented. Artifactual segments exceeding ± 100 µV were removed, resulting in an average artifact rejection rate of approximately 10–12%. Signal-to-noise ratio (SNR) was computed to confirm stable acquisition across channels. Given that the Muse four-channel headband offers limited spatial resolution, particularly for estimating high-frequency gamma coherence, all EEG results are therefore framed as exploratory indicators of engagement rather than confirmatory neurophysiological evidence. Although the Muse S is considered a consumer-grade EEG system, multiple validation studies have supported its research applicability, particularly for spectral and asymmetry analyses. Cannard et al. (2021) demonstrated that Muse yielded comparable results to research-grade systems such as the 64-channel BIOSEMI ActiveTwo in measures of power spectral density (PSD), individual alpha frequency (IAF), and FAA36. Similar feasibility has been shown in applied research domains including simulated flight operations, automotive brain–computer interface control, and language comprehension28. Relative band power analysis EEG data were sampled from four channels and normalized to the usual brain signal range of − 100 to 100 µV. Following filtering, a Fast Fourier Transform (FFT) was performed to calculate relative power between frequency bands delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–100 Hz) to evaluate cognitive reaction to the intervention28. Here, x, y, z, m, and n represent the delta, theta, alpha, beta, and gamma absolute band powers across all channels, respectively. For example, if x is the absolute power of the delta band, then the formula will yield the relative power of the delta band across all four channels. Here, “Ch” denotes the different channels of the MUSE band37, explicitly including two from the frontal (AF7, AF8) and two from the temporal (TP9, TP10) channels. Frontal alpha asymmetry Frontal alpha asymmetry (FAA) was calculated as an index of motivational and emotional states from EEG signals of the Muse headband. Alpha power (8–12 Hz) was derived after filtering (1–40 Hz) and calculating power spectral density using Welch’s method from AF7 and AF8. FAA was computed as: where positive FAA values are related to increased left-frontal activity, which corresponds to approach motivation and positive affect, whereas negative values point towards right-frontal dominance, which is also related to withdrawal and negative affect20,25,38. FAA values were integrated into the Engagement Index to measure learners’ motivational orientation during task performance. Gamma coherence Gamma coherence was calculated as an index of functional connectivity and cognitive integration based on EEG data from the Muse headband. Once the signals were filtered 30–50 Hz to remove other frequency bands except gamma, between AF7 and AF8 frontal electrodes, coherence was estimated with the magnitude-squared coherence. This metric is measuring how synchrony between the left and right frontal areas occurs, where increased coherence is indicative of greater neural coupling that is responsible for focused attention, working memory, and cognitive coordination26,39,40. Gamma coherence supplemented the Engagement Index by offering information about inter-hemispheric communication during the task. Here, x(t) be the bandpass-filtered EEG signal from AF7 in the 30–50 Hz gamma band, y(t) be the bandpass-filtered EEG signal from AF8 in the 30–50 Hz gamma band, f ∈26,40 Hz be frequencies within the gamma band, (f) = cross power spectral density (CPSD) between x(t) and y(t). (f), (f) = power spectral densities (PSD) of x(t) and y(t), respectively. Increased coherence is associated with increased frontal synchrony, an index of attentional engagement, working memory coordination, and cognitive integration. Gamma coherence hence complemented FAA in the Engagement Index by providing a connectivity-based measure of cognitive effort. Data analysis Quiz response data, usability questionnaire data, and engagement survey data were examined with traditional inferential statistical analysis. Performance score on the quiz was measured as the ratio of items correct out of the total items presented to the participant41. Survey instruments were designed and interpreted with attention to construct validity and response quality, following best practices in educational setups42. A two-way repeated-measures ANOVA with group (Picoh and Reading) as the between-subjects factor and transfer (from one to six questions to seventh question), time (pre-during) as the within-subjects factor was used to test main and interaction effects on quiz scores and neural variables. Where significant effects were detected, Tukey’s Honestly Significant Difference (HSD) post hoc tests were used to identify pairwise differences43. Pearson correlation coefficients were computed using IBM SPSS Statistics for Windows (Version 21.0) to assess the linear association between quiz scores and neural variables44. Statistical significance was established at p <.05, and power was kept at 0.80 to minimize Type II error risk45. For multiple correlation tests, the False Discovery Rate (FDR) correction was applied using the Benjamini–Hochberg procedure (q = 0.05) to control for the expected proportion of false positives within each group’s set of correlations46. Results Behavioral results Item-wise analysis and overall quiz performance The item-wise average score results for participants in the two groups Picoh, Reading was analyzed. The item 1 results showed no significant difference among groups, Picoh (M = 0.88, SD = 0.33, 95% CI [0.74, 1.02]), and Reading (M = 1, SD = 0.00, 95% CI [1.00, 1.00]), F (1, 49) = 3.27, p =.07, η² =0.06 (see Fig. 3). The item 2 results showed no significant difference among groups, Picoh (M = 0.92, SD = 0.27, 95% CI [0.81, 1.03]), and Reading (M = 1, SD = 0, 95% CI [1.00, 1.00]), F (1, 49) = 2.08, p =.15, η² =0.04. The item 3 results showed no significant difference among groups, Picoh (M = 0.88, SD = 0.33, 95% CI [0.74, 1.02]), and Reading (M = 0.84, SD = 0.37, 95% CI [0.69, 0.99]), F (1, 49) = 0.16, p =.69, η² =0.003. The item 4 results showed no significant difference among groups, Picoh (M = 0.96, SD = 0.20, 95% CI [0.90, 1.02]), and Reading (M = 1, SD = 0, 95% CI [0.94, 1.06]), F (1, 49) = 1, p =.32, η² =0.02. Item-wise analysis (representing each question) and overall quiz performance across participants among both the groups (Picoh and Reading). Error bars represent 95% CI around the point estimate. No significant differences between the Picoh and reading groups were found. The item 5 results showed no significant difference among groups, Picoh (M = 0.76, SD = 0.43, 95% CI [0.58, 0.94]), and Reading (M = 0.84, SD = 0.37, 95% CI [0.69, 0.99]), F (1, 49) = 0.48, p =.49, η² =0.01. The item 6 results showed no significant difference among groups, Picoh (M = 0.80, SD = 0.40, 95% CI [0.63, 0.97]), and Reading (M = 0.92, SD = 0.27, 95% CI [0.81, 1.03]), F (1, 49) = 1.47, p =.23, η² =0.03. The item 7 results showed no significant difference among groups, Picoh (M = 0.96, SD = 0.20, 95% CI [0.88, 1.04]), and Reading (M = 1, SD = 0, 95% CI [1.00, 1.00]), F (1, 49) = 1, p =.32, η² =0.02. Overall, in every question the performance of the groups was similar. Similarly, the overall quiz performance among participants in the two groups, Picoh and Reading, was analyzed. The results showed no significant difference among groups, Picoh (M = 0.88, SD = 0.20, 95% CI [0.80, 0.96]) and Reading (M = 0.94, SD = 0.10, 95% CI [0.90, 0.98]), F (1, 49) = 1.95, p =.16, η² = 0.03. The results indicate that both groups performed equally well in terms of retention of the moral concepts. Students showed comparable levels of moral comprehension in both robot-mediated and text-based learning, as evidenced by equivalent quiz scores and successful transfer to a new situation. This indicates that the delivery platform did not affect the acquisition or application of moral knowledge when content was held constant. Analysis of quiz performance during transfer The main effect of the two groups, Picoh and Reading, on the quiz score during transfer (average performance score on items 1–6 to performance score on item 7) was not significant, F (1, 48) = 2.05, p =.15, η² = 0.04. The main effect of transfer was not significant, F (1, 48) = 1.62, p =.20, η² = 0.03. The interaction between groups (Picoh and Reading) and transfer was not significant, F (1, 48) = 0.50, p =.48, η² = 0.01 (see Fig. 4). Both Picoh and Reading participants retained the moral content equally well on the first six items and effectively transferred it to a novel situation (seventh item), answering the RQ1, regarding moral transfer. Assessment of comprehension and generalization scores across Picoh and Reading groups. Error bars represent 95% CI around the point estimate. Both Picoh and Reading groups retained the moral content equally well on the first six items and effectively transferred it to the seventh item. Neural findings Frontal alpha asymmetry The frontal alpha asymmetry of participants in the two groups, Picoh and Reading, was analyzed. The main effect of the two groups, Picoh (M = –0.03, SD = 0.25, 95% CI [–0.13, 0.07]) and Reading (M = –0.02, SD = 0.25, 95% CI [–0.12, 0.08]), on frontal alpha asymmetry was not significant, F (1, 48) = 0.02, p =.87, η² = 0.001. The main effect of time (pre-intervention to during-intervention) was not significant, F (1, 48) = 0.21, p =.64, η² = 0.006. The interaction between groups (Picoh and Reading) and time was also not significant, F(1, 48) = 0.02, p =.87, η² = 0.001. Gamma coherence The gamma coherence of the participants in the two groups, Picoh and Reading, across AF7 and AF8 channels was analyzed. The main effect of the two groups, Picoh (M = 0.77, SD = 4.90, 95% CI [–1.25, 2.79]) and Reading (M = 0.73, SD = 4.90, 95% CI [–1.29, 2.75]), on gamma coherence was not significant, F (1, 48) = 0.001, p =.97, η² = 0.000. The main effect of time (pre-intervention to during-intervention) was not significant, F (1, 48) = 1.56, p =.21, η² = 0.04. The interaction between groups (Picoh and Reading) and time was also not significant, F(1, 48) = 0.001, p =.97, η² = 0.00. Frontal alpha asymmetry and gamma coherence were not significantly different between the robot-mediated and the text-based groups, nor over time. This implies that the two instructional modalities provided equal levels of emotional and cognitive engagement at the neural level. Correlation results Correlation between the Picoh group moral learning and neural data We correlated the quiz responses obtained in the Picoh group with neural data (frontal alpha asymmetry and gamma coherence) to explore relationships between them. Prior to correction, the Pearson correlation coefficient between average score (coming from the first six items) and change in frontal alpha asymmetry (during intervention − pre-intervention) was significant, r(48) = − 0.54, p <.05 (see Table 2). However, after applying FDR correction for multiple comparisons (q = 0.05), this effect did not remain statistically significant (p(adj) = 0.10) (see Table 2). The direction of the association nonetheless suggested that participants exhibiting a greater left-frontal shift in alpha power during the Picoh story tended to score higher on the first six items. The correlation between average score (coming from the first six items) and change in gamma coherence (during intervention − pre-intervention) remained nonsignificant both before (p =.43) and after FDR correction (p(adj) = 0.43). Thus, neither neural measure survived correction, though frontal alpha asymmetry showed a trend-level relationship with comprehension performance. Table 2Summary of pearson correlations between behavioural moral-learning measures and EEG-based engagement indices across instructional modalities. The table presents correlation coefficients (r), 95% confidence intervals (CIs), and both uncorrected and FDR-corrected p-values for associations between quiz-based measures of moral learning and neurophysiological engagement indices, frontal alpha asymmetry (FAA) and gamma coherence, across the Picoh (robot-mediated) and reading (text-based) conditions. FDR correction (Benjamini–Hochberg, q = 0.05) was applied within each group’s correlation set.GroupVariable 1Variable 2 r a 95% CIb p p (with FDR correction)cPicohAvg. Quiz (Q1-Q6)Frontal Alpha Asymmetry (FAA)–0.54[–0.71, − 0.28]0.050.10Avg Quiz (Q1-Q6)Gamma Coherence0.19[–0.12, 0.47]0.430.43Transfer (Q7)Frontal Alpha Asymmetry (FAA)–0.49[–0.68, − 0.21]0.050.10Transfer (Q7)Gamma Coherence0.24[–0.08, 0.52]0.290.29ReadingAvg. Quiz (Q1-Q6)Frontal Alpha Asymmetry (FAA)0.32[–0.04, 0.61]0.170.34Avg. Quiz (Q1-Q6)Gamma Coherence–0.04[–0.34, 0.28]0.860.86Transfer (Q7)Frontal Alpha Asymmetry (FAA)0.36[–0.02, 0.64]0.110.22Transfer (Q7)Gamma Coherence0.12[–0.22, 0.44]0.600.60PicohFAA ChangeGamma Coherence Change–0.19[–0.47, 0.12]0.440.66ReadingFAA ChangeGamma Coherence Change–0.19[–0.47, 0.12]0.440.66Notes.a The degree of freedom for all correlations was 48.b The 95% confidence intervals (CIs) were approximated from r values using Fisher’s z-transformation (two-tailed α = 0.05, N = 50).c The p-values were adjusted using the False Discovery Rate (FDR) correction (Benjamini–Hochberg procedure, q = 0.05), applied within each group’s correlation set (Picoh = 4 tests for moral-learning measures, Reading = 4; additional within-index correlations corrected separately). Summary of pearson correlations between behavioural moral-learning measures and EEG-based engagement indices across instructional modalities. The table presents correlation coefficients (r), 95% confidence intervals (CIs), and both uncorrected and FDR-corrected p-values for associations between quiz-based measures of moral learning and neurophysiological engagement indices, frontal alpha asymmetry (FAA) and gamma coherence, across the Picoh (robot-mediated) and reading (text-based) conditions. FDR correction (Benjamini–Hochberg, q = 0.05) was applied within each group’s correlation set. Notes. a The degree of freedom for all correlations was 48. b The 95% confidence intervals (CIs) were approximated from r values using Fisher’s z-transformation (two-tailed α = 0.05, N = 50). c The p-values were adjusted using the False Discovery Rate (FDR) correction (Benjamini–Hochberg procedure, q = 0.05), applied within each group’s correlation set (Picoh = 4 tests for moral-learning measures, Reading = 4; additional within-index correlations corrected separately). We examined correlations between the seventh (generalizability) quiz item and neural measures (frontal alpha asymmetry and gamma coherence) to explore the relationship between them. Before correction, the Pearson correlation between the seventh item and change in frontal alpha asymmetry (during intervention − pre-intervention) was significant, r(48) = − 0.49, p <.05 (see Table 2). However, after applying FDR correction for multiple comparisons (q = 0.05), this effect no longer remained statistically significant (p(adj) = 0.10). The trend suggested that participants showing a greater left-frontal shift in alpha power tended to score higher on the generalizability question. The correlation between the seventh item and change in gamma coherence (during intervention − pre-intervention) remained nonsignificant both before (p =.29) and after FDR correction (p(adj) = 0.29). Thus, neither neural measure survived FDR correction, although the frontal alpha asymmetry measure continued to show a directional trend consistent with enhanced comprehension. Correlation between the reading group moral learning and neural data We examined correlations between quiz performance in the Reading group and neural data (frontal alpha asymmetry and gamma coherence) to explore their associations. Before correction, neither correlation reached significance: r(48) = 0.32, p =.17 for frontal alpha asymmetry, and r(48) = − 0.04, p =.86 for gamma coherence (see Table 2). After applying FDR correction for multiple comparisons (q = 0.05), both results remained nonsignificant (p(adj) = 0.34 and p(adj) = 0.86, respectively). Thus, no reliable relationships were observed between changes in neural activity and average score (coming from the first six questions) in the Reading group, indicating that frontal and gamma-band neural measures did not predict quiz performance. We examined correlations between the seventh (generalizability) quiz item and neural measures (frontal alpha asymmetry and gamma coherence) in the Reading group. Before correction, neither correlation reached significance: r(48) = 0.36, p =.11 for frontal alpha asymmetry, and r(48) = 0.12, p =.60 for gamma coherence (see Table 2). After applying FDR correction for multiple comparisons (q = 0.05), both effects remained nonsignificant (p(adj) = 0.22 and 0.60, respectively). These results suggest that changes in frontal asymmetry and gamma coherence from baseline to intervention did not reliably predict performance on the generalizability item, indicating no measurable link between neural shifts and higher-order comprehension in the Reading condition. Correlation between frontal alpha asymmetry and gamma coherence in Picoh group The correlation between the frontal alpha asymmetry and gamma coherence was measured in the Picoh group to find the relationship between. The Pearson correlation coefficient between the change in gamma coherence (during intervention - pre intervention) and change in frontal alpha asymmetry (during intervention - pre intervention) was not significant (r (48) = −1.85, p =.44) (see Table 2). After applying FDR correction, the result remained non-significant (p(adj) = 0.66). Correlation between frontal alpha asymmetry and gamma coherence in reading group The correlation between the frontal alpha asymmetry and gamma coherence was measured in the Reading group to find the relationship between. The Pearson correlation coefficient between the change in gamma coherence (during intervention - pre intervention) and change in frontal alpha asymmetry (during intervention - pre intervention) was not significant (r (48) = −1.85, p =.44) (see Table 2). After applying FDR correction, the result remained non-significant (p(adj) = 0.66). Whereas between-group comparisons of overall EEG measures did not reveal any differences, within-group analysis of the Picoh group revealed a significant negative correlation between Moral Learning and frontal alpha asymmetry. This implies that those subjects displaying more left-frontal activation (approach motivation) when interacting with robots scored better, and it is inferred that FAA is able to measure individual emotional involvement more sensitively than group averages indicate. Discussion and conclusion This study examined how learning modality influences moral learning by comparing AI-supported robot interaction (Picoh) with traditional reading, using both behavioral and neurophysiological measures to explore underlying engagement processes. By relating quiz performance to FAA (affective–motivational engagement) and gamma coherence (cognitive integration), the study explored whether the two modalities differed in learning performance or in the neurocognitive systems supporting moral reasoning26,39,40,47. Item-by-item analysis indicated that the Picoh group and Reading group were equivalent on moral understanding and transfer. This answers RQ1 directly, meaning that equal content results in equal behavioral learning gain across modalities. This contradicts the assumption that greater interactivity is always superior for learning through richer social and affective cues6,8,10,17,43,48, as the findings show that standard reading can be just as effective for evoking moral understanding, particularly for fact-based or conceptually bounded content34.These findings suggest that interactivity is not inherently superior; the value of a modality depends on learner needs and the type of content6,8,17,48. The behavioral patterns align with moral reasoning theories suggesting that reflective processes can support understanding across modalities8,48. The tentative neural patterns additionally hint that embodied interactions may introduce an affective dimension not captured behaviorally10,12,24,32,43,49,50. At the neural level, the findings suggested tentative, non-significant modality-related patterns of engagement. FAA showed a trend-level association with comprehension and transfer performance in the Picoh group but not in the Reading group, offering preliminary and exploratory insight pertinent to RQ2 and RQ3. FAA quantifies the balance of activation in left- vs. right-frontal areas, which relates to approach motivation and emotional arousal20,25,38. Participants who displayed relatively higher left-frontal activation tended to score somewhat better on comprehension and transfer tasks within the robot-mediated group, although this pattern did not reach statistical significance. These tentative observations are consistent with dual-process models of moral thought10,51, which place affective intuition and motivational orientation at the center of moral thinking, and they provisionally position FAA as a possible neural indicator of these affective–motivational processes10,20,25,38,51. This trend aligns with the causal model put forth previously, in which Instructional Modality is hypothesized to influence FAA, which may in turn influence Moral Understanding and Transfer, suggesting, without confirming, that affective–motivational engagement could play a role in comprehension within embodied learning environments20,25,38,51. Drawing from Haidt’s social intuitionist theory and Greene’s dual-process theory10,51, these observations may indicate that embodied, socially interactive modalities engage intuitive–emotional processes differently than text-based reading, which is thought to rely more heavily on reflective cognitive processing6,8,17,48. Gamma coherence was not significantly associated with moral learning performance in either modality. Although gamma-band synchrony has been linked to attentional integration, its lack of association here suggests that gamma coherence may not be sensitive enough in this design to capture modality-related processing39. This result is in line with previous findings that FAA and gamma coherence reflect distinct dimensions of engagement47, with FAA relating more to affective–motivational orientation and gamma coherence indexing broader cognitive integration26,39,40. Taken together, these exploratory patterns suggest that emotional engagement may play a relatively stronger role than cognitive binding in this context6,8,10,17,48. However, given the exploratory nature of the EEG measures and the absence of statistically reliable effects, these interpretations remain tentative and highlight the need for follow-up work with higher-density EEG systems and larger samples. The overall trend indicates that similar behavioral outcomes may coexist with subtle, non-significant neural variations that require cautious interpretation. Although the two modalities facilitated similar comprehension and transfer, the Picoh condition showed a tentative, trend-level indication of engaging affective–motivational systems more strongly, as suggested by the exploratory FAA modulation20,25,38. This possibility aligns with Bandura’s social cognitive theory10, which highlights affective modeling, and with embodied cognition theories52, which propose that sensorimotor and affective involvement can support higher-order thought12,24,32. The Picoh robot’s expressive speech and multimodal cues may have fostered greater affective–motivational engagement, even though behavioral performance was equivalent12,24,32. If present, such differences would suggest that equivalent learning outcomes might arise through different underlying neurocognitive routes. Embodied interaction might offer motivational and affective scaffolding for generalization, while text-based learning may rely more heavily on reflective reasoning processes6,8,17,24,48–50,53. However, given the exploratory nature of the neural findings, these interpretations remain speculative and highlight the need for further empirical testing. A number of methodological concerns are worth highlighting for future research. EEG measures were taken using the Muse S four-channel headband, a validated, low-cost, portable EEG device intended for ecologically valid data collection in classroom and field settings28. Although it is not as high in spatial resolution as lab-grade systems, earlier work has shown it to be a reliable tool for FAA and spectral power measures at impedance levels held constant at or below 20 kΩ28,36. For the sake of ensuring data quality, artifactual segments greater than ± 100 µV were eliminated, with an average artifact rejection rate of about 10–12%, and signal-to-noise ratio (SNR) was calculated to ensure stable acquisition across channels28. Gamma coherence should be interpreted cautiously, as the limited electrode coverage likely reflects broad frontal synchrony rather than precise connectivity26,39,40,47. The Muse S affords an applied compromise between ecological viability and evidence-based accuracy for exploratory cognitive neuroscience studies28. These EEG results are best viewed as hypothesis-generating and can guide future research using higher-density, research-grade devices to record finer neural dynamics within ecological contexts28,36. Aside from device-level considerations, group-level equivalence at baseline affective and cognitive state was not assessed directly before assignment. The potential for novelty, social presence, and auditory stimulation to affect engagement in the robot condition cannot be discounted12,24,32. The expressive face, voice, and conversational interactivity of the Picoh robot would likely have enhanced participants’ sense of social presence and auditory engagement over the static reading condition12,24,32,49,50. These conditions may have driven differences in affective–motivational engagement (as captured by FAA) rather than in instructional effect20,25,38. These possible confounds might be addressed in future research by adding auditory-only (voice-over) conditions, pre-exposure sessions to reduce novelty, or within-subject crossover designs to more effectively balance personal factors of involvement12,24,32,49,50. The comparatively small sample that is gender-skewed may also restrict the consistency and applicability of results43. While power analysis suggested adequate sensitivity to identify medium effects at α = 0.05, smaller theoretically significant effects might not have been picked up43. Although all participants were at least familiar with one course of introduction to English literature, implicit first-language (L1) to second-language (L2) processing differences might have influenced reading fluency or comprehension; future research could include standardized language skills assessments so as to more fully control for this variable6,8,17,48. It is worth noting that the majority of between-group comparisons were not statistically significant (p >.05). Thus, the modality differences observed should be interpreted descriptively rather than inferentially6,8,17,48. The findings here can be considered preliminary and exploratory, guiding subsequent research instead of confirmatory proof in favor of directional hypotheses6,8,10,17,48. Combined, these factors suggest areas for methodological refinement without devaluing the study’s central contribution, exposing how neurocognitive measures can shed light on engagement mechanisms between instructional modalities26,39,40,47. The findings further current understanding regarding the potential of embodied AI and text-based instruction to differentially engage motivational and reflective systems, offering preliminary insight into how adaptive, context-responsive instructional design could be informed12,24,32,49,50. Acknowledgements This work was supported by the Indian Institute of Technology Mandi, who provided the necessary financial and computational resources for this experiment. This work was also supported by the project grant (IITM/DRDO-LSRB/VD/553) awarded to Prof. Varun Dutt. Author contributions SC contributed to designing the experiment and carried out data collection and analyses in this study. VD was the principal investigator who developed the idea of the study and served as a constant guiding light for this work. AB was the co-principal investigator who helped with the analysis based on behavioural and the neural experiments. All authors have contributed to the manuscript and approved the submitted version. Funding There is no specific funding received for this study. Data availability The corresponding author will provide access to the raw data and code supporting the conclusions of this article upon request without unnecessary restrictions. Declarations Competing interests The authors declare no competing interests. Ethical approval This study involved human participants and was conducted in accordance with the ethical standards and guidelines of the Indian Institute of Technology Mandi (IITM/IEC(H)/2025/VD/P2). The study protocol was reviewed and approved by the Institutional Ethics Committee of the Indian Institute of Technology Mandi. All participants provided informed consent prior to participation. No human tissue samples were involved. Informed consent Informed consent was obtained from all subjects involved in the questionnaire-based study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Citations Data Availability Statement The corresponding author will provide access to the raw data and code supporting the conclusions of this article upon request without unnecessary restrictions.