The longitudinal relationships among AI-assisted music classroom engagement, flow experience, and musical emotional intelligence: a parallel-process latent growth modeling study

Synthesis: Zhang and Tang surveyed 850 undergraduate music students at four Chinese universities five times across a semester, measuring classroom engagement in AI-assisted music classes, flow experience and self-perceived musical emotional intelligence, and modeled the three as parallel growth processes. All three rose significantly over the five waves, and their rates of change moved together: the slope of engagement correlated 0.39 with the slope of flow and 0.22 with the slope of emotional intelligence, while flow and emotional intelligence correlated 0.40. Growth in flow statistically carried the association between growth in engagement and growth in emotional intelligence, with no residual direct path. The design is the strongest longitudinal evidence in the corpus on AI-assisted arts teaching, and it establishes coordinated change rather than causation.

Key Findings

1. All three constructs rose across five waves. The mean slopes of classroom engagement, flow experience and musical emotional intelligence were each significantly greater than zero, and each slope variance was significant, so students also differed in how fast they grew.
2. Initial levels were tightly coupled. Intercepts correlated 0.45 between engagement and flow, 0.42 between engagement and emotional intelligence and 0.44 between flow and emotional intelligence, all at p < 0.001.
3. Change moved together as well. Slope correlations were 0.39 for engagement and flow, 0.22 for engagement and emotional intelligence and 0.40 for flow and emotional intelligence, which is what the growth-cascade hypothesis required.
4. Flow carried the engagement–emotion link. The engagement-to-flow slope path was 0.404 and the flow-to-emotional-intelligence path 0.297, leaving a non-significant direct path of 0.086 (p = 0.217) and an indirect association of 0.120 (95% CI [0.064, 0.205]).
5. Higher starts predicted slower growth. Intercept and slope correlated negatively within every process, from −0.38 for engagement to −0.42 for flow, so students who began most engaged gained least per wave.
6. Attrition was non-random and reported. 15.53% of the sample dropped out between the first and fifth waves, and dropouts had significantly lower baseline scores on all three constructs.

What the model can and cannot say

The three-process parallel-process latent growth model fitted well (chi-square = 107.94, df = 93, CFI = 0.998, RMSEA = 0.014, SRMR = 0.021), and longitudinal measurement invariance was established first, which is what allows the same scale to be compared across five occasions. The authors compared linear, latent-basis and quadratic growth forms and chose linear on BIC for all three variables, a defensible but consequential choice: a linear model reports one average rate of change and cannot detect a plateau.

The growth mediation is the part most likely to be over-read. Slopes are estimated over the whole observation period, so a significant slope-to-slope path means the two processes changed at coordinated rates, not that one preceded the other. The authors state this plainly and name random-intercept cross-lagged panel models as the design that would settle the ordering question.

Engagement, flow and the case for AI in the room

For arts teaching the practically interesting quantity is the level and trajectory of classroom engagement itself. This study cannot isolate what the AI contributed, because there is no non-AI comparison group — every classroom in the study was AI-assisted. What it can say is that in classrooms where AI was present, engagement, flow and emotional intelligence all grew, and that growth in engagement went with growth in flow rather than with growth in emotional intelligence directly.

That is consistent with a familiar instructional story: the activity that holds attention is the one that produces absorption, and absorption is what supports the emotional-reflective work music education cares about. It is also consistent with a much simpler story in which well-taught music classes produce all three, and the paper does not discriminate between them.

Where the sample limits the claims

850 students at first measurement and 718 at the fifth, from four Chinese universities, all self-report. Full-information maximum likelihood handled the missing data under a missing-at-random assumption, but the dropouts scoring lower at baseline means that assumption is doing heavy work, and the authors concede the mean growth slopes may be slightly overestimated. A single data source across five occasions can also inflate associations among the constructs through shared method variance, which the paper lists as a limitation rather than resolving.

The context is also specific: AI-assisted music classrooms in Chinese higher education, with no description of what the AI actually did in the lesson. That makes the study a measure of students' experience of AI-present teaching rather than of any particular AI practice.

What this means for practice

• Watch engagement trajectories, not just levels. Students who started highest gained least per wave, so a cohort that begins engaged can flatten without any drop in absolute scores.
• Design for absorption, then reflection. Growth in flow carried the link between engagement growth and emotional-intelligence growth, which argues for keeping the absorbing activity intact rather than replacing it with reflective writing.
• Expect attrition to bias your own data upward. Dropouts here scored lower at baseline on every measure, which is the usual pattern in voluntary surveys and the reason a course-level evaluation should track who stopped responding.
• Do not claim an AI effect from this design. Every class was AI-assisted; what grew was engagement in AI-present teaching, and comparisons against a non-AI classroom remain untested.
• Researchers. Five waves support latent growth modeling but not temporal ordering; a random-intercept cross-lagged design is the next step the authors themselves name.

Limitations

• No non-AI comparison group. Every participating classroom used AI, so the study describes trajectories within AI-assisted teaching and cannot attribute growth to the tool.
• All measures are student self-report. Engagement, flow and emotional intelligence came from one source at each wave, which can inflate associations between them.
• Differential attrition. 15.53% dropped out by wave five and had lower baseline scores, so the growth estimates may be optimistic under the missing-at-random assumption.
• Slopes cannot order events. Coordinated change over a semester is not temporal precedence, and no cross-lagged model was estimated.
• The AI practice is undescribed. The paper does not report what the AI did in these classrooms, so nothing here transfers to a specific instructional technique.

Citation

Zhang, Y., & Tang, Z. (2026). The longitudinal relationships among AI-assisted music classroom engagement, flow experience, and musical emotional intelligence: a parallel-process latent growth modeling study. Frontiers in Psychology, 17.