AI-mediated cognitive divergence in built-environment education: Evidence from a mixed-methods study

Synthesis: Crolla, Xia, and Jiang (2026) report a sequential mixed-methods study of what generative AI is doing to learning processes in design education, conducted across the built-environment Faculty of Architecture at the University of Hong Kong. Drawing on 24 faculty interviews, a survey of 32 instructors with multivariate regression modeling, and five clustered faculty discussions, they identify a pattern they term cognitive divergence: AI does not produce uniform improvements but amplifies existing differences in student readiness. Students with stronger foundations use AI to extend reasoning and accelerate iteration, while those with weaker foundations delegate formative cognitive work to AI, producing coherent outputs without corresponding understanding. This divergence is compounded by a loss of process visibility and by the erosion of frictional learning stages through which competence is built. Quantitative analysis identifies perceived pedagogical relevance rather than seniority as the primary predictor of faculty AI positivity, with a significant negative association between theoretical course orientation and AI positivity.

Core Finding

Generative AI acts as an amplifier of existing differences in students' cognitive readiness rather than an equalizing force. Consistent with Cognitive Offloading theory, stronger students use AI to extend reasoning and reach higher levels of elaboration, while weaker students outsource formative cognitive work and produce visually coherent but cognitively hollow outputs. The performance floor rises across the cohort while the ceiling does not correspondingly rise, and the divergence risks consolidating into a Matthew effect where existing advantage compounds into further advantage. Faculty positiveness toward AI is driven by perceived pedagogical relevance (β = 0.534, p = .004), not career stage, and theory-oriented courses predict skepticism (β = −0.380, p = .030).

Cognitive Divergence and the Amplifier Mechanism

Across all three strands, faculty consistently describe AI as amplifying rather than equalizing differences in student preparation. Students with stronger conceptual foundations engage AI as a generative extension of existing reasoning — iterating faster, exploring more widely, and reaching higher levels of elaboration. Students with weaker foundations accept machine-generated outputs without interrogating their logic, producing work that passes surface inspection but collapses under direct questioning, as one instructor put it, a "false sense of capability." At the level of capability formation, students who bypass effortful cognitive processes fail to build the transferable foundations needed to evaluate or extend their work independently. Read through Scaffolding theory, AI approximates productive scaffolding for strong students but substitutes for development for those without the foundations to evaluate what the support provides.

Loss of Process Visibility and Erosion of Friction

Two structural changes compound the divergence. First, a loss of process visibility: before widespread AI, a coherent submission implied sustained iterative engagement, but coherent outputs can now be produced through rapid prompting without the reasoning they appear to represent, severing the relationship between submission and learning. This is especially consequential in design education, which has long relied on visible process (sketches, intermediate artifacts, iterative models) as its primary indicator of learning. Second, the erosion of productive friction: competence is built through frictional stages — debugging, close reading, iterative drafting — that correspond to what learning research terms "desirable difficulties," conditions that impede immediate performance while enhancing long-term retention and transfer. When AI strips out these stages, students may complete tasks while failing to build the foundations needed to extend their work, putting the most professionally durable capabilities (independent learning, critical synthesis, problem framing, reasoning through model assumptions) at risk.

Structural Predictors of Faculty AI Positivity

Pairwise OLS regression across 595 item pairs (FDR-corrected) and a multivariate model (R² = 0.451; adjusted R² = 0.346; F(5,26) = 4.28, p = .006) identify two significant predictors of faculty AI positivity: perceived AI effect on teaching (β = 0.534, p = .004) and theoretical course orientation (β = −0.380, p = .030, negative). Seniority is non-significant throughout (all q > 0.05, all adjusted R² < 0.08), directly contradicting the assumption that resistance reflects generational or career-stage factors. Instructors who perceive AI as widening student performance gaps are also less positive about its future role (slope = −0.712, q = .030). The negative association between theoretical orientation and AI positivity reflects a genuine epistemological mismatch: current tools suit iterative, generative, and visualization-oriented workflows, while in theory-oriented courses requiring argumentation and critical reading, AI can enable students to produce fluent arguments that conceal absent understanding. Constructive alignment provides the frame — integration is supportive where tools align with intended epistemic outcomes and undermining where they allow outcomes to be simulated rather than achieved.

Adaptive Faculty Responses and Their Limits

Faculty responded with four overlapping strategies: Assessment redesign (open-ended tasks requiring problem definition, oral questioning, voice-over explanations, industry-linked briefs); process enforcement and visibility (intermediate submissions, hand-drawn diagrams, individual presentations within group work); deliberate friction and deautomation (time-constrained exercises, incomplete models requiring debugging, staged workflows); and reframing AI use as a scaffold (structured multi-step workflows where students evaluate, revise, and justify AI outputs). None is scalable under standard teaching conditions: effective interventions require sustained instructor observation and small cohorts, viable in studios but not across lecture cohorts of 70–140 students. This scale problem argues for institutionalizing process-visible assessment as a structural requirement rather than an individual workaround — for example, a minimum process-evidence requirement for AI-permitted assignments and a standard AI-use declaration recording tools, prompts, and what was accepted, rejected, or reworked and why.

What this means for practice

• Instructors. Differentiate AI integration by course epistemology as well as by student preparation: where a course's intended outcome is argumentation or critical reading, fluent AI-assisted prose can simulate the outcome rather than achieve it.
• Treat conceptual AI Literacy and shared technical foundations as a precondition for productive AI use, and place them early in the curriculum rather than alongside advanced AI work.
• Restrict AI automation in formative coursework until core habits of reading, writing, and iterative manual design are established, and rebuild productive friction deliberately through time-constrained exercises, incomplete models that require debugging, and staged workflows where students evaluate, revise, and justify AI output.
• Administrators. Treat process-visible Assessment formats as data infrastructure, not only as assessment reform: intermediate submissions, hand-drawn diagrams, and oral questioning restore the evidential basis of grading and generate the traces that process-level Learning Analytics needs to catch cognitive divergence early.
• Faculty developers. Coach faculty to read their own course designs for alignment gaps, because the study found that perceived pedagogical relevance, not seniority, drives AI adoption, and theory-oriented courses are where tool and outcome are most likely to mismatch.

Limitations

• The survey's n = 32 sets five predictors against only 26 residual degrees of freedom, and the authors caution that the adjusted R² of 0.346 should be read as indicating structural patterns rather than precise effect estimates.
• Participants were self-selected into a conversation about AI and pedagogy: the authors note that the qualitative and deliberative strands draw on faculty "already actively engaged with questions of AI and pedagogy," so these accounts likely reflect more adaptive and reflective orientations than a random faculty cross-section.
• Student learning is characterized entirely through faculty perception — an absence of student voice the authors name explicitly — with no student interviews, portfolio analysis, or task-based measures of cognitive engagement.
• It is one built-environment faculty at a single research-intensive university in Hong Kong (24 interviews, 32 survey respondents, five clustered discussions with 31 participants), so whether the amplification mechanism and the scalability constraints hold across other disciplines, resources, and governance cultures is untested.

Citation

Crolla, K., Xia, X., & Jiang, Y. (2026). AI-mediated cognitive divergence in built-environment education: Evidence from a mixed-methods study. Computers and Education: Artificial Intelligence.