Answers are grounded in this paper and cite their sources.
Generative AI (GenAI) has entered chemical-engineering classrooms faster than the evidence on its learning effects has matured. This critical review synthesizes the 2023–2026 literature under a single guiding principle. GenAI should be introduced without harming the student learning process first , and enhance learning second. The best causal evidence shows the harm is real but design-governed. In a pre-registered field randomized controlled trial, an unguarded chatbot raised assisted-practice performance yet left students about 1 7 % worse on a subsequent unassisted exam, whereas an answer-withholding tutor returned them to the no-harm baseline. Across peer-reviewed meta-analytic evidence, the average short-term effect on achievement is positive but moderate and heavily conditional (Hedges’ 𝑔 = 0 . 5 7 7 ). We calibrate what current models can and cannot do on chemical-engineering tasks: strong at lower-order recall and symbolic work, degrading sharply toward analysis and synthesis, and prone to confidently wrong numerics that feed safety-critical calculation chains. We then compare the available integration frameworks (UNESCO, AILit, DEC, IDEE, RCAIE, AIAS, two-lane assessment, and the authors’ own CASF, disclosed) through a dialogic-pedagogy lens cross-walked to Bloom’s taxonomy, cognitive load, and constructive alignment. We consolidate verified design conditions: engineered restraint and fading, reflection before help, two-lane assessment, integrity by design rather than unreliable detection, curriculum-embedded AI literacy, and discipline-specific applications from unit operations to process safety. We close with a research agenda anchored in confirmed gaps. Above all, no chemical-engineering-specific controlled learning-outcome study yet exists.
No takes yet. Share an insight, caveat, or question.
Reyes et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: