Research · Instructional Design
Generative AI is reshaping how summative assessments are built, delivered, and scored in higher education, but speed and scale are not the same thing as pedagogical soundness. As AI tools take on more of the mechanical work of writing exam questions, the question facing the higher education community is not whether AI can generate assessment items, but whether those items still serve the learning outcomes they are meant to measure. This is where instructional design frameworks become indispensable. One of the core features of Collage’s summative assessment design is the utilization of Bloom’s Taxonomy, where instructors can pick a corresponding Bloom’s Taxonomy level to go with their generated questions. Embedding instructional design principles like Bloom’s directly into an AI assessment workflow is not a cosmetic feature; it is what preserves faculty agency as AI becomes more deeply integrated into higher education.
By Megan Allen, Ed.M.
Partner Success Manager @ Collage AI
The Framework
A shared vocabulary for cognitive demand
Bloom’s Taxonomy is an instructional design framework that allows educators to categorize and evaluate learning objectives by their complexity. The taxonomy is based on the following levels, listed in order of increasing difficulty: Remember, Understand, Apply, Analyze, Evaluate, and Create (Stapleton-Corcoran, 2023). Each level corresponds to a different kind of cognitive demand, moving from simple recall of facts toward the ability to synthesize original work, and the framework has remained a durable reference point for curriculum designers precisely because it gives a common vocabulary for describing how difficult a task is meant to be, independent of subject matter.
01
Remember
02
Understand
03
Apply
04
Analyze
05
Evaluate
06
Create
The six levels, listed in order of increasing difficulty.
Why it matters for summative assessment
Bloom’s Taxonomy matters for summative assessment because it provides a shared map for aligning exam items with intended outcomes and cognitive depth. A recent systematic literature review found that the taxonomy functions as a foundation for building valid and meaningful evaluations, and that Bloom-based summative assessments show stronger validity and better alignment with learning outcomes, particularly when paired with rubrics and authentic assessment tasks (Pangga et al., 2025). The pedagogical reason for this is straightforward: students tend to study for the way they expect to be tested, so a summative exam built with Bloom’s levels in mind can push learning beyond recall toward analysis, evaluation, and creation, rather than allowing students to succeed through memorization alone (Alias et al., 2011). When exams are constructed without attention to cognitive level, they risk rewarding surface familiarity with material instead of the deeper competencies a course is designed to build.
AI-Generated Questions
Calibration is the variable that matters
This dynamic becomes more consequential, not less, once AI enters the picture. AI-based question generation tools have shown real promise in the research literature, but that promise has consistently depended on whether the system is explicitly conditioned to target a specific cognitive level rather than left to generate questions unconstrained. In one AAAI-published study, researchers found that teachers strongly preferred quizzes built with AI-generated questions that were explicitly aligned to Bloom’s taxonomy levels, and that these quizzes showed no loss in quality compared to questions written entirely by hand, with some metrics even favoring the AI-assisted versions (Elkins et al., 2024). The key variable across this line of research is not whether AI can produce a plausible-sounding question, but whether it can produce a question calibrated to the cognitive demand an instructor actually intends. Without that calibration, an AI system can generate large volumes of items that all cluster at the “Remember” or “Understand” level, quietly narrowing the assessment even as it appears to diversify it.
Human-Centered Design
Grounding AI in instructional design
So why does utilizing this framework matter specifically for summatives that are enhanced with AI? Bloom’s Taxonomy improves student learning outcomes in this context as well, and the research literature suggests that AI-generated summative assessments are considerably less effective for students when they are not grounded in an explicit instructional design framework. A 2025 framework paper goes further into this idea, arguing that a human-centered model should keep pedagogical criteria and faculty judgment central while AI handles suggestion and support rather than final decision-making (Martin et al., 2025). This human-centered framing is echoed in broader research on faculty experience with AI in higher education. A narrative review of nine empirical studies on faculty perceptions found that instructors are not simply passive users of AI tools; when educators exercise intentional, critical control over how generative AI is used in their teaching, they help validate, contextualize, and sometimes challenge the outputs those systems produce, rather than accepting them uncritically (Buele & Llerena-Aguirre, 2025). That same review also found that faculty who feel excluded from decisions about how AI is implemented in their classrooms report a diminished sense of professional agency and are more likely to resist adoption altogether, even when the underlying tool might otherwise help them.
Taken together, this body of research points to a consistent conclusion: instructional design is what turns a general-purpose AI system into a pedagogically constrained one. When Bloom’s Taxonomy is built into the assessment generation process itself, rather than applied as an afterthought, AI can support validity, alignment, and meaningful feedback while leaving the interpretive and evaluative judgment where it belongs, with the instructor. Without that scaffolding, AI can still speed up scoring and content generation, but the literature consistently flags real costs to that speed: opacity in how items are generated, the risk of algorithmic bias going undetected, and a general thinning of educational meaning as questions drift toward the lowest-effort cognitive levels.
In Practice
How Collage closes the gap
This is precisely the gap that the Collage platform is attempting to close. By requiring a Bloom’s level selection as part of question generation rather than treating cognitive complexity as an optional setting, an assessment platform keeps the instructor’s pedagogical intent visible and enforceable at every step, rather than buried inside a model’s default behavior. It also gives instructors a legible way to audit what they are getting: if a batch of AI-generated questions is supposed to sit at the “Analyze” level and instead reads like recall, the mismatch is immediately apparent rather than hidden inside an opaque generation process. That legibility is itself a form of faculty agency, since it lets instructors correct, override, or reject AI output on pedagogical grounds rather than simply trusting that the system got it right.
The Takeaway
The broader lesson for AI-enhanced higher education is that efficiency and pedagogical validity are not automatically in tension, but they also do not automatically align.
Bloom’s Taxonomy, and instructional design frameworks like it, are what make that alignment possible by giving both the AI system and the instructor a shared, explicit standard for what a given question is supposed to demand of a student. As AI becomes further embedded in summative assessment, frameworks like Bloom’s are not a legacy feature to be worked around; they are the mechanism by which faculty judgment remains the organizing principle of the assessment, even as the tools used to build it change.
References
1
Alias, A., Ariffin, K., Bhkari, N. M., & Zaini, A. A. (2011). Aligning assessment to course outcomes in OBE: Comparing the application of Bloom’s taxonomy in final examination papers in UiTM.
2
Buele, J., & Llerena-Aguirre, L. (2025). Transformations in academic work and faculty perceptions of artificial intelligence in higher education. Frontiers in Education, 10, Article 1603763. https://doi.org/10.3389/feduc.2025.1603763
3
Elkins, S., Kochmar, E., Cheung, J. C. K., & Serban, I. (2024). How teachers can use large language models and Bloom’s taxonomy to create educational quizzes. Proceedings of the AAAI Conference on Artificial Intelligence, 38(21). https://doi.org/10.1609/aaai.v38i21.30353
4
Martin, F., Kim, S., Bolliger, D. U., & DeLarm, J. (2025). Assessment types, strategies, and feedback in online higher education courses in the age of artificial intelligence: Perspectives of instructional designers. TechTrends, 69(6), 1330–1346. https://doi.org/10.1007/s11528-025-01115-8
5
Pangga, D., Ratnaya, I. G., Lanang Agung Parwata, I. G., Ayu Made Budhyani, I. D., & Hasna Hamiydah, S. (2025). The use of technology Bloom’s taxonomy in formative and summative evaluation: A systematic literature review. Journal of Innovative Technology and Sustainability Education, 1(2), 76. https://doi.org/10.63230/jitse.1.2.76
6
Stapleton-Corcoran, E. (2023). Bloom’s taxonomy of educational objectives. Center for the Advancement of Teaching Excellence, University of Illinois Chicago. Retrieved July 15, 2026, from https://teaching.uic.edu/blooms-taxonomy-of-educational-objectives/
See how Collage AI builds assessments.
Explore the platform or join a short walkthrough to see Bloom’s-aligned question generation in action.
Contact
