Flawed but Memorable: Student Critical Reception of Interest-Personalized GenAI Analogies in Computing Education

Synthesis: Bernstein and Sibia interviewed ten undergraduates who had completed CS2, using a think-aloud annotation task and a semi-structured interview grounded in the Paul-Elder Critical Thinking framework, to see how students critically receive GenAI-generated analogies for linked lists and recursion. Students did not judge explanations holistically: they separated structural accuracy, clarity, appropriateness, and trust, and flagged specific breakdowns in the source-to-target mapping. Most found the interest-personalized analogies more engaging and memorable than generic technical explanations, but trust moved in no single direction, and the sharpest auditors were those with deep knowledge of the analogy's source domain, not simply those whose interests matched. The authors name this two-sided analogy auditing — target-concept knowledge says which relations an analogy must preserve, source-domain knowledge says whether they hold — and argue a familiar source flips the student from learner into expert. Their design turn is to stop filtering flawed analogies out and instead assign them as objects to inspect and repair, turning the tool's least reliable property into a productive difficulty.

Why analogies, and why GenAI explanations

Analogies work by mapping a familiar source domain onto an unfamiliar target domain, using structural correspondence rather than surface similarity — the classic structure-mapping account. In computing education this makes them double-edged: because the mapping is always partial, a learner encountering linked lists or recursion for the first time may lack the grounding to see where the comparison breaks down. A linked-list analogy that works for traversal can quietly misrepresent insertion or deletion, and the "variable as a box" metaphor implies a variable might hold several values at once. Analogies also encode assumptions about whose experience counts as familiar, which can signal to some students that they do not belong. Despite generative AI tool use for concept explanation being widespread, prior research concentrated on code-level and explanation-level correctness, leaving students' critical reception of analogies almost unexamined. The paper asks three questions: what criteria students apply to accuracy and appropriateness (RQ1), what affective and social dimensions they perceive (RQ2), and how they see interest-personalized analogies differing from generic technical explanations in engagement, trust, and preference (RQ3).

Method: think-aloud annotation with CS2 students

Ten students who had completed CS2 were recruited from a large North American research university and paid $30 per session. Each participant saw four explanations — a personalized analogy and a generic technical explanation for both linked lists and recursion — after completing a pre-survey recording coursework, interests, prior GenAI use, and sense of belonging. Personalized analogies were generated by supplying each participant's self-reported interest (soccer, One Piece, Valorant, baking, Sudoku, and similar) as the requested source domain; the two generic explanations were fixed for everyone. Crucially, researchers did not edit outputs or plant errors, so the structural imperfections in the analogies were the model's own — a circular-route linked-list analogy and a badminton-rally recursion analogy with no guaranteed shrinking input both entered the study as genuine model artifacts. Presentation order was counterbalanced within each concept but the concept order (linked lists before recursion) was fixed. A reflexive thematic analysis in an interpretivist stance followed, without intercoder reliability or a consensus codebook, treating the researchers' own interest in personalization as a lens to interrogate rather than a bias to eliminate.

Evaluation criteria: structural correspondence and precise language

Students evaluated analogies as arguments with a point of view, not as true-or-false statements, and they audited at the level of mechanics. P6 corrected the One Piece linked-list analogy's structure rather than the computing concept: passing the last island and looping back to the first describes a circular linked list, not the singly linked list under study. P8 noted that a Sudoku recursion analogy's base case of 1 does not fit, since a board cannot always be reduced that way. P2 caught an agency error in a video-editing analogy — "the playhead knows what's next, rather than the clip." P6 rejected a badminton-rally analogy because a rally carries no guarantee of a shrinking input and proposed golf, where each stroke covers less distance, as structurally sound. A second recurring standard was precision of language: hedging phrases left authors "open to interpretation," and several students distinguished vagueness from missing elaboration, wanting the AI to justify claims ("this is here for this reason") rather than assert them. Clarity judgments bled into accuracy — one student doubted a text's accuracy simply because it was "a lot more bloated." Verification was active, not passive: cross-checking against lecture material, demanding provenance, and probing with follow-up questions for inconsistency. These behaviors echo known patterns of student–AI interaction and misalignment between what explanations claim and what learners can check.

Appropriateness, assumptions, and the imagined reader

Students read explanations for whose knowledge and background they assumed. The most direct failure was assuming Prior Knowledge the student did not yet have, as when an explanation reached beyond the syllabus into unseen theorems. A subtler failure was assumed domain familiarity through sport and gaming references. P9 recalled analogies from "the more male-dominated side of computing" and said that as someone who does not follow sports, "I'm not sure what to do with that." Even an avid Valorant player (P2) flagged that a Valorant analogy assumes in-game terminology ("what's an entry fragger?"). The concern was about the references AI reaches for by default, which links to equity and inclusion rather than to a single analogy. Excessive explanation was read as a social signal, not a stylistic one: over-explaining familiar material felt condescending and implied the tool assumed the student knew less than they did, complicating any "more Scaffolding is safer" heuristic. It also touches inclusion and belonging. Not every student applied the fairness lens by default — P10 self-attributed confusion to insufficient personal effort rather than interrogating the explanation's assumptions.

Personalization vs generic explanations: engagement, trust, and control

Most participants described the personalized analogies as more engaging and memorable. When personalization landed, the shift was qualitative: P6 said the One Piece analogy "might have even helped me to understand linked lists," and P8 said the study "changed my mind about analogies." But the pattern was not uniform. P10 rejected a baking analogy as non-generalizable ("not everyone bakes") — an interest-based analogy can exclude students who do not share the interest. Several students preferred the plain technical explanation even when the personalized one matched their stated interest, because it stood on its own while the analogy would have required follow-up prompts to reach the same depth. Trust moved in no single direction: one student was "less critical about it because I could trust it more," while another was made less trusting by heavy personalization and tried to disclose the minimum about themselves. What tracked scrutiny most consistently was domain expertise, not personal interest — students could tell when a mapping "was going wrong" because they knew the concept, and one noted that fluency in the source series let them appreciate an analogy's craftsmanship. Asked how a tutor should handle personalization, participants converged on learner control: offer a menu of candidate analogies, or reserve analogy for after a technical explanation has failed, or treat the two formats as complements.

Two-sided analogy auditing and critique-as-learning

The paper's conceptual contribution is two-sided analogy auditing. Target-concept knowledge identifies the relations an analogy must preserve (recursion's shrinking input and dependent calls, a linked list's pointer chain); source-domain knowledge determines whether those relations hold there, and neither side alone is sufficient. Interest personalization may therefore have an epistemic role in addition to its motivational one: a familiar source makes an analogy both more engaging and more inspectable, because the student who knows the source can audit it. This reframes a long-standing worry — Halasz and Moran's claim that learners cannot separate relevant from irrelevant inferences is, on this account, specific to one side of the mapping rather than a general deficit, and the same reframing turns the accuracy risk into an instructional resource. If locating where a mapping breaks requires knowing the concept's actual structure, then analogy critique is itself a check on understanding, and harder than reading a finished explanation: a student who shows that a badminton rally lacks a shrinking input has demonstrated a real component of recursion. The paper also argues the field's habit of scoring explanation quality as a single correctness judgment is inadequate, since learners assess concept, source, and correspondence as separate layers and may rate an analogy memorable while doubting its accuracy. It connects to existing work on Critical Thinking and Metacognition in AI-integrated learning.

Design implications and limitations

The authors recommend three design moves and one teaching move: personalize through source-domain structure rather than topical relevance alone; verify structural fit before delivering an analogy, because no amount of relevance rescues a broken mapping; and give learners control over personalization, including the option to see a generic explanation alongside a personalized one. The teaching move is the most consequential: assign flawed analogies as objects to inspect, question, and repair. The study's limits are stated clearly — personalization is confounded with explanation format (personalized items were analogies, generic ones were technical prose), so engagement effects cannot be attributed to personalization alone; a generic-analogy condition would be needed to separate them. Concept order was fixed, participants had already completed CS2 so their criteria may differ from true beginners', and their interests leaned toward gaming, sports, and popular media — the very reference space one participant found alienating. Four lessons (and several patterns resting on one or two accounts) follow from ten participants recruited by email from a pool of 667.

What this means for practice

• Instructors. Assign flawed GenAI analogies as objects to inspect and repair rather than filtering them out: locating where a mapping breaks, such as a rally with no shrinking input, is itself a check on understanding recursion.
• Instructors. Verify structural fit before delivering an analogy, because no amount of topical relevance rescues a broken mapping — students caught circular-route comparisons for singly linked lists and base cases that do not fit.
• Instructors. Give learners control over personalization: offer a menu of candidate analogies, or reserve analogy for after a technical explanation has failed, since some students preferred plain technical prose even when the analogy matched their stated interest.
• Designers. Personalize through source-domain structure rather than topical relevance alone, and audit the reference space — one participant flagged sport and gaming references from "the more male-dominated side of computing" as alienating.
• Researchers. Separate format from personalization in the next design: a generic-analogy condition is needed because the personalized items were analogies while the generic ones were technical prose.

Limitations

• Ten students recruited by email from a pool of 667 CS2 completers, each paid $30 per session, so several of the reported patterns rest on one or two accounts.
• Personalization is confounded with explanation format — personalized items were analogies and generic items were technical prose — so the engagement and trust effects cannot be attributed to personalization alone.
• Presentation order was counterbalanced within each concept, but the concept order (linked lists before recursion) was fixed for every participant.
• The reflexive thematic analysis reports no intercoder reliability or consensus codebook, and participants' interests leaned toward gaming, sports and popular media, the very reference space one participant found alienating.

Connected Concepts

• CS Education
• Generative AI
• Critical Thinking
• Personalized Learning
• Student-AI Interaction
• Misconceptions about AI
• Trust Calibration
• Prior Knowledge
• Student Engagement
• Learning Design
• Qualitative Research
• Constructivism
• Metacognition
• Desirable Difficulties
• Cognitive Offloading

Connected Articles

• With a Thermomix You Lose the Ability to Cook: A Kitchen Machine Analogy for Applications of Generative AI in Education — With a Thermomix You Lose the Ability to Cook: A Kitchen Machine Analogy for Applications of Generative AI in Education
• Exploring the Value of Diverse LLM Explanations in Introductory Programming — Exploring the Value of Diverse LLM Explanations in Introductory Programming
• When AI Is Wrong on Purpose: How Students Respond to Buggy GenAI Code — When AI Is Wrong on Purpose: How Students Respond to Buggy GenAI Code
• How Students (Mis)understand Conditionals and Loops -- A Taxonomy — How Students (Mis)understand Conditionals and Loops -- A Taxonomy
• Uncovering Students' Mental Models of Generative Artificial Intelligence — Uncovering Students' Mental Models of Generative Artificial Intelligence
• Calibrating Trustworthiness: Co-Designing Metrics and Visualizations for Evaluating LLMs in Education — Calibrating Trustworthiness: Co-Designing Metrics and Visualizations for Evaluating LLMs in Education
• Characterizing Students' LLM Usage Behaviors and Their Association with Learning in Critical Thinking Tasks — Characterizing Students' LLM Usage Behaviors and Their Association with Learning in Critical Thinking Tasks
• Did Alice Do Wrong? Cross-Cultural Differences in Student Perceptions of Generative AI Use in University Computing Education — Did Alice Do Wrong? Cross-Cultural Differences in Student Perceptions of Generative AI Use in University Computing Education
• Distilling Black-Box Machine Learning into a Small, Self-Explaining Language Model for Learning Analytics — Distilling Black-Box Machine Learning into a Small, Self-Explaining Language Model for Learning Analytics
• The critical-thinking paradox in generative AI-integrated learning: distinguishing efficiency from cognitive depth — a differentiated framework and testable propositions — The critical-thinking paradox in generative AI-integrated learning

Citation

Bernstein, S., & Sibia, N. (2026). Flawed but Memorable: Student Critical Reception of Interest-Personalized GenAI Analogies in Computing Education. Koli Calling 2026 (26th International Conference on Computing Education Research). arXiv:2609.06095.