Case study links AI reading workflow to feasibility, not learning gains

A one-semester practitioner case study found that an AI-assisted reading-assessment system combining generated passages and questions, rapid optical-mark-recognition grading, and same-session electronic feedback could be operated repeatedly by one instructor in an EFL university course. The approach, called the AI-Mediated Motivational Reading Assessment framework, was used across six quizzes involving 53 students in two sections. Its clearest demonstrated achievement was logistical rather than educational: the complete cycle was sustained throughout the semester, with same-session grading and score publication maintained. Instructor observations suggested that students sometimes discussed quiz content and anticipated grades, but the study did not measure motivation with a validated instrument and could not determine whether any score differences were caused by AI-generated material.

The research was a single-cohort descriptive case study conducted in a required undergraduate reading-comprehension course at a university in the Middle East during the 2025–2026 academic year. It was not a controlled experiment. The cohort comprised 53 predominantly 18-to-22-year-old Arabic first-language speakers working at approximately B2 proficiency under the Common European Framework of Reference; Section 1 enrolled 33 students and Section 2 enrolled 20. The first quiz used a textbook passage with publisher-written questions, while the remaining five used passages and multiple-choice questions generated by a large language model. Quizzes 1 through 3 were delivered online without proctoring, whereas Quizzes 4 through 6 were completed on paper under invigilation. Passage source therefore changed alongside delivery mode, monitoring conditions, quiz position, and cumulative exposure, leaving no comparison that isolated an AI effect.

Across all six quizzes, the mean score was 76.5%, and 83.0% of the students who sat each assessment met the study’s pass threshold of 60%. The textbook-based first quiz produced a mean of 69.4% with a standard deviation of 25.1, while the five AI-generated quizzes averaged 77.8% with a standard deviation of 19.8. The resulting 8.4-percentage-point difference is the finding most likely to attract attention, but it cannot safely be described as an AI advantage. The textbook assessment was a single instrument placed first in the semester, its difficulty was not independently calibrated, and students subsequently became familiar with the multiple-choice routine and the instructor’s item conventions. Attendance also varied, with 49 students present for Quiz 1 and as few as 41 present for Quiz 3, potentially altering the composition of the group represented in each mean.

Delivery conditions complicated the comparison further. Unproctored online quizzes averaged 78.9%, compared with 73.6% for invigilated paper quizzes. Among AI-generated quizzes alone, the online average was 84.4%, about 10 points above the paper-based average. Online students had unrestricted access to dictionaries, translation applications, browser searches, and generative AI assistants, so their scores may reflect a mixture of independent reading and external assistance in an unknown proportion. The researchers therefore treated the invigilated results as the more defensible estimate of independent reading comprehension. When only paper-based AI quizzes were included, their mean was 73.6%, narrowing the gap with the textbook quiz to 4.2 points; even this smaller difference remained entangled with quiz sequence, instrument difficulty, and changing attendance. No inferential tests were conducted because the design lacked a control group, random assignment, counterbalancing, and a pre-test, so the study cannot show whether its reported differences exceed ordinary variation or apply beyond this cohort.

The framework’s six-stage workflow was designed to make frequent reading assessment practical for an individual instructor. In the first stage, the instructor chooses a topic connected to the class and prompts a language model for a passage of suitable length, vocabulary level, and proficiency. The second stage generates multiple-choice items spanning literal comprehension, vocabulary in context, inference, main idea, and author’s purpose. Session-adaptive loading uses 20 questions for a 45-minute class or 40 for a 90-minute class; the framework recommends a three-passage arrangement for extended paper assessments, and the final quiz was implemented that way with three thematically independent passages and 40 items. Students are told they may submit their answer sheet and leave without requesting permission, eliminating the conventional period in which early finishers wait for peers. After collection, a smartphone optical-mark-recognition application scans the sheets, produces scores and item-level information, and the instructor publishes results through Blackboard or a comparable learning-management system, typically 15 to 20 minutes after completion.

These components were intended to address four motivational conditions named interest, achievement, practice, and focused attention. Tailored subject matter was meant to trigger situational interest, rapid and visible results were intended to provide evidence of competence, repeated quizzes were designed to create practice, and permission to leave was intended to direct attention toward the text. The framework draws on Self-Determination Theory, Expectancy-Value Theory, research on situational interest, and evidence that timely task-level feedback strengthens the connection between effort and consequences. Those theories supplied its design rationale, however, rather than demonstrated mechanisms. The study included no validated measure of intrinsic motivation, situational interest, autonomy, competence, or flow, and it did not vary the framework’s elements independently. As a result, the case cannot show which component produced any behavioral or score difference, or whether the package had an effect at all.

Qualitative observations provided a suggestive but methodologically fragile signal. The instructor reported that students discussed quiz content with one another after sessions, checked the learning-management system for anticipated grades, and continued reviewing answers under the departure protocol. One student described the final, three-passage quiz as “interesting.” No comparable behaviors were recalled from textbook activities. Yet the instructor was also the framework’s designer, teacher, quiz administrator, observer, and sole author, making the qualitative record highly dependent on one informed perspective. Observations were not collected under a predefined coding scheme, no second observer rated the behaviors, and textbook sessions were remembered rather than observed contemporaneously. Behaviors inconsistent with the framework, including complaints, indifference, or disengagement, were not systematically recorded. The observations are therefore grounds for investigation rather than evidence that the framework increased intrinsic motivation, and continuing to check answers could reflect grade incentives instead of voluntary interest.

Generative AI also transferred substantial editorial work from writing to judgment. Large language models can produce fluent prose and plausible questions, but the source article identifies several failure modes: fabricated dates, statistics, attributions, or causal claims; stereotyped portrayals arising from demographic prompting; grammatical but unnatural combinations or inconsistent registers; and prose that favors explicit topic sentences, regular paragraph structures, and controlled vocabulary. The instructor checked every passage for factual accuracy, cultural appropriateness, linguistic quality, and curriculum alignment. Verifiable claims were checked against independent sources, while unverifiable claims were removed. Every question was examined for a correct key, plausible distractors, ambiguity, inferential demand, and independence from other items. The instructor reported revising or replacing items at a rate broadly consistent with published figures of 15% to 30%, although systematic item-level records were not retained. The workflow is therefore not an autonomous teaching system: it changes the teacher’s task from authoring every element to evaluating polished content that can look correct while remaining wrong.

The case also exposes a tension between making reading immediately engaging and preparing students to handle unfamiliar material. Culturally relevant passages can activate background knowledge, but a curriculum composed exclusively of familiar content may underdevelop the ability to infer meaning without schema support. Generated prose may also be easier to process partly because it lacks the syntactic and rhetorical variety of authentic academic writing. The textbook quiz had the widest score dispersion in the cycle, which is compatible with authentic prose discriminating more sharply among readers, although one passage cannot establish that explanation. A sound evaluation would compare multiple matched textbook and AI-generated passages under identical monitored conditions, counterbalance their order, measure motivation with validated pre- and post-intervention instruments, and include independent, blinded observers. Results should also be tested on authentic, schema-unsupported texts to determine whether easier AI-assisted practice improves transferable comprehension or primarily creates comfort with familiar material.

For now, the case study supports a deliberately modest conclusion: a feedback-rich reading-assessment cycle can be built from widely available tools and maintained across six sessions for 53 EFL university students, with same-session grading and publication preserved throughout. It does not establish that AI-generated passages improve comprehension, raise intrinsic motivation, or outperform textbooks, and it cannot attribute observed behavior to any individual component of the package. Its practical value lies in documenting a replicable workflow and generating testable hypotheses about how interest, autonomy, practice, and rapid feedback may interact. The next stage of evidence requires matched materials, controlled proctoring, validated measurements of motivation and comprehension, independent observation, and assessment on authentic texts, rather than treating the apparent engagement or score difference as an established effect.

Subject of Research: AI-assisted EFL reading assessment workflow and motivational engagement in a university case study

Article Title: Can AI make reading more engaging? A motivational framework for AI-assisted reading assessment in EFL/ESL classrooms: a practitioner case study

Article References: Almirabi, M. (2026). Can AI make reading more engaging? A motivational framework for AI-assisted reading assessment in EFL/ESL classrooms: a practitioner case study. Frontiers in Education, 11, 1908499. https://doi.org/10.3389/feduc.2026.1908499

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1908499

Keywords: artificial intelligence, EFL reading, reading comprehension, student motivation, formative assessment, generative AI, multiple-choice quizzes, optical mark recognition, learning management systems, make, reading, engaging