From One-Size Texts to Tailored Readings: Student Experiences with AI-Generated Course Materials

Course and participants

This study was conducted in Fall 2025 in LEAD 503: Organizational Systems and Human Resources, a masters-level course in an educational leadership program. It is a fictitious identifier to reduce deductive disclosure while preserving the instructional design features relevant to the study. The course focused on examining, synthesizing, and applying human resources and organizational management practices necessary for leadership in educational settings. Learning outcomes emphasized applying organization theory to educational environments, cultivating safe and productive working and learning climates, conducting recruitment and selection processes, and developing skills for performance evaluation.

LEAD 503 was delivered in a hybrid format combining in-person and synchronous online sessions. The course integrated applied simulations, case-based discussion, and structured policy-analysis activities that provided repeated occasions for students to translate reading-based knowledge into decisions and written products.

The study’s primary evidence about student perceptions comes from an end-of-course survey administered after course grades were submitted. For LEAD 503, 24 students completed the survey and consented to participation, including consent for analysis of AI-related assignments that documented interaction with the course assistant. The course enrollment may have been larger than the survey respondent count, but the analytic sample for this paper is limited to consenting survey respondents in LEAD 503.

Class Companion: platform and configuration

The Class Companion is a course-tutor bot built using OpenAI’s Custom GPT feature within ChatGPT and configured to generate weekly readings, answer syllabus questions, test comprehension via scenarios, and provide assignment guidance while maintaining academic integrity. The system is explicitly framed as fallible (“powered by AI and can make mistakes”) and is designed to remain course-focused and to defer to the instructor when uncertain. See Appendix 2 for bot behavior instructions.

The content-generation framework directs the bot to write extended readings (target 2,000 words) for working professionals in California higher education, emphasizing immediate applicability, explicit learning objectives, and structured elements such as reflection questions and practical exercises. The engagement style is interactive, with Socratic prompts and requests for examples from the student’s work setting. The configuration specifies boundaries including redirection of off-topic requests and refusal to complete assignments directly. The full instruction set appears in Appendix 2.

Assistant role and integration with the course schedule

The intervention consisted of an AI-powered “Class Companion” embedded as a standing course resource. Students were instructed to use ChatGPT or an equivalent major AI platform, and they were encouraged to set up access in a way that enabled consistent functionality across the term. The Class Companion was positioned as a built-in tutor and course guide. Its intended functions included: (a) generating weekly readings aligned to the syllabus sequence, (b) answering questions about course requirements and deadlines, (c) assessing student understanding through scenarios and questions, and (d) providing assignment support by suggesting approaches and strategies.

For the purposes of this paper, the focal design feature is the weekly reading generator. Weekly readings were not ancillary; they functioned as the core “text” for the course. Readings were aligned to the topical sequence: recruitment and selection; governance and decision-making structures; policy development and implementation; information systems and communication flows; employee relations; crisis management and organizational communication; stakeholder relations and community engagement; performance management and employee development; strategic communication and public relations; organizational change and systems improvement; and leadership and organizational culture in higher education. In class, students repeatedly used these readings as shared preparation for discussion and as a conceptual substrate for applied tasks. Early in the term, class time included explicit demonstrations of the reading workflow and the reading submission format, followed by small-group discussion activities structured around the AI-generated readings.

Readings were operationalized as “reading logs” submitted to the learning management system as PDFs. Students were instructed to request the week’s topic, generate a reading, and ask at least three follow-up questions before submitting. This requirement treated reading as an interaction rather than a one-shot consumption activity. Weekly postings were assigned course points (3 points each, 33 points total) but were not formally graded for quality, functioning as a preparation mechanism rather than an evaluative essay.

Definition and operationalization of “dual tailoring”

The course design emphasized a specific form of individualization that this study refers to as dual tailoring. The term is a descriptive design label, not a new theory. It names an instructional configuration that combines two well-established dimensions of individualization within a single AI-mediated workflow. Dual tailoring has two dimensions.

1. First, interest tailoring refers to adapting the reading to the learner’s professional context, sector, and topical curiosities. In a graduate leadership course, “interest” is often expressed as a preference for a particular institutional setting (community college, CSU, UC, private nonprofit), a professional role (department chair, HR analyst, dean, student affairs administrator), or a problem type (hiring, conflict resolution, governance breakdowns, crisis response). Interest tailoring can be implemented through changes in examples, cases, and applied prompts without changing the conceptual backbone of the reading.

2. Second, comprehension-level tailoring refers to adjusting the accessibility of the reading to students’ readiness and background knowledge. In practice, this can involve altering pacing, adding definitions and intermediate steps, reducing vocabulary density, providing additional scaffolding (advance organizers, summaries, check-for-understanding questions), or offering a deeper “second layer” when a student requests more nuance. Importantly, comprehension-level tailoring does not imply remediation. In graduate professional learning, it often means enabling students to calibrate depth and explanation to the cognitive demands of their work context and prior exposure to HR, policy, and organizational theory.

Both forms of tailoring were supported by the same mechanism: iterative prompting. Students could request a first-pass reading and then ask targeted follow-ups to reframe, deepen, simplify, localize, or apply. The instructor controlled key constraints through the reading prompt template and course norms, including expected length, required elements (theory plus applied examples plus reflection questions), and the expectation that students should engage in multiple follow-up queries rather than accept a single generated text.

Instructional norms, verification expectations, and accountability

The course adopted an explicit AI use policy that normalized AI as part of the learning workflow while preserving student responsibility for quality, veracity, and originality. Verification norms were reinforced through specific assignments, including a “policy compliance bot” task requiring accuracy checks, and through the reading workflow itself, which expected students to request explicit topics and constraints, ask clarifying follow-ups, and treat readings as inputs to higher-control activities (case analysis, simulations, writing tasks) assessed by the instructor. These applied activities served as a quality-control layer where misconceptions surfaced through discussion, peer critique, and instructor feedback.

Data sources

Two primary sources inform the analysis: (1) an end-of-course survey with closed-response items and open-ended responses, and (2) weekly reading logs submitted as PDFs. The survey supplies student-perceived value, acceptability, and risk perceptions. See the actual instrument in Appendix 1. The logs supply two distinct but linked artifact types produced in normal course activity: the AI-generated reading text and the student–AI interaction that generated, revised, and extended that text. The reading logs are 4,487 pages total, or about 939,000 words.

Reading logs as dual artifacts

Each weekly log is a combined record with two analyzable layers. The first layer is the generated “reading,” a textbook-style narrative with headings, learning objectives, summary points, and applied exercises. The second layer is the interaction trace, typically alternating student prompts and Class Companion responses. This structure enables two kinds of claims: claims about properties of the generated instructional text itself (for example, sector anchoring, role framing, definitional density) and claims about adaptive behavior in response to student prompts (for example, elaboration after “explain more,” re-scoping after “that is not my job,” or requests for cases and sources).

A minority of PDFs contained generated text with little visible follow-up interaction; these segments are usable for artifact-level analyses but excluded from analyses requiring prompt-response adjacency.

The artifact corpus includes logs B through K. Log A is excluded because it used chat links rather than the standardized PDF format, and mixing capture formats would introduce avoidable measurement noise.

Survey sampling and measures

Survey analysis is restricted to consenting respondents who completed the survey items tied to Class Companion use and perceived outcomes. Closed-response items are treated as descriptive measures of perceived utility and acceptability, rather than as latent constructs supporting strong inference. Open-ended responses are treated as qualitative evidence of perceived benefits, concerns, and self-reported norms for responsible use.

The survey was a purpose-built descriptive instrument designed for this specific course context, not a previously validated scale. Items were constructed to capture perceived utility, acceptability, and boundary awareness rather than to measure latent constructs. Internal consistency (Cronbach’s alpha) is reported only as a diagnostic for aggregation where item blocks plausibly cohere. This approach is consistent with the study’s exploratory, proof-of-principle scope.

Quantitative analysis

Quantitative reporting is descriptive. For closed-response items, analysis focuses on response distributions and summary statistics. Where blocks of items plausibly form a coherent index, internal consistency can be reported as a diagnostic (for example, Cronbach’s alpha) to justify aggregation for descriptive purposes. Exploratory contrasts by usage intensity are only appropriate if the survey includes an interpretable and non-sparse usage measure; when that condition is not met, usage-based subgroup claims are omitted.

Qualitative analysis

Qualitative analysis uses a focused thematic approach guided by the paper’s conceptual claims. Coding targets (a) perceived value and acceptability (clarity, preparedness, relevance, credibility), (b) risk awareness and skepticism, and (c) perceived boundaries such as reliance concerns or demands for instructor oversight. Code development follows an iterative process: an initial pass to generate candidate codes, consolidation into a compact codebook aligned with the Findings structure, and a confirmatory pass to ensure coverage and identify deviant cases. In a single-course pilot, transparency and traceability of excerpts matter more than elaborate coder infrastructure; agreement checks can be applied to a subset if multiple coders are used, but the primary emphasis is on auditability of claims through short, attributable excerpts.

Artifact analysis

Artifact analysis addresses two adaptation dimensions that are structurally encouraged by the bot configuration: interest tailoring and comprehension-level tailoring. Interest tailoring is operationalized through indicators of sector anchoring, role framing, topic emphasis, and between-log differentiation. Comprehension-level tailoring is operationalized through triggered increases in scaffolding after student prompts that request explanation, definition, or elaboration. The detailed operational definitions and any sampling strategies used for tractability are reported in the corresponding Findings subsections to minimize redundancy. One element warrants advance notice: between-log differentiation is assessed in part through TF-IDF cosine similarity, a standard computational measure of vocabulary overlap between documents. Readers unfamiliar with this technique will find a full explanation alongside the results in Finding 2. Table 1 summarizes the operational definitions used for both dimensions (see Table 1).

Integration and excerpt selection

Integration uses joint displays and paired narrative evidence. Tables provide log-level indicators, while excerpts document mechanisms that counts alone cannot show. Excerpts are selected using maximum-variation logic, prioritizing short passages with unambiguous interpretive warrants.

Ethics and confidentiality

The study involves a small class setting and artifacts that can contain rich professional context. Deductive disclosure is therefore a primary risk. All excerpts are de-identified, unique workplace identifiers are removed or generalized, and log labels are used instead of student names. Where a passage is distinctive enough to plausibly identify a participant, the passage is omitted or paraphrased while retaining the analytic claim through alternative evidence. The study was reviewed and approved by the university’s Institutional Review Board prior to data collection.

Data availability. The survey instrument is provided in Appendix 1. De-identified summary data from the survey are available from the corresponding author upon reasonable request. The full reading logs cannot be shared publicly because they contain professional context sufficient to enable deductive identification of participants in a small class setting.

Finding 1: Student-perceived value and acceptability

Students generally treated the Class Companion as a practical support tool. Across five rated functions (4-point effectiveness scale), ratings clustered in the upper range (Table 2). The strongest perceived value was assignment support: 83% rated the tool “Effective” or “Very Effective” (M = 3.38, SD = 0.77). Providing information about course requirements was also rated highly (88% Effective/Very Effective). Generating quality reading materials was positive but more mixed (75% Effective/Very Effective), suggesting that students accepted the readings as “good enough” while noticing variability. The lowest ratings were for assessing knowledge and skills (71% Effective/Very Effective), the only function with any “Not effective” selections.

On broader acceptability items (4-point agreement scale), most respondents endorsed the overall learning value and future willingness to use a similar tool. Seventy-five percent agreed that they learned more than they would have in a comparable course without an AI companion (29% fully agree; M = 2.96, SD = 0.91). Seventy-one percent said they would take another course with an AI class companion or equivalent (29% fully agree; M = 2.92, SD = 0.93). The strongest endorsement was skill development: 88% agreed that their AI skills increased significantly, and two-thirds fully agreed (M = 3.50, SD = 0.83). These results suggest that acceptability was not only about convenience. Students experienced the intervention as an opportunity to build transferable competence in using AI systems for academic work, which likely contributed to their willingness to see the approach repeated.

Open-ended comments sharpen how students connected value to clarity, preparedness, relevance, and credibility. On clarity and preparedness, multiple respondents described the tool as a fast, responsive “second set of directions” that helped them interpret assignments and organize work. One student emphasized the value of targeted explanation, noting that it “helped explain assignments and textbook readings,” and another framed it as a time-saver that reduced friction in week-to-week preparation. These remarks align with the stronger quantitative ratings for course requirements and assignment completion, and they reinforce an interpretation of the assistant as a just-in-time support layer rather than a substitute instructor.

Perceived relevance appeared in two forms. First, students valued the ability to tailor examples to their professional context and local setting. A respondent explicitly requested that the tool use “examples related to Sac State and the California education system,” implying that relevance was not merely topical alignment with the syllabus but contextual specificity. Second, students appreciated the ability to adapt the reading experience to their needs, including summarizing, re-explaining, or extending content. At the same time, a few comments suggested diminishing returns when the tool expanded beyond what the student asked for, for example describing it as “wordy” and inclined to “keep adding information.” That minor friction is consistent with the more mixed reading-material ratings: usefulness was common, but not uniform.

Perceived credibility was a conditional stance rather than a rejection. Students articulated that the tool was helpful but should not become an unquestioned authority. One wrote, “I do not think we should get used to using it for everything,” while still calling it “a great tool for learning.” Another wanted clearer sourcing, noting that citing authors directly would make it “more credible.” These comments signal bounded trust: students accepted the tool for scaffolding and navigating tasks but remained alert to limits and wanted cues supporting verification.

Finding 2. Evidence of interest tailoring

Interest tailoring was a recurring and observable feature of the AI-generated reading logs. It appeared both as overt customization moves (explicit prompts to adapt examples, cases, or exercises to the student’s setting) and as embedded contextualization (institutional names, sector-specific constraints, and role-relevant tasks). Table 3 summarizes these indicators across reading logs B–K, using a systematic sampling and a set of operational markers that translate qualitative differences into comparable numeric signals.

Method and operationalization were designed to balance feasibility with defensible inference across a very large corpus. Rather than hand-coding the full set of logs, we used a systematic automated 1/5-page sample from each log, producing samples ranging from 13,728 to 22,034 words per log (Table 3). Within each sample we counted a dictionary-based set of “tailoring markers” and normalized them per 10,000 words. These markers included second-person role anchoring (for example, “in your role”), direct adaptation offers (for example, “Would you like me to tailor…”), and concrete situational prompts (for example, “describe your case”). We also counted explicit sector cues via institution-type references, including “community college,” “CSU,” and “Sacramento State,” again normalized per 10,000 words (Table 3). Finally, to estimate how different the logs were from one another in overall language use, we computed TF-IDF vectors for each sample and calculated cosine similarity across logs, reporting each log’s average similarity to the rest and its minimum and maximum similarity values (Table 3). Cosine similarity provides a compact estimate of textual proximity: higher values indicate that two samples rely on more similar vocabularies and phrasing patterns, while lower values indicate greater divergence in word choice and emphasis.

Three patterns in Table 3 support the claim that interest tailoring was not merely anecdotal, but structurally present and measurably variable. First, the overall rate of tailoring markers was substantial, with a mean of 52.24 markers per 10,000 words (SD = 13.73), and a wide observed range from 38.74 (Log K) to 74.29 (Log B) (Table 3). In plain terms, customization language was frequent, but not uniform. Some logs leaned heavily into prompts, personalization offers, and role anchoring, while others relied more on generic exposition.

Second, sector focus varied in ways consistent with the interest-tailoring hypothesis. Log B is the clearest community-college-centered case, with 33.43 community college references per 10,000 words (Table 3). In contrast, most other logs are framed as “CSU/regional public,” showing higher Sacramento State reference rates (e.g., Log F at 9.53 per 10,000 words). The AI selected plausible organizational subunits and missions fitting each institutional context, not merely changing labels.

Third, role framing varied across logs. Some are anchored in HR and employee relations (Logs E and H), others in IT and information systems (Logs C and K), and several in communications and public relations (Logs D, F, G, I, J) (Table 3). Logs often made role framing explicit through conversational repair: in Log F, the student rejects a compliance-focused action plan, triggering a reframing to a nonprofit student-success role.

Representative cases show these dimensions co-occurring. In crisis communication materials, the AI embeds sector variation by offering parallel examples differing by institution type, paired with role-relevant tasks. In stakeholder engagement materials, the AI switches to a UC system lens when prompted, expanding into institution-specific narratives driven by the student’s selected case.

The similarity results help quantify “degree of differentiation” beyond marker counts. Average cosine similarity across logs ranges from 0.50 to 0.61 (Table 3). Minimum similarity values fall as low as 0.41 (Logs B and H), indicating that at least one pairing for those logs diverges notably in vocabulary and emphasis, while maximum similarities reach 0.78 (Logs F and I), indicating that some logs share substantial shared language patterns (Table 3). Interpreted conservatively, this pattern supports a hybrid structure: a common instructional backbone across logs, with meaningful contextual overlays that shift the center of gravity toward a given sector and role. The result resembles a template with adjustable dials. The template keeps the course-aligned scaffolds stable, while the dials tune examples, institutional references, and practical prompts toward the student’s declared setting.

As a robustness check, we recomputed similarities under alternative preprocessing choices: with and without stopword removal, and using bigram rather than unigram TF-IDF vectors. The pattern of results remained stable across these alternatives. Average similarity values shifted by no more than 0.03, and the rank ordering of logs by average similarity did not change, supporting the conclusion that the differentiation pattern is not an artifact of a particular preprocessing decision.

Finding 3. Evidence of comprehension-level tailoring

The reading logs contain repeated sequences where a student asks for clarification, expansion, or a definition, and the Class Companion responds by increasing scaffolding. In these episodes, scaffolding takes a recognizable form: explicit definitions, paraphrases (“in simple terms”), concrete examples, stepwise breakdowns, and occasional check-for-understanding prompts that ask the student to apply the idea to their local context.

Finding 4: Boundaries, failure modes, and student verification responses

Student verification behaviors

Student responses to these boundary conditions were visible in the interaction logs. Across 837 student turns extracted from logs B–K, 23 turns (about 2.7 percent) contained explicit risk-aware moves using conservative markers (corrections of AI assumptions, requests for evidence or cases, statements of confusion functioning as comprehension checks). The distribution was uneven: two logs contained no marker-coded turns at all, while most logs clustered in a narrow band of roughly 2 to 4 percent. This pattern indicates that a critical stance toward AI output was present but not uniform, supporting the claim that such stance is learnable and improvable rather than an automatic byproduct of using AI-generated text.

The clearest evidence of verification is direct correction of AI assumptions. In Log B, the student writes, “that is inaccurate I work in student housing at UC Davis,” triggering a reset and revised framing. In Log F, the same stance appears as role-bounding: “this is inaccurate, that is not my job.” Both cases illustrate that the risk in personalization is not only factual error but category error: the assistant may generate plausible text fitting the wrong job function. The students interrupt that failure at the prompt level.

A second, distinct form of critical stance is the demand for concrete, externally anchored evidence. In Log I, the student asks, “can you please show me actual case studies?” The phrasing is instructive. The student does not merely request more examples; they request actual cases, which implicitly raises a standard of warrant and traceability beyond generic vignettes. In the broader context of AI-generated instructional text, this kind of request is one of the few observable behaviors that directly increases epistemic quality, because it pressures the system toward specificity that can be checked against institutional reports, peer-reviewed literature, or publicly available documentation.

A brief sequence from Log I illustrates the verification cycle concretely. The student prompts: “Can you please show me actual case studies?” The Companion responds with a narrative about a named university’s crisis communication strategy, citing general circumstances but providing no document title, date, or retrievable source. The student follows up: “Can you give me the sources for these case studies?” The Companion then supplies a list mixing plausible-looking citations with at least one that could not be independently confirmed. This three-turn sequence captures the core dynamic: the student raises the evidentiary standard, the system attempts to comply, and the result is partially but not fully traceable, leaving the student in a position where further external verification is needed.

A third form is the explicit comprehension check. In Log E, the student writes, “I have no idea what you are talking about.” Functionally, this is a safeguard against unearned assent: it forces re-explanation and stepwise progression, which is precisely the mechanism through which comprehension-level tailoring becomes observable (Faber et al., 2024).

Pedagogically, this suggests that “critical thinking toward probabilistic text” can be taught as a situated literacy rather than asserted as a general virtue. In this intervention, the teachable unit is the move, not the attitude: correcting a mismatch in sector or role; requesting a checkable case; asking for citations that can be traced; and stating confusion early enough to trigger re-scaffolding. The logs show that some students already performed these moves while others did not, which implies room for explicit instruction, modeling, and peer norming grounded in authentic excerpts from the course’s own interaction archive.

Dual tailoring as a distinctive instructional mechanism

The dual-tailoring framework addresses the longstanding gap between the pedagogical value of individualized instruction and the practical barriers to implementation at scale (Bloom, 1984; Cronbach & Snow, 1977). Prior personalization approaches typically adjusted along a single dimension, such as difficulty level in adaptive problem sets or topical examples in intelligent tutoring systems (Walkington & Bernacki, 2014). Fan et al. (2025) found that engineering students recognized efficiency gains from generative AI but expressed concerns about accuracy and domain-specific reliability, a pattern closely mirrored in this study’s bounded trust findings. This intervention demonstrates simultaneous adjustment along both dimensions described earlier.

The evidence confirms that this tailoring was structural, not cosmetic. Readings shared a common instructional backbone but diverged meaningfully in vocabulary and emphasis, as confirmed by cosine similarity analysis. The result resembles a template with adjustable dials that tune examples and institutional references without abandoning course-aligned learning objectives. Scaffolding increased in response to student signals of confusion, delivered just-in-time rather than predetermined by the instructor.

What makes this mechanism distinctive is the pairing of generative capacity with iterative prompting. Students were required to ask follow-up questions, creating repeated opportunities to request reframing and test understanding. Reading became an interaction rather than a consumption activity, with implications for the visibility of comprehension gaps that would otherwise remain latent.

The hidden pedagogy of verification

A finding that extends beyond the initial research questions is that the intervention functioned as an implicit training ground for critical engagement with probabilistic text. The corpus contains repeated instances of students correcting the assistant’s assumptions, requesting concrete evidence, and refusing to proceed when explanations became unintelligible. These moves were not uniformly distributed across students, which creates instructional leverage. When some students enact verification behaviors in visible ways, instructors can make those moves explicit, name them as desirable practice, and build routines that transfer them to peers. This finding aligns with Corbitt et al.’s (2024) observation that graduate students vary in how much they trust even instructor-provided claims, and extends it by showing that AI-generated text can make verification a recurring, observable practice rather than an occasional disposition. The broader educational community has recognized this challenge: a recent collective manifesto on generative AI in education calls for critical engagement rather than passive acceptance of AI outputs (Bozkurt et al., 2024).

The primary risk of AI-generated text is not that it is always wrong but that fluency can mask uneven quality. The AI-generated reading model distributes verification responsibility: students must learn to treat generated text as provisional and to recognize when claims exceed the evidence provided. Survey data confirm that students adopted this stance. They valued the readings for clarity and preparedness but also expressed caution about dependence and requested instructor oversight. This is bounded trust, not blanket acceptance, and it is arguably a more productive epistemic posture than the unexamined authority that commercial textbooks often enjoy.

The intervention suggests replicable design principles. First, treat generation as draft production; the instructor must review outputs and correct weak citations. Second, maintain explicit norms requiring follow-up questions and verification. Third, use applied activities (case analysis, simulations) as the quality-control layer. Fourth, position the assistant as a fallible guide rather than an authoritative text. Fifth, build iterative refinement into the workflow so that outputs improve across weeks.

A broader hypothesis emerges from these findings. Commercial textbooks hide their limitations beneath consistent formatting and publisher prestige; AI-generated text makes quality variance visible because the source is explicitly probabilistic. If we take seriously the idea that students should learn to evaluate claims and seek evidence, then a reading format that requires verification may prove more pedagogically productive than one that naturalizes acceptance through professional presentation. This remains a hypothesis, not a finding of this study.

Limitations and boundaries of the claims

The study documents one particular way of introducing AI-generated readings in a single graduate course with a small sample and limited artifact capture. The findings cannot support claims about universal effectiveness or transferability to undergraduate courses, large lectures, or disciplines with different epistemological norms. Transferability to undergraduate or large-enrollment settings is untested and constitutes a priority for future research, given the differences in student autonomy, prior knowledge, and the instructor’s capacity for individual oversight that characterize those contexts. The survey data are self-reported perceptions, which are consequential for acceptability but do not establish learning gains relative to a control condition. The study does not support causal claims about learning outcomes. The artifact analysis uses conservative marker-based counts and systematic sampling, which provide defensible lower-bound estimates but may undercount subtler forms of tailoring and critical engagement.

The intervention also relies on instructor time for prompt design, output review, and verification scaffolding. That time cost is lower than authoring multiple textbook-length documents from scratch, but it is not zero. Instructors without experience in prompt engineering or without institutional support for AI integration may find the approach difficult to implement with fidelity. The findings are therefore best interpreted as feasibility evidence rather than as an endorsement for wholesale replacement of textbooks without corresponding changes in instructional infrastructure and faculty development.

Finally, the course took place in Fall 2025, at a moment when generative AI capabilities and institutional norms around AI use were still evolving rapidly. What is feasible and acceptable today may shift as platforms change, as students develop more sophisticated strategies for using AI, and as institutions develop clearer policies about AI-generated content in academic work. The boundary conditions documented here should be treated as context-specific rather than universal.

Implications for future work

The next phase of this work addresses the most consequential limitation: the uneven quality and weak sourcing visible in the Fall 2025 logs. In the subsequent semester, the instructor used ChatGPT’s Deep Research workflow to assemble a curated set of course-aligned reference materials with explicit sourcing, then incorporated those materials into the Class Companion’s knowledge base. This redesign shifts the generation model from predominantly prompt-driven text creation toward retrieval-augmented generation anchored in vetted sources. The expectation is that this approach will reduce over-specific unsourced claims, improve citation quality, and increase traceability while preserving the dual-tailoring capacity that students valued.

Broader implications extend to AI-integrated curriculum design. If dual tailoring is viable in one graduate course, it may be viable in other professionally oriented programs where relevance to work context is a persistent challenge. Future work should examine scalability to larger courses, transferability across disciplines, and whether the bounded credibility stance students adopt toward AI-generated text influences their approach to other authoritative sources. The findings here do not prove that AI-generated readings are superior to commercial textbooks, but they provide enough evidence to take seriously the possibility that a flexible, context-aware reading format may be a better fit for how graduate professional learning actually works.