Teaching with AI, Not Against It: A Thinking Partner Framework for Working Adult Learners in Compressed Online Formats

Synthesis: Packard, Fuller, Holmes and Kayanuma (2026) describe the Thinking Partner Framework, built and revised across six semesters of six-week, fully asynchronous graduate research methods courses serving roughly 440 K-12 teachers. Instead of banning generative AI or policing it with detection tools, the design positions AI as a Socratic partner and requires students to document the dialogue: checkpoint submissions reached a 96% average completion rate across the six semesters. The paper's sharpest claim is an integrity reframing — once learning processes are visible, integrity becomes alignment between claimed and actual process rather than a matter for surveillance — while its evidence stays honestly where it belongs, in systematic observations from one instructor's practice rather than measured effects.

Key Findings

1. The prohibition-versus-permission binary fails the students these programs serve. The authors argue that banning AI is unenforceable for working professionals who already use it in their jobs, prevents the very competencies graduate study should build, and rests on tools that Weber-Wulff and colleagues (2023) found inconsistent across fourteen instruments, with Turnitin missing about a quarter of unedited AI text.
2. Detection-based policy carries documented bias. Liang et al. (2023) found the same tools disproportionately flag writing by non-native English speakers and neurodivergent students, which the authors convert into a policy question: not "Did you use AI?" but "How did you use it, and what did you learn from it?"
3. Five design principles are presented as one system. Transparency, process documentation, authentic assessment, scaffolded autonomy, and metacognitive reflection are meant to reinforce each other — clear expectations, visible process, professional tasks, fading support, and reflection that outlives the course — rather than to be adopted one at a time.
4. Checkpoint documentation was near-universal, though depth varied. Across six semesters averaging 73 students each (about 440 in total), checkpoint completion averaged 96%, submissions ran 300 to 600 words, and later iterations with explicit prompt templates produced longer and visibly more substantive reflections.
5. Grades spanned 57% to 96% under unlimited resubmission. Because the course lets students correct and resubmit for the highest score, the authors read the upper end as a mastery-oriented grading policy rather than as evidence of learning gain, and note that improvement across attempts may reflect AI use they did not track.
6. Prompt templates target four moves. Concept clarification, research design thinking, article evaluation, and application to the student's own classroom — for example, "Instead of defining it for me, help me figure it out by asking me questions about what I already know and where I'm confused."
7. Some students carried the questioning into their own classrooms. Field notes record a special education teacher using AI-assisted literature search for an IEP meeting, a science teacher presenting a critical evaluation of a district-mandated curriculum to her principal, and several students reporting Socratic questioning with their own pupils.

Why the compressed format drives the design

The six-week format is an institutional constraint, not a pedagogical preference: these students cannot absorb fifteen-week semesters alongside full-time teaching. The authors cite Price (2024), who found accelerated-course students performing at least as well as traditional-length students on effort-based assessments while topic mastery declined, and Holzweiss, Polnick and Lunenburg (2019) on the focused intensity compressed online programs create for motivated adults. Their conclusion is that compression requires redesign rather than compression, and that AI dialogue extends instructional presence into hours an instructor cannot cover — the 10 p.m. window after a teacher has finished grading her own students' papers. The course also reorients research methods toward intelligent research consumption rather than doctoral-style production, since these students will sit in faculty meetings where administrators propose AI mandates and vendors quote accuracy rates.

Reciprocal transparency is the mechanism, not a policy add-on

Documentation here runs in both directions: students develop metacognitive awareness of their own process, and instructors see where misconceptions arise and how understanding develops. The authors distinguish this from a disclosure requirement — a student who documents how Socratic dialogue clarified the difference between correlation and causation has produced evidence of learning, not compliance. It also matches professional habit, since these adults already document lesson plans, patient notes, and project reports. The design leans on Black and Tomlinson (2025), who found that students documenting AI use mostly distinguished AI-assisted editing from their own contributions, and on process-product assessment models that evaluate the decisions students make during human-AI interaction (Awadallah Alkouk and Khlaif, 2024). Under this account a student who accurately documents substantial AI assistance can still hold integrity, while one who conceals AI use and claims independent work does not.

Socratic dialogue as the interaction pattern

The framework's core pedagogic move is to instruct students how to question rather than to detect when they have not. The authors ground it in Fakour and Imani (2025), who found ChatGPT functioning as a Socratic tutor, and in Favero et al. (2025), whose Socratic chatbot — designed to pose questions instead of answers — produced significantly higher scores than conventional chatbots (p < .001). Supporting evidence is assembled for the surrounding design: Cai, Msafiri and Kangwa (2025) reviewed 158 empirical studies on AI-supported self-directed assessment and motivation; Gurung et al. (2025) found AI plus human tutoring outperforming AI alone on critical thinking; Escalante, Pack and Barrett (2023) found no significant difference between AI and human tutor feedback on learning outcomes, which the authors use to justify reserving instructor time for evaluative judgment while AI handles structural feedback on drafts. Knowles's (1984) andragogy and Adarkwah's (2024) extension of it into the generative AI era supply the adult-learning rationale for prompts that begin from the student's own classroom problem.

What six semesters showed, and what they cannot show

The reported observations are systematic but not controlled: instructor field notes, review of checkpoint submissions, and course evaluations, with no comparison group and no standardized measures. Engagement varied in the way the authors expect. Some students wrestled with questions, documented confusion, and returned to concepts repeatedly; others optimized for the fastest path through required interactions, which the authors treat as rational triage by adults juggling teaching, family, and graduate study rather than as failure. Because assignments could be resubmitted until the score was maximal, the grade range (57% to 96%) is a policy artifact as much as a learning signal, and the authors say so. The transplant evidence is anecdotal by their own description — individual teachers applying frameworks in IEP meetings, district curriculum review, and their own questioning — and they close by calling for controlled comparisons, analysis of the checkpoint corpus, longitudinal follow-up, and equity research they did not conduct.

What this means for practice

• Instructors. Replace prohibition language in the syllabus with documented process, and teach the questioning pattern explicitly: publish Socratic prompt templates, require checkpoint submissions that paste the initial prompt and describe what changed in the student's understanding, and grade the reasoning shown in that documentation rather than the polish of the product.
• Administrators. Rewrite policy around "how did you use AI, and what did you learn" instead of "did you use AI," retire detection-tool contracts, and expect course redesign to cost roughly 15 to 20 hours per course before any ongoing adjustment; fund that time rather than assuming goodwill covers it.
• Faculty developers. Treat AI capability as a prerequisite for teaching it: run hands-on development and standing communities of practice, and keep the distinction between AI and instructor roles explicit in course communication, because immediate AI replies can make expert feedback look redundant.

Limitations

• The evidence comes from one instructor's routine teaching at one institution across six semesters, with no comparison group, no standardized outcome measures, and no student identification; the designer of the AI integration is also its evaluator.
• The 96% figure measures checkpoint submission, not the depth or honesty of the reflection inside it, and the authors report that depth varied widely.
• De-identified course records collected through normal educational practice meant no IRB approval was required, so there is no consent, attrition, or demographic detail against which to test differential effects.
• Access equity is unexamined: the authors assume working adults have sufficient connectivity while acknowledging that phone-only access, aging hardware, and first-time AI exposure are unmeasured, and they recommend building equity assessment into future implementations from the start.
• Tool use is not tracked and the paper names ChatGPT and Claude without reporting which model versions students used or over what period, so the observations cannot be tied to a specific generation of AI capability.

Citation

Packard, A., Fuller, R., Holmes, G., & Kayanuma, B. (2026). Teaching with AI, Not Against It: A Thinking Partner Framework for Working Adult Learners in Compressed Online Formats. Online Journal of Distance Learning Administration, 29(3).