Design rationale
For supervisors, examiners and research collaborators. Thesis: Generative AI and the Organisation of Dental Education (Design-Based Research; pragmatism; institutional theory). This is Iteration 0 (content version 0.3.0); every element is a draft for review.
1. Purpose and the causal chain
The platform exists because of a specific causal chain. Every page, feature and piece of text must be traceable to it.
- Demand. Accreditation bodies are under pressure to specify how AI should be incorporated into dental curricula (Study 1).
- Gap. An international competency framework has existed since 2022 (Schwendicke et al., 2022), yet curricular implementation remains patchy and supplementary (Aura-Tormos et al., 2026).
- Binding constraint. Study 1 indicates that the limiting factor appears to be educator capacity, not student readiness or technology. Most surveyed faculty report no routine use of AI in assessment, and in some samples fewer than 20% understand machine-learning fundamentals (Study 1, Tier 3).
- Institutional risk. Standards that specify student outcomes without specifying institutional conditions may operate only as exhortation (Study 1 conclusion). Under isomorphic pressure (DiMaggio and Powell, 1983), schools may respond ceremonially: documenting AI integration without changing practice. This is the decoupling of formal structure from activity (Meyer and Rowan, 1977).
- Intervention. The platform is designed to build substantive educator capacity and to generate evidence that distinguishes substantive integration from ceremonial integration.
- Return to accreditation. That evidence informs what accreditation standards can reasonably require of institutions, not only of students. This closes the loop back to the accreditation dilemma.
Consequence for design: The platform must never reproduce the decoupling it studies. Any feature that could generate documentation without practice (certificates, completion badges, "evidence packs" built from self-ratings) is out of scope.
Unit of analysis
The thesis studies the organisation, not the individual learner (the education–learning distinction in the methodology chapter). The platform therefore works at two levels:
- Individual practice: how one educator uses AI in teaching. This is the entry point.
- Institutional conditions: the regulative, normative and cultural-cognitive conditions that enable or constrain that practice (Scott, 2014). This is the analytic focus.
Individual data are only ever a route to understanding institutional conditions. Aggregated, cohort-level views (Part B) are the research unit.
Place in the design-based research programme
- Proposed placement: Study 7 (faculty development), drawing on Study 1 (scoping review) and the REAL-AI framework paper. [AYMAN TO SUPPLY: confirm placement with supervisors]
- Status: Iteration 0, a conjectural prototype. Studies 2 (interviews) and 3 (survey) are in progress. Their findings must be able to revise the design principles, instruments and scenarios without code changes. This is why all content moves to versioned data files with provenance (A1).
- Validation path: design principles and instrument items go to the Delphi study (Study 5); field use in the pilot (Study 6).
2. Philosophical commitments
Bracketed codes A1–A10 refer to the implementation steps of the v3 design brief (A1 content architecture, A3 capacity profile, A5 accreditation bridge, A6 practice log, A7 AI coach, A8 this page, A9 research folder). SC1–SC7 are the scenarios.
| Commitment | Source | Design consequence in the platform |
|---|---|---|
| Pragmatism: ideas are tested by their consequences in practice | Dewey (1938); Hoadley and Campos (2022) | Practice evidence is structured as intention → enactment → consequence (A6). Reflection prompts ask what would count as evidence in the educator's own setting. |
| Consequential validity | Messick (1998) | Instruments are judged by what their use leads to, so each output states what it should and should not be used for (A3). |
| Institutional theory: three pillars | Scott (2014) | A second instrument profiles institutional conditions across regulative, normative and cultural-cognitive pillars (A3). |
| Decoupling: formal structure vs actual activity | Meyer and Rowan (1977) | No certificates or badges. "Documented" and "enacted" are kept visibly distinct (A6). One scenario confronts the inspection-pressure dilemma directly (SC3). |
| Isomorphism | DiMaggio and Powell (1983) | The accreditation bridge (A5) shows that the demand comes from accreditation pressure, and warns against copying "AI integration" displays from other schools. |
| Boundary object | Wenger (1998); Star and Griesemer (1989) | The same content is presented to three audiences (educators, programme or quality leads, researchers and examiners), each with a view suited to them (A5, A8). |
| DBR: iterative, theory-generating, documented design decisions | McKenney and Reeves (2019); Obczovsky et al. (2025) | Conjecture map, design principles with provenance, an iteration log, and content versioning (A1, A9). |
| REAL-AI framework | Khalifah et al. (REAL-AI paper, under review) | The four pillars structure the practice instrument and the scenarios. Pillar definitions must be the paper's verbatim definitions (A1). |
3. Design requirements
| ID | Finding | Source | Design requirement |
|---|---|---|---|
| DR1 | Educator capacity, not technology, appears to be the binding constraint | Study 1, Tier 3; Discussion | The platform targets educators, not students, and treats capacity as institutionally conditioned. |
| DR2 | Standards specifying student outcomes without institutional conditions act as exhortation | Study 1 Conclusion | Profile institutional conditions alongside individual practice (A3). |
| DR3 | The only controlled study testing explainable AI in novice clinicians found automation bias | Study 1, Tier 4 [AYMAN TO SUPPLY: Table 1 ID] | Scenario SC1; the coach invites critique of its own output (A7). |
| DR4 | The only RCT in the corpus found no transfer of AI training to unaided performance | Study 1, Tier 4 [AYMAN TO SUPPLY: Table 1 ID] | Scenario SC2; practice evidence asks about performance without the tool (A6). |
| DR5 | Foundational AI literacy has the most defensible evidence base; applied clinical competencies do not yet | Study 1 Conclusion | Coach and scenarios prioritise literacy-level change and avoid overclaiming applied outcomes (A7, SC4). |
| DR6 | No included intervention reached behaviour change in practice or patient outcomes | Study 1, outcome-level analysis | The platform captures enacted practice and its consequences, not only attitudes (A6). |
| DR7 | The risk of ceremonial compliance under accreditation pressure | Meyer and Rowan (1977); Study 2 Theme 2 (documentation) | No certificates; a decoupling probe item; scenario SC3 (A3, A6). |
| DR8 | Design must be revisable from Studies 2–3 and the Delphi | DBR methodology chapter | All content in versioned data files with provenance (A1). |
4. Conjecture map (Iteration 0)
Conjecture map from the v3 brief, Section 4, following the conjecture-mapping approach (Sandoval, 2014).
High-level conjecture
If dental educators engage with evidence-grounded AI dilemmas and record how their practice actually changes, within a structure that makes institutional conditions visible, then programmes will develop substantive (not ceremonial) AI-integration capacity that accreditation can meaningfully recognise.
Embodiment
- E1 Two-level profile: individual practice (REAL-AI) and institutional conditions (Scott's pillars).
- E2 Evidence-anchored scenarios, each with a visible provenance note.
- E3 Reflection before rating.
- E4 A reflexive AI coach that is transparent about its limits and invites critique of its own output.
- E5 Practice evidence structured as intention → enactment → consequence, with no certificates.
- E6 An accreditation bridge showing where the demand comes from.
Mediating processes
- M1 Educators articulate their own reasoning before seeing expert judgement (E3).
- M2 Educators distinguish what they can change from what depends on institutional conditions (E1, E4).
- M3 Educators critically evaluate AI output, including the coach's (E2, E4).
- M4 Educators plan, enact and review concrete changes rather than only documenting intentions (E5).
Outcomes
- O1 Enacted changes to teaching or assessment practice, with consequences recorded.
- O2 Shifts in the institutional-conditions profile at cohort level over time.
- O3 Reduced gap between documented and enacted AI integration (decoupling probe).
What would count as the conjecture holding; to be operationalised in Studies 6–7.
Conjectures to test. Each pairs an embodiment and a process with an outcome: E3→M1, E5→M4→O1, and E1→M2→O2. These are the questions the pilot must answer, and the iteration log records evidence for or against each.
5. Design principles
Statements and provenance are from the v3 brief (Section 5). Each rationale is assembled only from the findings of the design requirements it cites (research/design-requirements.md) and is for Ayman to review or rewrite.
DP1 Capacity is institutional
Support individual educators in ways that make their institutional conditions visible and discussable.
Assembled from the cited requirements. DR1: educator capacity, not technology, appears to be the binding constraint. DR2: standards specifying student outcomes without institutional conditions act as exhortation.
DP2 Enactment over documentation
Count only what is enacted and reviewed as evidence; never produce artefacts that could stand in for practice.
Assembled from the cited requirements. DR6: no included intervention reached behaviour change in practice or patient outcomes. DR7: the risk of ceremonial compliance under accreditation pressure.
DP3 Judgement before authority
Educators state their own reasoning before the platform's or the AI's view is shown.
Assembled from the cited requirements. DR3: the only controlled study testing explainable AI in novice clinicians found automation bias.
DP4 The tool must model what it teaches
The AI coach discloses that it is AI, states its limits and invites critique.
Assembled from the cited requirements. DR3: the only controlled study testing explainable AI in novice clinicians found automation bias. DR5: foundational AI literacy has the most defensible evidence base; applied clinical competencies do not yet.
DP5 Claim no more than the evidence supports
Prioritise foundational literacy, and frame applied outcomes as open questions.
Assembled from the cited requirement. DR5: foundational AI literacy has the most defensible evidence base; applied clinical competencies do not yet.
DP6 Every element is traceable
Every item, scenario and principle shows where it comes from and its validation status.
Assembled from the cited requirement. DR8: design must be revisable from Studies 2–3 and the Delphi.
DP7 Demand comes from accreditation
Make the link from accreditation pressure to capacity explicit without reducing practice to compliance.
Assembled from Section 1.1 of the brief. Accreditation bodies are under pressure to specify how AI should be incorporated into dental curricula (Study 1); under isomorphic pressure, schools may respond ceremonially.
6. Iteration log
Each entry records what changed in the design, why, and any evidence for or against the conjectures in the conjecture map (E3→M1, E5→M4→O1, E1→M2→O2).
Iteration 0 (v3): conjectural prototype from Study 1 and the REAL-AI paper; supersedes v1–v2 researcher-drafted content
- Date: 1 October 2026
- Content version: 0.3.0
- Status: draft; not yet used in the field
Changes from v2
- A1 Content architecture. All framework content moved from code into versioned JSON files in
content/, each object carrying an ID, version, status, provenance and Arabic-review status. Pillar descriptions are shown as working definitions until the published definitions are supplied. Content changes are recorded incontent/CHANGELOG.md. - A2 "Why this?" panels. Every scenario and instrument section shows its provenance in plain language, its validation status and its version.
- A3 Capacity profile. The readiness check became a two-level capacity profile: the 16 practice items (draft instrument v0.3) and a new institutional-conditions instrument (nine items across Scott's three pillars, two of them reverse-scored, plus a separately reported decoupling probe). Results show the two profiles separately with a fixed "How to read this" box; no combined score exists. The CSV export and the development-plan request include both profiles.
- A4 Scenario library. Seven scenarios in two groups, each with its own reflection prompt. SC1–SC4 are new and evidence-anchored; SC5–SC7 keep their v2 text and gain provenance. The v2 radiograph scenario is retired, superseded by SC1. Reflection before rating is unchanged; each saved reflection now records the exact prompt and scenario version shown.
- A5 Accreditation bridge. New
/accreditationpage relating the REAL-AI pillars to accreditation themes, with the GDC-alignment framing for Saudi programmes and a fixed caution that outputs are not accreditation evidence. - A6 Practice log. "My record" became the practice log, with a new practice-change entry (intention → enactment → consequence → evidence type) and status Planned → Enacted → Reviewed. No certificates, badges, completion percentages or evidence packs. The printout is headed "Personal development record — not accreditation evidence."
- A7 Reflexive coach. The system prompt now states that the coach is an AI system, limits its claims to what the evidence supports and separates individual action from institutional conditions. Every reply shows the model name and a "How useful was this?" rating, stored with the entry.
- A8 Design rationale page. New English-only
/rationalepage for supervisors, examiners and research collaborators, linked from the footer. - A9 Research documentation. New
research/folder: conjecture map, design requirements, design principles, instrument codebook and citation register (generated fromcontent/bynpm run research:docs), this iteration log, and the Part B ethics and data design (ethics-and-data.md, design only; no server-side storage). - A10 Home page. Hero reframed around accreditation and educator capacity ("Building the capacity that accreditation assumes"), with the REAL letter animation kept, a link to the accreditation page under the pillars, and three steps: profile your practice and conditions; work through evidence-based dilemmas; log what you change, and what happens.
Revisions within Iteration 0
- 0.3.1 (2 October 2026). Accreditation page: removed the Saudi GDC-alignment paragraph at Ayman's request (it did not convey meaning), departing from brief requirement A5; added the GDC and NCAAA platform URLs.
Evidence on the conjectures
None yet. Iteration 0 has not been used in the field. The pilot (Study 6) and Study 7 are expected to provide evidence for or against E3→M1, E5→M4→O1 and E1→M2→O2.
7. Limitations
- Both instruments are draft self-report measures. Validation is planned through expert review (Delphi) and psychometric testing.
- Scenario options and their fit ratings reflect the researcher's reading of the REAL-AI framework and Study 1. They are design conjectures for review, not settled judgements.
- The evidence base in Study 1 is dominated by attitude surveys and small pilots, so several scenarios address open questions rather than established findings.
- The AI coach can be wrong. Its use in a platform about critical AI use is deliberate, but it carries the risk of the automation bias the platform warns against. User ratings of coach feedback are collected to examine this.
- Arabic text is a draft translation pending expert review.
- In this iteration, no data leave the browser except anonymous score summaries and reflections sent to generate feedback.
8. References
Sources from the citation register that are cited on this page, author–date (Manchester-Harvard). Sources marked V were added in the v3 brief and must be verified before any thesis use. [AYMAN TO SUPPLY: full Manchester-Harvard references from the thesis reference list]
- Aura-Tormos et al. (2026).
- Dewey (1938).
- DiMaggio and Powell (1983).
- Hoadley and Campos (2022).
- Khalifah et al. (REAL-AI paper, under review).
- McKenney and Reeves (2019).
- Messick (1998).
- Meyer and Rowan (1977).
- Obczovsky et al. (2025).
- Sandoval (2014): conjecture mapping, Journal of the Learning Sciences. [V: verify before thesis use]
- Schwendicke et al. (2022).
- Scott (2014).
- Star and Griesemer (1989): boundary objects, Social Studies of Science. [V: verify before thesis use]
- Study 1: Khalifah et al., scoping review (under review).
- Wenger (1998).