Research Article
Generative artificial intelligence in doctoral education: a participatory action research evaluation of the Dr. I.D.E.A. model
Synthesis: Padrón Alvarez and Estrada Molina (2026) ran a participatory action research study with 34 doctoral candidates in Cuba who were also university teachers, putting generative AI inside an integrated research-and-teaching development sequence rather than a stand-alone workshop. Self-rated competence rose sharply in research, AI-supported digital work and creative pedagogy — but with no comparison group and nothing but self-report, the design cannot attribute that rise to the intervention. The durable contribution is the operational layer: all 34 put GenAI into a course, three-quarters sustained a documented redesign for at least three consecutive weeks, and nearly as many were still using AI-supported practice at follow-up.
Key Findings
- One integrated sequence produced gains in three domains at once — research, AI-supported digital and creative-pedagogical competence — with the digital domain, starting lowest, gaining most.
- The gains are perceived, not demonstrated. Every quantitative measure is self-report, and the design has no comparison group, so maturation and response-shift bias remain plausible explanations.
- Behavioural evidence is stronger than the competence ratings: all 34 participants integrated GenAI into a course, 26 sustained a documented redesign over at least three weeks, and 25 still used AI-supported practice at follow-up.
- Verification did not get displaced. Indexed-database searching continued alongside weekly GenAI use, which the authors read as evidence that the AI layer supplemented rather than replaced the verification layer.
- A quarter of participants did not sustain the change, which makes follow-up support as consequential as the initial intervention.
What the study did
The Dr. I.D.E.A. model has four layers: active learning in collaborative digital environments; doctoral training as the articulating axis; generative AI as a foundational layer; and an outer participatory action research cycle in which participants co-design, implement and refine. The study was run at a Cuban public technological university with associated institutions, from September 2025 to January 2026, with follow-up to March 2026.
Participants were 34 doctoral candidates in the active predoctoral phase who also teach — 41 were invited, 39 expressed interest, 37 were eligible, and 34 completed consent and all three measurement waves (82.9% of the active population; 91.9% of those eligible). Mean age was 29.8 (SD 4.2); 58.8% were in educational sciences, 26.5% in engineering, 14.7% in computer sciences.
Measurement used three waves — T1 in September 2025, T2 in November after the workshops and guided implementation, T3 in January after a refinement phase — with three composite instruments: MRCS for research competence (28 items), a 29-item digital composite built from SDiCoS plus one adapted MAILS item, and a 44-item creative-pedagogical composite combining K-DOCS and PCS. Three propositions guided the analysis: P1 co-development, P2 transfer to teaching, P3 scientific socialisation.
The dosage was modest: eight synchronous hours plus asynchronous work.
Competence gains, and what they do and do not mean
Every domain moved, and moved early:
- Research competence: M 3.21 (SD 0.74) → 4.43 (SD 0.59), ΔM = 1.22, dₐᵥ = 1.83, 95% CI [1.31, 2.35], p < 0.001.
- AI-supported digital competence: M 2.98 (SD 0.91) → 4.65 (SD 0.68), ΔM = 1.67, dₐᵥ = 2.10, 95% CI [1.52, 2.67], p < 0.001. This domain had the lowest baseline, which the authors note leaves more room for change.
- Creative-pedagogical competence: M 3.29 (SD 0.80) → 4.53 (SD 0.67), ΔM = 1.24, dₐᵥ = 1.69, 95% CI [1.19, 2.18], p < 0.001.
Repeated-measures ANOVA confirmed the time effect in each domain (partial η² between 0.69 and 0.79), and the interval comparisons show gains in both halves of the study — with one exception worth noticing: the creative-pedagogical gain decelerated (T1–T2 dₐᵥ = 0.99, T2–T3 dₐᵥ = 0.69), whereas research and digital gains held steady (0.89 → 0.93 and 1.07 → 1.03).
Change scores across the three domains correlated positively (r = 0.54 to 0.61), which the authors read as support for the co-development proposition — with the caveat that shared self-report method inflates correlations among change scores.
These are perceived competence gains. Effect sizes near 2 SD from an eight-hour intervention without a control group are a claim about self-perception, and response-shift bias — participants recalibrating what "competent" means as they learn — is a live alternative explanation the authors name.
What participants actually did
The study's stronger evidence is behavioral and independently verified:
- 34 of 34 (100%) incorporated GenAI into at least one university course.
- 26 (76.5%) implemented a sustained course redesign — a documented revision of session plans, materials and assessment over at least three consecutive weeks.
- 25 (73.5%) continued AI-supported practices at short-term follow-up.
- 30 (88.2%) had committee-verified substantial dissertation progress; 21 (61.8%) gave conference presentations; 17 (50.0%) submitted manuscripts to indexed journals; 18 (52.9%) initiated new research projects.
Use at the end was routine rather than occasional: 32 of 34 (94.1%) used GenAI at least weekly, with ChatGPT at M = 4.2 sessions per week, Perplexity 2.7, and Gamma or Genspark 1.9. Indexed-database searching continued in parallel — Web of Science M = 3.8 searches per week, Scopus 3.1, SciELO 2.9, ERIC 2.4 — which matters because it shows the AI layer did not displace the verification layer.
Triangulation across the quantitative, qualitative and product strands produced 11 meta-inferences: five classified as agreement and six as partial agreement, with no dissonance. It also surfaced the failures the averages hide: 8 participants (23.5%) never implemented a sustained redesign, 9 (26.5%) did not maintain AI practices at follow-up, 4 (11.8%) showed no substantial dissertation progress, 17 (50.0%) did not submit a manuscript, and AI-assisted qualitative pre-coding remained a conditional use requiring supervision for 47.1%.
The operational layer
The model's transferable content is less the four-layer diagram than the routines that made it work:
- One integrated sequence, one authentic task. Research, AI-supported digital and creative-pedagogical competencies were exercised on the same object — a thesis chapter, a teaching design, a manuscript — rather than in separate workshops.
- Mandatory transfer to practice. Each candidate adapted an AI-supported research or learning activity to one of their own courses and documented the redesign, implementation and reflection.
- Structured peer review with designated academic discussants, giving written critique on authentic products before AI-supported revision, so revisions were socially validated rather than individually adopted.
- A responsible-AI layer as a rubric, not a module: co-constructed categories of authorized, conditional and non-authorized uses, individual AI-use logbooks, and mandatory verification against indexed databases.
What this means for practice
- Programme designers. Integrate rather than stack. Design one development sequence in which research, digital and pedagogical competence are exercised on the same authentic task; separate tool workshops do not reproduce this design.
- Programme designers. Make transfer to teaching a requirement rather than an invitation: every participant redesigns one of their own courses and documents the redesign, its implementation and their reflection.
- Faculty developers. Institute peer critique before AI revision. Designated discussants give written critique on draft products first, which keeps judgment with people.
- Programme designers. Replace a stand-alone ethics module with an operating rubric — authorized, conditional and non-authorized uses, AI-use logbooks, and mandatory verification in indexed databases.
- Programme leaders. Budget for follow-up. A quarter of participants did not sustain the redesign and a quarter did not maintain AI-supported practice.
Limitations
- No comparison group. Single-group participatory action research with a three-wave design: maturation, history, testing effects, response-shift bias, regression to the mean (especially in the lowest-baseline digital domain) and demand characteristics all remain plausible; the changes are associative, not causal.
- All quantitative measures are self-report, capturing perceived rather than demonstrated competence, with social desirability and common-method variance in play; the product-based evidence documents activity, not competence level, and has no baseline.
- Composite indices with uneven construct coverage: the AI-specific content of the digital composite rests mainly on one adapted MAILS item, some K-DOCS items relate only indirectly to pedagogical creativity, and the PCS has only preliminary validation.
- N = 34 is too small for the measurement work the claims imply — Rasch modeling, confirmatory factor analysis and longitudinal invariance testing were all judged infeasible, so the composites are treated as theory-driven indices rather than validated latent scales.
- Insider research on a voluntary sample from one program, with a short follow-up and a modest formal dosage, limits transferability; the authors' own position as program insiders is part of that limitation.
Citation
Padrón Alvarez, A., & Estrada Molina, O. (2026). Generative artificial intelligence in doctoral education: A participatory action research evaluation of the Dr. I.D.E.A. model. Trends in Higher Education, 5(1), 141.