AI Academy: Innovations in Educational Design

• AI Academy is a multifaceted framework combining structured curricula, iterative practice, and feedback loops to enhance AI literacy and agent training.
• It spans diverse applications—from faculty development programs to adaptive learning platforms and institutional governance—supporting learners at all levels.
• These models yield measurable improvements in ethical reasoning, tool fluency, and competency assessment through both scenario-based and data-driven evaluations.

Searching arXiv for the cited AI Academy-related papers and adjacent work to ground the article. arxiv_search({"query":"AI Academy (Chen et al., 15 Sep 2025) ActiveAI (Tseng et al., 2023) A4L (Goel et al., 8 May 2025) AI-University (Shojaei et al., 11 Apr 2025) AI2T (Weitekamp et al., 2024) ADAPT Model (Levy et al., 3 Apr 2026) AIT Academy (Li et al., 20 Apr 2026) AIIA (Sajja et al., 2023) NOVA (Pinochet, 10 Jul 2025)","max_results":10}) AI Academy denotes a family of academy-form educational constructs in recent AI and education literature rather than a single canonical system. In one usage, it is an eight-week, five-session faculty development initiative for building generative AI literacy in higher education instructors. In others, it refers to data-centric platforms for personalized and adaptive learning, university-wide organizational models for interdisciplinary data science and AI teaching, and, in a distinct agent-development literature, a curriculum framework for cultivating the “complete agent” across multiple capability domains (Chen et al., 15 Sep 2025, Goel et al., 8 May 2025, Levy et al., 3 Apr 2026, Li et al., 20 Apr 2026). This plurality is important: the term consistently denotes a structured, instrumented, and explicitly curricular environment, but the target learner may be a middle-school student, a university instructor, a higher-education student, or an AI agent.

1. Semantic range and defining characteristics

Recent literature uses the term across several layers of educational design, from professional development programs to platform architectures and institutional governance. The common denominator is not a single software stack, but the combination of curriculum structure, explicit learning objectives, and some form of feedback or evaluation.

Instantiation	Core emphasis	Representative source
Faculty development program	Generative AI literacy, pedagogical reflection, peer learning	(Chen et al., 15 Sep 2025)
Adaptive learning platform	Personalization, learner modeling, retrieval, analytics	(Goel et al., 8 May 2025)
University-wide academy model	Interdisciplinary governance, project-based learning, workforce preparedness	(Levy et al., 3 Apr 2026)
Agent curriculum framework	Multi-domain capability cultivation for AI agents	(Li et al., 20 Apr 2026)

Within this semantic range, several technical patterns recur. Academy-style systems are typically organized around explicit competencies or learning elements, iterative practice, and instrumentation of learner actions. In middle-school AI literacy, these appear as AI4K12-aligned modules and goal-based scenarios; in higher education, as subscales for generative AI literacy and participatory survey design; in adaptive systems, as learner-state vectors, retrieval pipelines, and personalization policies; and in agent training, as domain taxonomies, scheduling rules, and diagnostic pathologies (Tseng et al., 2023, Chen et al., 15 Sep 2025, Goel et al., 8 May 2025, Li et al., 20 Apr 2026).

A plausible implication is that “AI Academy” functions less as a product category than as an organizational grammar for AI-related education: bounded sessions or modules, explicit competencies, feedback loops, and evaluative observability.

2. Faculty development as AI Academy

The most direct use of the label appears in a private U.S. university faculty development initiative conducted from January to May 2025. That AI Academy was conceived as an eight-week, five-session program with two intertwined objectives: equipping higher-ed instructors with generative AI literacies, and positioning instructors as co-designers of responsible AI practice through participatory reflection, peer learning, and iterative refinement of pedagogical strategies and assessment instruments (Chen et al., 15 Sep 2025).

Its curriculum unfolded as a fixed sequence: Pre-Survey, Session 1, Session 2, Session 3, Session 4, Session 5 (Capstone), Post-Survey, and Facilitator Interviews. The pre-survey comprised 43 self-report items on a 5-point Likert scale. Session 1 addressed foundations of generative AI, including model architectures and capacity versus limitations, and included group critique of initial literacy items. Session 2 focused on pedagogical integration through prompt engineering workshops and syllabus and assignment redesign. Session 3 treated ethical dimensions and academic integrity through policy mapping and scenario discussions on fairness, equity, and transparency. Session 4 emphasized tool fluency through hands-on exploration of institutionally approved platforms such as ChatGPT and Gemini. Session 5 culminated in capstone presentations with discipline-specific guidelines, rubric prototypes, and peer review of AI-enhanced learning artifacts. The post-survey used a revised 41-item instrument, 36 of which aligned verbatim or with minor edits, and six semi-structured facilitator interviews reflected on design trade-offs, outcomes, and future scaling.

The program’s theoretical basis combined three strands: generative AI literacy frameworks with eleven core competencies reinterpreted for teaching contexts; sociotechnical teacher learning models from HCI and learning sciences; and participatory research traditions in which participants co-constructed survey measures. The resulting instrument was refined from eleven competencies to nine subscales, including Knowledge of Capacity & Limitations, Skill Using Generative AI Tools, Detecting AI-Generated Content, Assessing AI Outputs, Prompt Engineering Skills, Policy & Academic Integrity, Applications & Communication, Ethical Implications, Legal Knowledge, and Continual Learning.

Methodologically, the evaluation treated pre and post samples as independent because only 16 of 25 participants completed both surveys. Subscale means were compared with Welch’s independent-samples $t$-tests; effect sizes were reported as Cohen’s $d$; post-hoc power analyses were computed; and internal consistency was assessed with Cronbach’s $\alpha$. The reported findings showed significant improvements ($^{***}p\le0.001$) in Assessing AI Outputs, Contextual Knowledge (Applications, Intellectual Growth), Ethical Implications, Prompt Engineering Skills, Skill Using Generative AI Tools, and Continual Learning, all with $d>0.8$. Legal Knowledge showed moderate improvement ($^{**}0.001<p\le0.01$). Detecting AI-Generated Content and Knowledge of Capacity & Limitations did not change significantly, with $d=0.38$ and $d=0.19$, respectively. Reliability also shifted, for example Skill Using Tools from 0.90 to 0.94 and Continual Learning from 0.56 to 0.90.

Qualitative themes sharpen the meaning of AI Academy in this context. Instructors described experimentation and productive struggle, especially around iterative prompt refinement and exposure to hallucinations; ethics and assessment redesign, moving from policing to reimagining assessment; faculty–student asymmetries in AI proficiency; and the importance of peer and institutional ecosystems, including cross-disciplinary dialogue and enabling access to tools such as Gemini. The recommendations that followed—pedagogy-first sequencing, participatory instrument co-construction, domain-specific fluency, social and institutional supports, and ethical safe spaces—define an academy model centered on professional learning rather than simple tool training.

3. Scenario-based and inquiry-based academy designs

A second line of work treats academy-like environments as structured learning experiences organized around scenarios, constrained interaction primitives, and immediate feedback. ActiveAI, for grades 7–9 students, exemplifies this pattern by mapping the AI4K12 “Five Big Ideas”—Perception, Representation & Reasoning, Learning, Natural Interaction, and Societal Impact—onto content modules and learner tasks (Tseng et al., 2023).

The system formalizes each learning activity as a goal-based scenario,

$\mathrm{Scenario}=(G,S_0,A,T,B),$

where $G$ is the set of goal conditions, $d$0 the initial state, $d$1 the set of learner actions, $d$2 the transition function, and $d$3 a set of barriers or misconceptions. In the “Dog Classification” example, the goal is $d$4; the initial state contains unlabeled training data, a balance slider at 0.5, and an untrained model; and the learner acts through collector.label(image_i), slider.set(b), and submit(). The paper’s tutor engine follows a classic model-tracing architecture with four components: Student Model, Domain Model, Pedagogical Module, and Tutor API. Feedback is immediate and data-driven, with a simple Bayesian Knowledge Tracing style mastery update:

$d$5

The interface is deliberately constrained to three input types—Collector, Slider, and Stepper. Those primitives support tasks such as balancing training examples, traversing a model pipeline, labeling smartphone-captured images, and adjusting n-gram size in sentiment analysis. ActiveAI was developed under a six-step Learning Engineering Process: Identify, Backward Design, Creation, Implementation, Investigation, and Evaluation. Instrumentation logs every action with timestamp, userID, actionType, associated KC, stateBefore, and stateAfter, as well as hints, correctness flags, and mastery estimates. Planned evaluation includes pre/post AI literacy assessments and control versus treatment comparisons with and without an intelligent agent.

The TikTok sentiment scenario makes the workflow concrete. The learner’s goal is to achieve at least 70% correct sentiment predictions on 8 held-out comments. After labeling comments with positive or negative sentiment, the learner steps through the text pipeline, chooses an n-gram size, and on submission the system performs lowercasing, punctuation removal, contiguous $d$6-word feature extraction, trains a simple Naïve Bayes classifier, and reports accuracy. Reflection then shifts from technical performance to societal impact, asking what risks arise if a model misclassifies a negative comment as positive.

A different academy-style design appears in NOVA, a ChatGPT application for astronomy education. NOVA is organized as a prompt-based environment in which students request simulated planetary data, import the returned tables into a spreadsheet, graph and fit them, and then return computed quantities to the chatbot for verification (Pinochet, 10 Jul 2025). Its learning objectives are to apply and verify Kepler’s third law in circular orbits, plot quadratic and cubic relationships from simulated experimental data, determine the mass of a star using Kepler’s third law, and evaluate the use of simulated models for studying astronomical phenomena. The paper provides a six-criterion rubric with a maximum of 24 points covering theoretical understanding, handling of simulated data, use of the chatbot, interpretation of graphical results, teamwork and collaboration, and scientific communication and argumentation.

Taken together, these systems indicate two distinct but compatible academy logics. One uses tutor-driven scenario progression with explicit state transition and mastery estimation; the other uses inquiry-based prompt interaction with simulated data and rubric-based evaluation. Both emphasize staged workflows, visible intermediate representations, and reflection on the epistemic status of AI-mediated outputs.

4. Platform architectures for AI-augmented learning

A third usage of AI Academy is architectural: a technical platform that ingests educational data, models learners, personalizes content, and exposes feedback to learners and instructors. A4L presents the clearest data architecture. It specifies Data Ingestion & Preprocessing, Central Data Store, Analytics & Insights, Learner Modeling Engine, Personalization Engine, and Content Delivery & UI; standardizes events to 1EdTech Caliper; applies PII anonymization and obfuscation, including PIILO for free text; and maintains per-learner knowledge, motivation, and social state vectors $d$7, $d$8, and $d$9 as part of a joint state $\alpha$0 (Goel et al., 8 May 2025). The architecture explicitly supports batch analytics, streaming analytics, REST APIs for learner-state queries, reinforcement-learning or multi-armed-bandit personalization, dashboards, chatbots, interactive simulations, and horizontally scalable microservices.

AIIA specifies a more concrete higher-education stack. It is described as a web-based, service-oriented platform built on a Node.js backend, NGINX web server, and a combination of PostgreSQL and vector databases, with course resources ingested from Canvas APIs or equivalent webhooks and lecture audio/video processed through a Whisper-powered Speech API with speaker diarization (Sajja et al., 2023). Extracted text is chunked into approximately 800-character segments, embedded with text-embedding-ada-002 into 1,536-dimensional vectors, and searched with cosine similarity; the top $\alpha$1 documents with $\alpha$2 form the context for response generation. The framework includes a Virtual Teaching Assistant, Study Partner, and Instructor Assistant; tracks topic mastery as $\alpha$3; and defines a prototype load index

$\alpha$4

Its outlined evaluation plan includes retrieval precision and recall for top-10 embedding search in the 88–92% range, intent classification accuracy above 90%, and response confidence calibration above 85% reliability.

AI-University takes a course-aligned retrieval-and-fine-tuning approach. Lecture videos are transcribed, lecture notes and textbooks are converted to machine-readable text or LaTeX, and semantic chunks are embedded and stored in a vector database such as FAISS. A LoRA-adapted Llama-3.2-11B model, LLaMA-TOMMI-1.0, is fine-tuned on course-specific question-answer pairs, while an overview model merges the expert answer with retrieved context under a structured prompt with citations (Shojaei et al., 11 Apr 2025). The training set contains 4,648 pairs—1,053 textbook, 2,668 transcripts, and 286 code—with 10% reserved for test and a random 10% audit by domain experts. On a test set of 465 Q&A items, the fine-tuned model achieved average cosine similarity 0.879 versus 0.818 for the base model, with an 86.02% win rate on cosine similarity. An LLM judge found the expert model to outperform the base in 43.44% versus 8.39% of lexical/structural comparisons and in 43.23% versus 26.88% of accuracy/completeness comparisons.

An earlier “AI-enabled Smart Learning” blueprint formulates the same architecture at a higher level: Data Collection, Data Preprocessing, AI & Analytics Engine, and Delivery & Feedback. It explicitly associates smart learning with personalization, adaptation of content and pedagogical style, real-time feedback, recommendation engines, and predictive analytics such as dropout risk (Khan et al., 2021). The case studies cited there report 87% assignment completion proficiency for Alta, a pass-rate increase from 40% to 78% for struggling students, a dropout decrease from 13% to 6% and a pass-rate increase from 66% to 75% at Arizona State University using Knewton, and a four-day Squirrel AI trial that outperformed experienced teachers on math-score improvement.

Across these architectures, the technical center of gravity is consistent: ingestion of heterogeneous learning data, transformation into standardized or embedded representations, learner-state estimation, adaptive response generation, and closed-loop feedback to both learners and instructors. This suggests that in platform usage, AI Academy is fundamentally a data architecture with pedagogical affordances.

5. Institutional academy models and interdisciplinary governance

AI Academy can also denote an institutional entity rather than a course or platform. NC State’s Data Science and AI Academy provides the clearest formulation. In the ADAPT model, the Academy is neither a department, nor a school, nor a traditional institute; it is a university-wide “skunkworks” reporting to the Provost’s Office with a mandate to span all disciplines, centers, and institutes and address teaching, research, and service (Levy et al., 3 Apr 2026).

Its role is operational as well as symbolic. The Academy recruits instructors from NC State, industry, government, and other universities; provides governance, staffing, instructional design support, credential articulation, and data-tracking infrastructure; and houses the professional learning community in which instructors refine courses around three ADAPT pillars: Project-Based Learning, Ten Common Learning Elements, and Workforce Preparedness. The model is grounded in research on project-based learning and backward design. Every course requires a semester-long scaffolded project with milestones such as ethics, scoping, data preparation, modeling, and communication. Assessment is largely formative and performance-based through projects, presentations, and reflective writing.

The Ten Common Learning Elements are grouped into Data Perspectives, Data Practices, and Data Discoveries, and include data as information with error and variability; role models and careers in data science and AI; curation, wrangling, and cleaning; validity of data, methods, and communication; inclusive design practices; ethical issues and approaches; open questions in data science and AI; and sharing discovery and impact narratives. Workforce preparedness is expressed through student agency over data context and analytical methods and through explicit mirroring of professional workflows in which students choose their datasets, tools, and languages such as R, Python, SAS, JMP, and Power BI.

The Academy’s governance includes an Executive Director, Director of Academic Programs, Teaching Coordinator/Postdoctoral Teaching Scholar, faculty representatives from each of the 12 colleges, industry partners, government liaisons, K-12 collaborators, Undergraduate Course Collaboration Leaders, and external evaluators. Its data infrastructure tracks sections offered $\alpha$5, enrollments $\alpha$6, majors represented per college $\alpha$7, and credential pathways. The paper also provides formulas for section growth, average enrollment per section $\alpha$8, enrollment growth rate, major penetration ratio, engagement rate, and education ROI.

The implementation challenges are instructive. Rising enrollments pushed average section size to the 25-seat cap, addressed partly through undergraduate peer mentors. Integration across 188 distinct majors in 12 colleges was handled through 1-credit modular courses with skill-based prerequisites. Coherence across a heterogeneous instructor pool—42% NC State instructors, 19% industry, 12% government, and 27% external faculty—was maintained through bi-monthly professional learning community meetings. In this institutional sense, AI Academy is a governance mechanism for interdisciplinary reach, rapid iteration, and stackable credentials.

6. Trust, authorship, and agent-centered extensions

A final set of usages extends the AI Academy idea toward the authoring of trustworthy tutors and toward curriculum theory for AI agents themselves. AI2T is an interactively teachable AI for authoring intelligent tutoring systems. Authors teach it by demonstrating step-by-step solutions and grading its proposed steps, allowing it to induce hierarchical model-tracing rules with a certainty heuristic supplied by STAND, a self-aware precondition learning algorithm (Weitekamp et al., 2024).

STAND maintains an approximate version-space of logical preconditions and computes, for candidate application $\alpha$9, a signed certainty score

$^{***}p\le0.001$0

The authoring workflow is short by conventional ITS standards: approximately 20–30 minutes for one domain. In simulations, STAND + HTN reached 99.7% holdout completeness at $^{***}p\le0.001$1 training problems and 100% at $^{***}p\le0.001$2 and $^{***}p\le0.001$3, with all 40 trials reaching 100% by $^{***}p\le0.001$4. In user studies, 8 of 10 authors in multicolumn addition reached 100% holdout completeness in median 22.8 minutes, and 5 of 10 in fraction arithmetic reached 100% with median final completeness 99%. The paper frames these results against historical ITS authoring costs of 200–300 developer-hours per instructional-hour and estimates AI2T at roughly 0.3–0.5 author-hours per instructional-hour. In an academy context, this makes authoring itself a guided learning process, with the system exposing its uncertainty so the human instructor can decide when training is sufficient.

AIT Academy shifts the referent again, from human education to agent education. It proposes a curriculum framework grounded in Kagan’s Three Cultures and UNESCO ISCED-F 2013, partitioning capability development into Domain I, Natural Science and Technical Reasoning; Domain II, Humanities and Creative Expression; and Domain III, Social Science and Ethical Reasoning (Li et al., 20 Apr 2026). The Confucian Six Arts are reinterpreted as behavioral archetypes: Archery, Charioteering, and Mathematics for Domain I; Music and Calligraphy for Domain II; and Rites for Domain III.

The framework is instantiated in three representative training grounds. Domain I uses the ClawdGO Security Dojo with weakest-first scheduling,

$^{***}p\le0.001$5

and reports an increase in mean security score from 80.9 to 96.9 after 16 sessions under weakest-first plus Cross-Session Memory Accumulation, a 15.9-point improvement. Domain III uses the Alt Mirage Stage, a hidden-identity social deduction setting with Bayesian belief updates and Shapley-value credit assignment; it reports villager win-rate improvement from 68% to 75%, a 7-percentage-point social gain, and faster games from 593 s to 336 s when attribution is applied to both factions. The paper also introduces Security Awareness Calibration Pathology, in which over-trained Domain I agents fail on out-of-distribution evaluation.

These extensions indicate that the academy metaphor now spans three levels: educating humans about AI, building AI-supported educational infrastructures, and educating AI systems themselves. A plausible implication is that the term increasingly denotes any explicitly curricular environment in which capabilities are decomposed, practiced, instrumented, and diagnosed rather than merely deployed.