From AI Assistance to Learning Ecosystem Architecture
Summary: Artificial intelligence (AI) can help learners produce stronger work at the point of need. But assisted performance is not the same as transferable capability. This article proposes four design questions for deciding what AI may do, what learners still need to understand, and what evidence can show that learning endures beyond the support.
Imagine a learner using artificial intelligence to search, summarize, compare sources, and suggest a response. The work becomes faster and may become better. Now remove the system. Can the learner explain why the response is sound, recognize an error, adapt the reasoning to a new context, or decide when the recommendation should be challenged?
This is not primarily a question of detecting AI use. It is a question of design. A system can improve performance at the point of need without necessarily developing a stable, transferable capability in the person using it. If digital learning environments do not distinguish between those two outcomes, successful assistance can be mistaken for learning.
When Faster Support Is Mistaken for Learning
The tension predates generative AI. The move from just-in-case training to just-in-time support made information available during the task rather than before it. Electronic Performance Support Systems were already designed to give workers or students the information and tools needed to perform in real time (Maughan, 2005).
That shift was not a mistake. It would be inefficient to precede every operational need with a course or require learners to memorize information that can be retrieved reliably. The difficulty begins when saving training time leads us to treat successful performance as evidence of learning.
AI intensifies the issue because it can do more than retrieve a relevant document. It can construct an answer, compare options, and contribute to a decision. A 2025 survey of 319 knowledge workers found that higher confidence in generative AI was associated with less self-reported critical thinking, while participants described a shift toward checking AI outputs, integrating responses into the task, and overseeing the work (Lee et al., 2025). The study does not show that AI necessarily weakens critical thinking. It does show why we must identify which capabilities the person is expected to exercise.
Support can close an immediate performance gap without closing the capability gap.
The same problem appears when we assume that a learner-centered experience also builds learner autonomy. A platform can recommend resources and make choices easier without helping learners understand why one option is better than another. Ease of use is valuable, but it is not evidence that learners can make those choices independently.
Design the Boundary Deliberately
The solution is not to turn every interaction with technology into a lesson. Sometimes the appropriate goal is simply safe, fast, and accurate performance. At other times, learners need knowledge and judgment that will remain available across tasks and contexts.
Four questions can help faculty, instructional designers, learning leaders, and technology teams make that boundary explicit:
- What is the intended outcome? Decide whether the activity is meant to enable immediate performance, develop a durable capability, or do both. Do not use completion or polished output as automatic evidence of competence.
- What may the system do? Delegation should follow the learning purpose. If searching, comparing, summarizing, or recommending is itself part of the intended capability, learners need opportunities to practice and understand that process before or alongside assistance.
- What must the learner still understand? If AI searches, what criteria should learners use to judge the sources it returns? If it summarizes, how will omissions be noticed? If it recommends, who can challenge it and on what basis? The European Digital Competence Framework, DigComp 3.0, likewise frames digital competence as critical and responsible, not merely operational (Cosgrove & Cachia, 2025).
- What evidence would show that learning endures? Ask learners to explain a decision, transfer a principle to a new case, compare an AI response with another source, revise an output, or complete an unaided checkpoint when independence is part of the outcome. The relevant evidence is not simply whether AI was used, but whether the intended capability can still be demonstrated.
In practice: Consider a university course in academic writing or research methods. If the goal is for students to evaluate scientific literature and build a defensible argument, AI may help generate keywords, map concepts, or suggest an outline. But if literature searching is itself part of the intended capability, students must still carry out and justify the search themselves in academic databases. They should be able to explain why a source is credible and relevant, detect inaccurate or fabricated citations, and justify their thesis. Evidence of learning might combine a critique of an AI-generated literature review with an unaided task or oral defense in which students must apply the same criteria independently. The boundary is therefore set by the capability to be developed, not by the tool itself.
These questions also change how learning data should be interpreted. Completion, clicks, and polished AI-assisted outputs describe activity or system-mediated performance. They become evidence of learning only when they are tied to the intended capability and to a task that makes it observable. Otherwise, dashboards may make an efficient environment look more educationally effective than it is.
From Connected Tools to Coherent Architecture
Learning management systems, learning experience platforms, analytics, skills platforms, and AI can be technically integrated without forming a coherent learning ecosystem. Integration creates connections; it does not determine what those connections should mean for learning. A recent reference architecture developed for the Netherlands armed forces identifies six components of a learning ecosystem: learning content, teaching methods, people, technology, data, and policy and governance—not a technology stack alone (van der Pal et al., 2025).
This is where learning ecosystem architecture matters. It means designing how people, technologies, data, activities, and responsibilities work together, including deciding where technology should support performance and where learners still need to develop capability. Rather than requiring a new job title, this design function can be shared across roles.
Technology can make some actions easier. The key design question is what can be delegated, what humans must still be able to understand and challenge, and which capabilities the environment should develop over time.
The decisive question is not how much the system can do for learners. It is this: what do we want to remain in them when the system’s support is withdrawn?
References
Cosgrove, J., & Cachia, R. (2025). DigComp 3.0: European Digital Competence Framework (Fifth edition). Publications Office of the European Union. https://doi.org/10.2760/0001149
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 1121, pp. 1–22). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713778
Maughan, G. R. (2005). Electronic performance support systems and technological literacy. The Journal of Technology Studies, 31(1), 49–56. https://doi.org/10.21061/jots.v31i1.a.7
Van der Pal, J., Buiel, E., Korteling, H., & Oprins, E. (2025). Architecture for a large scale learning ecosystem. In R. A. Sottilare & J. Schwarz (Eds.), Adaptive instructional systems (Lecture Notes in Computer Science, Vol. 15812, pp. 148–156). Springer. https://doi.org/10.1007/978-3-031-92967-0_10
AI Disclosure
The author used a generative AI assistant to support restructuring, language revision, translation, and source discovery. The article’s central argument draws on the author’s earlier research; the author reviewed and approved the final text and takes full responsibility for its content.
Greta Baldanzi is an educator and former corporate Learning & Development and human resources professional with more than twenty years of experience in training, employee development, and international learning coordination. Her research and writing focus on learning technologies, performance support, digital and AI competence, and learning ecosystem architecture.
