Human-AI collaboration in education: continual learning systems as adaptive instructional partners

Modern educational systems face increasing pressure to respond dynamically to evolving student needs, curricular shifts and rapid changes in instructional technologies (Acar et al., 2025; Akavova et al., 2023). While artificial intelligence (AI) has made significant inroads in adaptive learning platforms (Beck et al., 1996; Holmes and Tuomi, 2022) most existing systems remain static, trained once on historical datasets and unable to evolve in response to new learners, environments or pedagogical goals. This limitation has prompted growing interest in continual learning (CL), a subfield of machine learning focused on enabling models to learn incrementally from a stream of tasks or data without forgetting previously acquired knowledge (Parisi et al., 2019).

Recent advances in CL have expanded its application from traditional machine learning tasks to adaptive intelligent systems that must continuously incorporate new knowledge while preserving prior learning (Wang et al., 2024). As AI systems become increasingly integrated into dynamic environments, researchers have emphasized the importance of incremental learning mechanisms that support adaptation without complete retraining (Yang et al., 2025; Bell et al., 2025). These capabilities are particularly relevant in education, where learner needs, instructional practices and curricular requirements continually evolve.

In the educational context, CL systems are well suited for adaptive technologies that need to respond to ongoing learner feedback, curriculum updates and instructional innovations (Liu, 2017). However, the greatest potential of these systems lies not in replacing human educators, but in augmenting them. Emerging research supports the view that the most effective AI systems will operate within human-AI partnerships where machines provide dynamic, data-informed recommendations and human instructors apply context, ethics and pedagogical judgment. This collaborative model reflects a shift from AI as an autonomous tutor to AI as a co-designer and co-facilitator in the learning process.

Recent studies have examined the growing role of generative AI in education (Barari et al., 2024, 2025). Evaluated large language models across multiple engineering disciplines and Bloom's cognitive levels, demonstrating their increasing capability to support engineering problem-solving and instructional tasks. Similarly (Barari et al., 2026a, b), investigated student performance in programming-based engineering coursework before and after the emergence of generative AI, suggesting that AI tools may help reduce performance disparities by providing additional procedural support and learning assistance. Together, these studies highlight the expanding influence of AI in engineering education and reinforce the need for adaptive, human-centered systems that can evolve alongside advances in AI capabilities.

Research increasingly shows that perceptual experience is shaped by underlying neural computations, with implicit memory representations, formed unconsciously through environmental regularities and contextual knowledge, playing a central role in processing sensory input (Czigler and Winkler, 2010). CL, also referred to as lifelong learning, is a growing area within machine learning that addresses a critical limitation of traditional AI systems: the inability to learn from new data over time without forgetting prior knowledge (Hinton et al., 2015; De Lange et al., 2021). Most conventional models are trained once on a fixed dataset and deployed without the ability to adapt to novel tasks or changing environments, a setup ill-suited for dynamic domains like education. This limitation leads to catastrophic forgetting, a well-documented phenomenon where performance on previous tasks deteriorates as new information is learned (McCloskey and Cohen, 1989).

Recent work by Wang et al. (2024) provides a comprehensive synthesis of CL, integrating theoretical foundations, methodological developments and application domains into a unified framework. The authors emphasize that CL systems must balance the stability–plasticity tradeoff, ensuring that new knowledge can be acquired without degrading previously learned information, while also maintaining strong generalizability across tasks and domains.

Recent advances have expanded CL beyond theoretical machine learning challenges into practical adaptive systems. Applications have been explored in Edge AI (Wu et al., 2026) and aviation systems (Barari and Barari, 2025; Barari et al., 2026c), while studies on representation learning and incremental concept formation have demonstrated promising approaches for mitigating catastrophic forgetting and supporting continuous adaptation (Barari and Kim, 2021; Barari et al., 2025). At the same time, interest in intelligent tutoring systems (ITS) continues to grow, with recent work highlighting their potential to provide personalized support for STEM learners and promote cognitive development through adaptive interactions (Zarestky and Gleason, 2026; Mutalliyevich et al., 2025). Together, these developments suggest that CL offers a promising foundation for educational AI systems capable of evolving alongside learners, instructors, and curricula.

To address this, researchers have introduced a range of CL strategies. Regularization-based methods, such as Elastic Weight Consolidation (EWC) (Kirkpatrick et al., 2017), add a penalty to the loss function that discourages changes to model parameters important for earlier tasks. Replay-based methods, like Gradient Episodic Memory (GEM) (Lopez-Paz and Ranzato, 2017), store a subset of data from previous tasks or generate pseudo-samples using models such as Generative Adversarial Networks (GANs) (Goodfellow et al., 2014). Architectural approaches, including Progressive Neural Networks (Rusu et al., 2016), expand model capacity with each new task while preserving access to earlier representations.

Techniques such as EWC and experience replay help prevent catastrophic forgetting, a challenge where AI models lose previously acquired skills when learning new tasks. By preserving prior knowledge, these methods enhance operational efficiency and safety, enabling more accurate fault prediction, improved routing decisions and adaptability to evolving regulations (Li and Hoiem, 2017).

A comprehensive survey by the authors' prior work categorizes these methods based on theoretical foundations, operational constraints and application domains, highlighting the growing interest in CL for complex, real-world systems. As CL research matures, its applicability in education is increasingly compelling, particularly for adaptive learning platforms, curriculum analytics and real-time instructional support (Barari, 2025).

Sustainable education equips learners to address present and future challenges by rethinking educational goals, methods and collaboration for lasting societal and environmental impact (Lin et al., 2023). AI-assisted systems are designed to support sustainable education in teaching and learning processes by leveraging intelligent algorithms to deliver personalized feedback, recommend content and automate assessment (Hemachandran et al., 2022; Nwana, 1990; Létourneau et al., 2025; Maity and Deroy, 2024; Gobert et al., 2023). These systems range from ITS to learning analytics dashboards and curriculum recommendation engines. However, the majority of such systems operate under static intelligence; they rely on fixed models that lack the ability to evolve with new learner data or shifting educational contexts (Goodfellow et al., 2014).

The paradigm is now shifting toward human–AI collaboration, where AI systems are not positioned as autonomous instructors but as decision-support tools that complement human expertise (Chiu and Rospigliosi, 2025; Yan, 2025). This approach mirrors trends in other high-stakes domains, such as aviation and medicine, where AI is most effective when working alongside humans in a feedback-rich, adaptive environment (Barari and Sanders, 2024, 2025; Barari and Siebold, 2026; Siebold and Barari, 2026). In education, this manifests in several ways:

1. Instructors use AI-generated insights to monitor learner engagement and adjust pacing or instructional strategies.

2. Curriculum designers partner with AI to evaluate long-term trends in learning outcomes and inform iterative content revisions.

3. Students interact with AI tutors that adjust difficulty or delivery style based on real-time responses.

These collaborative models may be strengthened when the AI component employs CL, allowing systems to adapt to local student needs, individual teaching styles and institutional changes over time. As such, the integration of CL into AI-assisted education tools opens the door to systems that co-evolve with the humans they support.

CL systems embedded in AI-assisted tools offer unique advantages in education: they facilitate personalized learning at scale, enable curriculum co-design through longitudinal insights and support real-time instructional adaptation. By moving beyond static intelligence, these systems can align more closely with the naturally evolving nature of teaching and learning. Yet, integrating CL into education also introduces challenges, such as ensuring transparency in adaptive behavior, managing data privacy and maintaining pedagogical alignment. These challenges highlight the importance of human-in-the-loop frameworks, where educators retain agency over decision-making while benefiting from AI's adaptive support. Combining human judgment with machine adaptability enables the design of educational ecosystems that are responsive, sustainable and deeply learner-centered.

We begin by presenting the methodology that guided the development of our AI-enhanced educational models. This includes the conceptual design framework, data considerations and intended scale of implementation. Following the methodology, we describe a pilot study on the integration of instructional chatbots as a foundational practice within university courses. This implementation offers practical insights into the design of human–AI interactions in authentic learning contexts. Building on this foundation, we then introduce two conceptual case studies: an AI-enhanced ITS and an AI-assisted curriculum co-design tool, both leveraging CL to support long-term adaptability and human-in-the-loop collaboration in education.

This study proposes a conceptual design model for integrating CL into AI-assisted educational systems, with the goal of enhancing adaptability, instructional alignment and long-term effectiveness. The model is situated in the context of higher education and envisions a dual-framework architecture that supports both individual learning pathways and program-level curriculum refinement. The proposed system includes two components: (1) an AI-enhanced ITS designed to evolve with each student cohort while supporting instructor oversight and (2) an AI-assisted curriculum co-design tool capable of analyzing longitudinal learning patterns across a pilot program in engineering fundamentals.

While the two proposed systems are conceptual in nature, their design is grounded in prior instructional implementations and longitudinal observations from courses delivered between 2023 and 2025. These experiences informed the architectural choices, human-in-the-loop design and data structures proposed in this framework.

This study is a conceptual and design-oriented framework grounded in prior instructional implementations. As such, formal statistical analysis is deferred to future work focused on empirical validation of these systems.

The tutoring system is envisioned for use in a foundational engineering course with a class size of approximately 30 students. The system would operate as a layer within an online learning management environment, providing adaptive instructional feedback based on a student's progression, performance and common cohort patterns.

At a conceptual level, the tutoring system is designed to:

1. Implement a replay-based CL architecture to retain pedagogically valuable intervention patterns from prior terms.

2. Adapt feedback mechanisms as it receives new data, without discarding effective earlier strategies.

3. Support human-in-the-loop decision-making, where instructors remain responsible for interpreting and applying system recommendations.

Instructors would interface with the system via a dashboard that summarizes:

1. Student progression and knowledge gaps

2. Emerging misconceptions at the class level

3. AI-generated recommendations for scaffolding and intervention

These recommendations are not applied autonomously but are intended to support instructor insight and prompt timely, context-aware teaching responses.

The conceptual model assumes the following data relationships:

1. Individual student progression metrics, including mastery trajectory and learning behavior patterns

2. Cohort-level trends, such as shared areas of difficulty or delayed conceptual acquisition

3. Interaction logs, tracking the evolution of AI-generated support over multiple cohorts

These data sources would inform the system's continual updates while helping preserve previously learned instructional patterns.

A schematic of ITS is shown in Figure 1, and its temporal evolution across cohorts is illustrated in Figure 2. This diagram illustrates the architecture of a CL-based ITS designed for engineering education. It begins with student input and diagnostic assessment, followed by real-time personalization driven by a regularization-based CL model (e.g. EWC). Instructional feedback, hint generation, and adaptive content sequencing are informed by longitudinal learning profiles. Human-in-the-loop oversight allows instructors to review system suggestions, adjust scaffolding and feed insights back into the model for future adaptation. The system balances knowledge stability and learning plasticity to support individualized, evolving instruction across multiple cohorts.

Figure 2 illustrates the temporal progression of the ITS across multiple student cohorts while highlighting the internal learning cycle within a representative cohort. Rather than repeating identical processes for each cohort, the figure emphasizes that the same interaction loop, comprising student input, misconception identification, instructor feedback and model updating, operates consistently over time. Through CL mechanisms such as replay-based memory and regularization, knowledge from earlier cohorts is retained and integrated into subsequent iterations. This results in progressively refined feedback, improved instructional alignment and increased system effectiveness across cohorts.

The second component of the model is intended for use in a pilot engineering fundamentals program composed of multiple aligned courses. The tool is designed to support instructional designers and curriculum coordinators by identifying opportunities for curricular improvement based on historical and current student data.

Conceptually, the curriculum co-design tool would:

1. Use regularization-based CL (e.g. EWC) to integrate new semester data without overwriting key insights from past cohorts.

2. Track content alignment across courses, identify shifts in assessment strategies and detect persistent curriculum gaps.

3. Generate visual summaries and flags that faculty can use to guide review cycles and instructional planning.

The tool does not independently modify curriculum but serves as an analytic and diagnostic aid. Faculty would engage with system outputs in collaborative review meetings, refining the AI's insights with contextual expertise and instructional goals.

The model proposes the following data structures as essential for effective curriculum analysis:

1. Curriculum effectiveness indicators, including pass rates, retention and progression patterns

2. Longitudinal student performance trends, highlighting when and where learning gaps emerge

3. Course-level metadata, such as delivery modality, assessment changes and topical sequencing

These data streams would be analyzed to identify systemic inefficiencies and to inform decisions about course redesign, pacing or scaffolding.

Figure 3 presents a swimlane flowchart illustrating the architecture and data flow of the proposed AI-assisted curriculum co-design system. The diagram is structured across five horizontal lanes, each corresponding to a major functional layer of the model: Data Sources, Data Engineering, AI Core, Analytics Layer and Faculty Collaboration.

The flow begins in the Institutional Data, where five key data sources, including course metadata, learning outcomes, assessment scores, student cohort data and faculty/student feedback, are collected and passed downstream. These raw inputs are channeled into the Data Engineering Lane, where a series of preprocessing modules perform normalization, outcome aggregation and temporal alignment across semesters. The final component in this lane, the Content Drift Detector, acts as a decision node that determines whether structural changes in curriculum delivery warrant human intervention or re-weighting.

If no drift is detected, data proceeds to the AI Core, where the Regularization-Based Learner (e.g. using Elastic Weight Consolidation) incorporates new outcome performance data while preserving past knowledge. The Trend Analyzer identifies long-term underperformance patterns, and the Change Impact Evaluator examines whether previous instructional interventions resulted in measurable improvement.

Insights from the AI Core are passed to the Analytics Layer, where outcomes are flagged, misalignments between assessments and intended learning outcomes are detected, and visual dashboards are generated. These dashboards, combined with system-generated revision suggestions, are then sent to the faculty lane, where instructors review, validate and refine the recommendations.

Finally, human judgments, whether annotations, overrides or confirmations, are fed back into the system to close the CL loop. The flow concludes with a Cycle Complete node, preparing the system to incorporate new data from the next academic term. This visual structure emphasizes both the modular composition of the system and its commitment to human-AI collaboration, ensuring that faculty oversight remains central while leveraging AI for scalable pattern detection and curriculum refinement.

To support CL across academic terms while preserving valuable instructional knowledge, we propose applying the EWC framework to our curriculum co-design model. This approach enables the system to update its performance evaluations and recommendations based on new semester data while retaining insights from previous cohorts.

EWC helps stabilize known-effective curriculum elements, for example, topics that have consistently demonstrated stable outcome attainment patterns over multiple semesters, by discouraging unnecessary changes to well-functioning components. At the same time, it allows the system to adapt to new signals, such as a decline in student performance on concepts like Moment Analysis observed across recent cohorts. This method encourages explainable adaptation, ensuring that each recommended curriculum adjustment is grounded in both empirical trend shifts and prior instructional decisions, fostering a transparent and evidence-based process for ongoing improvement.

The EWC-regularized loss function is expressed as:

Each component of this equation is redefined in the context of curriculum-level CL:

1. $L(θ)$ is the total objective function that combines the new term's learning outcome performance with a stability constraint to prevent overfitting to recent data.

2. $Lnew(θ)$ represents the error between predicted and observed outcome performance for the current term.

3. $θi$ denotes the current learned parameter for outcome $i$⁠, representing the importance or effectiveness weight of a particular curriculum feature (e.g. sequencing of content or method of assessment).

4. $θi*$ is the previously learned parameter for outcome i, typically inferred from a running average or stored from the past semester's model state.

5. $Fi$ is the Fisher information approximation for parameter i, indicating the relative importance of preserving the prior value. This may be empirically estimated based on outcome sensitivity to changes in instructional patterns.

6. $λ$ is a hyperparameter that governs the plasticity-stability tradeoff. A high $λ$ penalizes deviations from historical trends, enforcing curricular consistency, while a lower value allows greater adaptation in response to emerging issues (e.g. a steep performance decline in a key topic like Moment Analysis).

By incorporating this loss formulation, the system mimics a human expert who adjusts instructional design cautiously, preserving well-functioning curricular elements while tuning only those areas where clear, sustained underperformance is detected. In practical terms, EWC ensures that modifications to suggested curriculum revisions are anchored in longitudinal evidence rather than reactive to single-term anomalies.

The methodology described here presents a conceptual architecture for a dual-purpose educational AI system grounded in CL. The proposed model is intended to guide future development efforts that aim to embed adaptive, human-aligned intelligence into instructional technologies. By enabling AI systems to learn over time while collaborating with educators, this approach supports long-term responsiveness in both student-facing and program-level educational processes.

These design principles were further informed by prior AI integrations in our curriculum, particularly through targeted chatbot implementations that provided critical insights into structured, role-based AI–student interaction, laying the groundwork for the pilot study discussed in the following section.

In the next section, we present a pilot study demonstrating the integration of instructional chatbots as a foundational practice within university courses. Building on this foundation, the article then introduces two proposed case studies: an AI-enhanced ITS and an AI-assisted curriculum co-design tool, both leveraging CL to support long-term educational adaptability and human-in-the-loop collaboration.

Before moving into the two proposed case studies on ITS and curriculum co-design, it is important to ground this discussion in an established proof-of-concept for human–AI collaboration in learning environments. Our university has implemented a range of conversational AI agents, custom-built chatbots, that serve as role-specific instructional partners across disciplines. These implementations not only offered targeted skill-building experiences for students but also provided valuable insights into how narrowly scoped AI systems could meaningfully contribute to course learning outcomes.

While the pilot implementations provided qualitative and design-level insights into human-AI interaction patterns, systematic quantitative evaluation and statistical analysis were beyond the scope of this initial study and represent an important direction for future research.

The chatbots were designed to operate within curated scenarios tied to disciplinary practices, ensuring relevance, structure and pedagogical alignment. Importantly, these tools avoided the pitfalls of general-purpose AI agents by embedding tight guardrails, targeted knowledge bases and structured dialogue patterns. Their use in supporting reflection, domain-specific application and communication skill development laid the conceptual groundwork for more advanced and systemic approaches like those proposed in the following case studies.

In particular, lessons learned about scaffolding, hallucination prevention, student motivation and instructor trust directly informed the architecture and human-in-the-loop feedback mechanisms embedded in our intelligent tutoring and curriculum co-design models. These earlier chatbot pilots demonstrated that AI systems, when thoughtfully integrated into instructional design, can amplify human judgment, promote metacognition and support meaningful disciplinary engagement. The next section builds on these foundations to propose two scalable, AI-assisted innovations for continuous improvement in higher education.

Financial-Advising Simulation (Money and Banking): Students play the role of financial advisors while the chatbot adopts a client persona facing a monetary dilemma that hinges on the present-value (PV) formula. The bot's prompts require learners to articulate the PV calculation in plain language, justify their recommendation and respond to follow-up “what-if” questions. After the exchange, students submit both the transcript and a reflective memo, which the instructor grades for conceptual accuracy and sound financial reasoning. Figure 4 captures the interactive environment where students engage with a chatbot simulating a financial client scenario in Money and Banking course. The Canvas-integrated interface supports structured learning, encourages financial reasoning in plain language and facilitates transcript-based reflection assignments.

Reflective Practice Coach (American Literature): Here the chatbot acts as an academic coach that leads a guided reflection on major course projects. A step-by-step prompt script nudges students to connect insights from course assignments with future career applications while asking probing “tell-me-more” questions. Instructors evaluate the depth of critical thinking shown in the submitted transcript, reinforcing metacognition and transfer of learning.

Theory-Persona Dialogues (Globalization and World Politics): Six separate chatbots personify major international-relations theories (e.g. realism, constructivism). Each bot's knowledge base includes core texts, canonical authors and contrasting positions, enabling students to interrogate theories, compare perspectives and test applications to current events. Post-interaction reflections are graded on students' ability to synthesize and apply complex theory in context. Figure 5 displays the Canvas-integrated interface of one of the instructional chatbots used in Globalization and World Politics course.

Project-Management Supervisor (Space Tourism): In this capstone-style scenario, the chatbot embodies the student's boss at a fictional space tourism company. As shown in Figure 6, Max Mustermann, a busy executive who provides just-in-time feedback on project charters, cost estimates and risk registers for a fictional space-tourism firm. The bot's guardrails compel students to surface assumptions, cite evidence (APA) and iterate deliverables. While students utilize Max throughout the term, two required interactions plus a reflective summary count toward project proposal grades.

Table 1 provides a comparative summary of four instructional chatbots integrated into different university courses. Each chatbot is designed with a discipline-specific persona and tailored interaction model to support targeted learning outcomes.

Taken together, the cases illustrate design patterns common to the chatbots produced at our university: (1) a carefully scoped persona anchors the dialogue, (2) a curated knowledge base and conversation flow reduce the propensity for hallucinations and (3) structured post-chat reflection closes the loop between practice and assessment. These design moves are deliberate responses to lessons learned in our early chatbot pilots. Generic, unanchored or unstructured bots yielded shallow exchanges. Grounding our chatbot implementations in an authentic disciplinary role persona drove domain-specific discourse and engagement. Moreover, having a chatbot facilitate student participation in a curated scenario allowed students to apply course concepts to real-world conditions. We also dealt with the issue of hallucinations and overconfidence, prompting us to adopt tightly scoped knowledge bases and guardrails. Further, narrowing the focus of the chatbot (often through a curated persona and scenario) allowed us to adopt a tightly scoped knowledge base for the chatbot and predict and mitigate user behaviors likely to produce hallucinations.

In addition to managing technical limitations, we also had to address challenges in student engagement and usage patterns. Initial student use skewed toward answer getting, leading us to build structured prompt scaffolds that press for explanation, translation across audiences and scenario testing. In response to early student confusion, we also learned to foreground the introduction of the chatbot with a detailed overview explicitly outlining the expected conversational structure, delineating the chatbot's functions and boundaries, and explaining the pedagogical rationale behind the activity and (if relevant) applicability to discipline-specific workplace competencies. Finally, we began requiring transcript as well as reflection artifacts to make thinking inspectable and assessable and to encourage metacognition and critique of chatbot output. Adding a reflection component also helped to address the concerns of instructors who were unsure of how to interpret student thinking within a chatbot conversation or who were reluctant to spend the amount of time necessary to parse through lengthy dialogue sessions.

Building on the insights gained from the chatbot pilot study, we now turn to two proposed case studies that extend the role of AI beyond scripted interactions toward adaptive, continually evolving educational tools. While the chatbots demonstrated the value of structured, persona-driven engagement in course-specific contexts, the following conceptual models explore how CL can enhance ITS and curriculum design processes at scale.

To demonstrate the applicability of the proposed conceptual model, this section presents two case studies drawn from common educational settings in higher education engineering programs. Each case illustrates how CL-based AI systems could support human–AI collaboration in instruction and curriculum refinement, and highlights the potential for sustained adaptability, enhanced instructional clarity and meaningful contributions to long-term student achievement.

This case explores the conceptual design of an AI-enhanced ITS integrated into a first-year engineering fundamentals course with approximately 30 students. The course covers core concepts such as engineering problem-solving. Given the technical complexity and sequential knowledge structure of the course, it offers an ideal setting for exploring how a CL-based tutoring system could support evolving learner needs across semesters.

The envisioned ITS operates alongside the course's learning management system and draws on a replay-based CL algorithm to identify common misconceptions, monitor student progression and adapt its recommendations over time. Importantly, the model does not act autonomously but functions as a decision-support tool for instructors, generating weekly feedback reports that highlight:

1. Persistent knowledge gaps at the cohort level

2. Patterns of stagnation in individual student learning

3. Recommended interventions or scaffolding strategies

Over multiple cohorts, the model is expected to retain instructional insights, such as which scaffolds work best for specific types of problems, while learning from the evolving behavior of new students. For example, if previous cohorts struggled with conceptual heat engine efficiency and benefitted from visual simulations, the system may prioritize suggesting such resources to future instructors when it detects similar struggle patterns. The CL layer ensures that these insights are not overwritten with each new term, enabling the system to become progressively refined over time.

Through its transparent, adaptive design, this ITS aims to develop a sustainable human-AI partnership in instruction, empowering faculty while creating opportunities for early, targeted intervention.

A simplified implementation of Case Study 1 is designed through a prototype for a sample engineering fundamental course, Statics, with concepts like force systems, equilibrium and trusses (See Appendix A for core architecture overview). The tutoring system logs student answers and misconceptions, uses past data (from prior cohorts), and suggests feedback/remediation strategies. The summary of demonstrated features in the script is presented in Table 2.

While the current implementation is presented at a conceptual and prototype level, the system is explicitly designed to support longitudinal deployment across multiple semesters, enabling the accumulation of instructional knowledge and progressive refinement of feedback strategies over time.

The second case involves the proposed deployment of an AI-assisted curriculum co-design tool in a broader engineering fundamentals program comprising multiple interrelated courses. The tool is conceptually designed to support reviewing cross-course alignment, identifying potential systemic gaps and iterating on program structure over time.

The program context for this case study reflects a typical Engineering Fundamentals program, which serves as a two-year foundational pathway for most bachelor's degree programs in engineering. The program consists of approximately 63 credit hours and includes a combination of mathematics and physics courses, along with core engineering and engineering science courses. These typically include Introduction to Engineering, Engineering Graphics, Statics, Dynamics, Solid Mechanics and introductory programming. The courses are intentionally aligned through shared learning outcomes and progressive skill development, forming an integrated curriculum that supports student transition into upper-level engineering coursework. The proposed tool is designed to operate across this multi-course structure, enabling cross-course analysis of student progression, outcome alignment and curriculum coherence over multiple academic terms.

Leveraging regularization-based CL, the model would process curriculum effectiveness data across multiple semesters, including course performance metrics, student retention, pass/fail patterns and longitudinal outcome mastery. These data are used to identify:

1. Misalignments between instructional content and learning outcomes

2. High-impact bottlenecks in course sequencing

3. Shifts in performance linked to delivery modality or assessment design

The system proposes dashboard-style reports to visualize performance trends across courses, flagging areas that consistently exhibit lower levels of observed student achievement. For example, if students consistently underperform in fluid mechanics after a redesign of the statics course, the model will surface this correlation, prompting faculty to examine potential content drift or prerequisite gaps.

Unlike conventional curriculum analytics tools, the CL engine retains prior patterns while incorporating new cohort data. This allows it to learn over time which recommendations have historically led to positive curricular outcomes, thus enhancing the relevance of its future insights.

The curriculum co-design tool is not intended to replace instructional decision-making but to facilitate informed faculty collaboration, reduce manual analysis burden and guide iterative improvement grounded in evidence and context.

An essential output of the AI-assisted curriculum co-design tool is the longitudinal tracking and visualization of historical performance trends. This component functions as the tool's institutional memory, preserving critical insights from past academic terms while enabling adaptive refinement of curriculum strategies. It is designed to help faculty and program leaders interpret outcome trajectories over time and inform evidence-based adjustments to instructional practices.

Historical performance trends refer to the system's ability to monitor and visualize key longitudinal curriculum indicators. These include:

1. Pass Rates: The percentage of students who successfully complete courses with passing grades. A decline from 85% in Fall (2023) to 77% in Spring (2025), for example, may indicate increasing difficulty or misalignment in course content.

2. Retention Rates: The proportion of students who continue in the program from one semester to the next. A decrease from 92% to 85% suggests potential issues in student engagement or curriculum pacing.

3. Progression Rates: The percentage of students advancing to subsequent courses in the sequence, such as from Statics to Dynamics. A drop from 78% to 70% may reflect underpreparedness or ineffective instructional scaffolding.

These metrics are presented in an integrated, multi-line chart interface that enables easy identification of declining, improving, or stable trends. Faculty can observe these three dimensions simultaneously to recognize broader patterns and isolate performance bottlenecks. To illustrate the types of longitudinal analytics envisioned for the curriculum co-design tool, Figure 7 presents an example dashboard showing representative curriculum indicators across multiple academic terms.

The trends would be generated using course metadata, assessment outcomes, student enrollment records and cohort progression data elements already modeled in the system's longitudinal dataset. These data streams are aligned with outcome-level scores tracked by the tool and provide the foundation for structured trend analysis.

The historical performance trends also inform the EWC mechanism embedded within the system. Curriculum elements associated with stable or improving indicators would receive greater preservation via EWC regularization, while declining trends may serve as signals for potential adaptation in the co-design model. In this context, the EWC penalty discourages unnecessary modifications to historically strong elements while allowing flexibility in areas showing consistent decline. Figure 8 shows an illustrative EWC-based curriculum adaptation example.

This component serves several high-impact functions:

1. Early Warning System: Identifies systemic issues before they affect program quality or accreditation targets.

2. Intervention Timing: Allows faculty to correlate instructional or assessment changes with observed shifts in student outcomes.

3. Evidence-Based Decisions: is intended to support data-informed curriculum review, avoiding overreaction to short-term anomalies.

4. Recognition of Success Patterns: Enables replication of effective teaching strategies by highlighting what has worked consistently.

Ultimately, the historical performance trends component directly supports the co-design tool's broader goals by informing the EWC model parameters, generating actionable recommendations and validating the outcomes of previous interventions. It also empowers faculty through transparent visual evidence, making collaborative curriculum review sessions more focused, efficient and impactful.

The core architecture of this system is presented in Appendix B. The system represents an innovative approach to curriculum analytics that combines the pattern recognition capabilities of AI with the pedagogical expertise and contextual knowledge of experienced faculty, creating a sustainable framework for continuous curriculum improvement across engineering programs.

The two case studies presented illustrate how CL-enabled AI systems can serve as collaborative partners to educators, addressing both instructional support and curriculum refinement in higher education. While these models are proposed at the conceptual design level, their structure reflects a growing need for AI systems that can evolve over time, adapt to complex instructional contexts and function as decision support tools rather than autonomous replacements.

While both case studies are grounded in CL and human–AI collaboration, they operate at different levels of the educational system and serve complementary functions. The ITS focuses on micro-level instructional support, providing real-time, student-centered insights that help instructors address immediate learning needs and misconceptions. In contrast, the curriculum co-design tool operates at a macro level, analyzing longitudinal performance trends across courses to inform program-level decisions and curriculum alignment. Together, these models illustrate a multi-scale approach to educational AI, where localized, adaptive instruction and broader, data-driven curriculum refinement reinforce one another. This complementarity highlights the potential of CL systems not only to enhance individual learning experiences but also to support systemic, sustainable improvement in engineering education. Future implementations may further integrate these two layers, enabling feedback from instructional-level systems to inform curriculum-level adaptation, creating a fully connected CL ecosystem.

One of the most significant advantages of these models lies in their ability to retain valuable patterns from past cohorts while simultaneously integrating new data. In the tutoring system, this means that effective instructional strategies, such as the use of visual simulations or targeted scaffolding, are not forgotten when new groups of students enter the course. In the curriculum co-design tool, the system is envisioned to track systemic inefficiencies and successful revisions across semesters, generating increasingly informed recommendations over time.

Crucially, both models emphasize human oversight and contextual interpretation. Rather than making instructional decisions autonomously, the systems surface patterns and generate recommendations that faculty can evaluate, override or refine. This aligns with the broader goal of human-AI teaming, where machines handle pattern recognition and trend detection at scale, while educators provide meaning, ethics and pedagogy. Another benefit is scalability. As programs grow in size and complexity, it becomes increasingly difficult for faculty to track longitudinal trends or tailor instruction at an individual level. CL systems can help address this gap by providing a layer of memory and adaptability that enhances responsiveness without requiring complete retraining or manual recalibration.

CL systems introduce some challenges, particularly in the context of education. First is the issue of interpretability: faculty must be able to understand and trust the system's outputs in order to act on them. Without careful interface design and transparent algorithm behavior, the systems may be underutilized or misunderstood. Second, the risk of unintended bias in continually adapting models must be considered. If past cohort data reflect systemic inequities, the model may inadvertently reinforce them unless checks and balances are built into the update process. Faculty input, domain expertise and diverse training data are essential to mitigate this risk. Another limitation is the need for institutional data infrastructure capable of supporting longitudinal tracking across cohorts, courses and outcomes. Many educational settings still face barriers to integrating student data across systems in a privacy-compliant and accessible way.

Ethical considerations are central to the implementation of AI-assisted systems in educational settings, particularly when student data are involved. In the instructional contexts informing this work, AI-supported activities were integrated as part of regular course design, and no personally identifiable information was used in the development of the conceptual models presented. Data used to inform system design and insights were aggregated and de-identified, ensuring that individual student privacy was preserved. Instructors maintained full oversight of AI-supported interactions, and students engaged with these tools as part of structured learning activities with clear pedagogical intent. Future implementations of such systems should continue to adhere to institutional guidelines, including transparency in data use, informed participation and compliance with applicable data protection policies.

The chatbot implementations described in this article were conducted as instructional activities integrated into regular course delivery and are discussed for illustrative and design purposes only. The present manuscript does not report the collection, analysis or publication of identifiable student research data, nor does it present empirical findings derived from human-subjects research. Consequently, institutional review board approval was not required for the conceptual framework and design-focused discussion presented in this work. Future empirical evaluations of the proposed CL systems will be conducted under appropriate institutional review procedures, including participant consent and applicable ethical oversight requirements.

Looking ahead, these conceptual models open multiple avenues for future research and development. As a future direction, the insights generated by these AI-assisted systems, particularly the curriculum co-design tool, can be leveraged to support program-level assessment and accreditation processes. By continuously analyzing longitudinal data on student performance, outcome alignment and instructional effectiveness, the system can help identify trends that inform evidence-based improvements. This can enhance curriculum mapping, guide faculty development initiatives, and support documentation for accreditation bodies such as Accreditation Board for Engineering and Technology (ABET). Integrating CL into program assessment not only increases analytical precision but also fosters a culture of continuous improvement grounded in data-driven decision-making.

Another promising direction is the integration of virtual laboratories as simulation environments that can generate controlled, diverse and rich data streams for CL. Our prior work in this area demonstrates how such environments can support both formative assessment and model training, offering a valuable complement to real-world student data. Additionally, as these systems mature, future iterations could incorporate multimodal learning analytics (e.g. video, voice, eye-tracking) to deepen understanding of student engagement and learning processes. This could further expand the system's capacity for personalization and adaptation.

Finally, cross-program deployment and meta-level learning, where insights from one course inform recommendations in related courses, could enable AI systems to support broader institutional learning goals while maintaining relevance at the local level.

Future work should focus on the empirical validation of the proposed systems through controlled studies and longitudinal data collection. This includes the use of statistical analyses to evaluate learning gains, intervention effectiveness and curriculum improvements across cohorts. Metrics such as performance distributions, mastery thresholds and retention rates could be analyzed using appropriate inferential methods to assess the impact of CL-based AI systems in educational settings.

This article proposes a conceptual framework for integrating CL into AI-assisted educational systems, with a focus on fostering meaningful human–AI collaboration in higher education. By presenting two illustrative case studies, an ITS designed to evolve with each student cohort and a curriculum co-design tool that supports long-term instructional refinement, we demonstrate how CL can enable systems to become more responsive, context-aware and sustainable over time.

Unlike traditional static AI tools, the models described here are designed not to replace educators but to enhance their decision-making by providing insights that are accumulated, refined, and contextualized through continual exposure to learning data. This human-centered approach aligns with emerging best practices in AI ethics and educational technology, where transparency, adaptability and human oversight are foundational principles.

The proposed value of these models lies in their potential to track cohort-level patterns, identify persistent challenges and retain instructional strategies identified as useful within the CL framework. At the same time, they invite ongoing reflection about the importance of data governance, faculty engagement and model interpretability to ensure that these tools are both effective and equitable.

Importantly, the development of these case studies was informed by prior experimentation with domain-specific chatbot implementations across multiple courses. These chatbots, which were designed around structured roles and guided conversational flows, provided valuable insights into how AI can scaffold disciplinary reasoning, facilitate reflective thinking and support just-in-time feedback. Their implementation provided initial insights into the pedagogical potential of human-AI teaming, but also established a practical foundation for scaling up to more complex, adaptive AI systems that align with CL principles.

Looking forward, future development can build on this foundation by integrating virtual learning environments that generate synthetic data for early-stage model training and validation. Such simulations can accelerate the development of AI systems in education while protecting student data and providing high-resolution, task-specific feedback.

Ultimately, by embedding CL within a human–AI partnership model, this framework offers a path toward adaptive, scalable and pedagogically aligned educational systems that grow and improve with the people they are designed to support.

The supplementary material for this article can be found online.