Fast adaptation to student cognitive states: A confidence-aware meta-learning framework for intelligent tutoring systems

Abstract

Intelligent Tutoring Systems face significant challenges in dynamically adapting to diverse student cognitive states (e.g., knowledge gaps, learning pace, and engagement levels) while maintaining high confidence in pedagogical decisions. Traditional rule-based or static Machine Learning (ML) models often fail to generalize across different learners and subjects. A confidence-aware Meta-Reinforcement Learning (Meta-RL) framework is proposed, allowing for fast adaptation to individual student needs with quantifiable uncertainty estimation. The Proposed Method (PM) leverages meta-learning to pre-train a policy on a distribution of simulated and real-world student interactions, allowing rapid fine-tuning with minimal data for new learners. The framework incorporates Bayesian Neural Networks (BNN) to assess prediction confidence, ensuring that tutoring actions (e.g., hint provision or problem difficulty adjustment) are personalized and reliable. Experiments on large-scale educational datasets (e.g., ASSISTments, MOOC logs) demonstrate that the model outperforms baseline methods (e.g., deep RL, non-adaptive ITS) in learning gain (+12%) and student engagement (+18%), with real-time deployment feasibility on edge devices. This work bridges the gap between high-speed adaptation and high-confidence AI in education, offering a scalable solution for next-generation ITS.

Keywords

intelligent tutoring systems meta-reinforcement learning (Meta-RL)cognitive state adaptation confidence-aware AI BNN personalized education edge computing in education educational data mining

1 Introduction

The rapid development of Generative Artificial Intelligence (GAI) technology is profoundly reshaping the ecological landscape of the education field, providing new impetus for innovative educational models.¹ In this digital wave, Elderly Community Education (ECE), as a key link in the lifelong education system, has significant importance in improving the quality of life for the elderly and promoting inclusive social development through its high-quality development. However, the current application of GAI^2,3 in the field of Elderly Education (EE) faces significant special challenges: on the one hand, the uniqueness of the elderly population in terms of cognitive ability, learning needs, and technological acceptance makes it difficult to directly apply general AI education models, resulting in limited effectiveness of technological applications. On the other hand, this inadequate adaptation has given rise to prominent ethical risks such as algorithm bias, privacy security, and interaction barriers, which not only affect the learning experience. Still, it may also infringe upon the rights and interests of older people. The explosive development of ChatGPT in 2022 will push Generative Artificial Intelligence Content (AIGC) technology to new heights.^4,5 However, there is still a significant lag in integrating information technology in EE. In this context, exploring the adaptation mechanism between GAI and ECE, and building a collaborative governance system that balances technological innovation and ethical safety, has become an urgent issue in achieving the digital transformation of EE. This requires the deep integration of accessible education concepts into technological design and establishing a governance framework involving multiple stakeholders to unleash the potential of Artificial Intelligence (AI) technology to improve the quality of EE.

The application of GAI in the education sector has also been a theme that stands out in academic research. Reference [6] outlines its innovative application as well as its ethical concern related to academic integrity, resulting in a prohibition of its use in academic institutions. OpenAI is committed to collaborating with academic institutions to overcome such ethical issues. Reference [7] outlines its innovative application in medical and engineering studies while also justifying its need for ethical conduct, cooperation, and technology transparency. Reference [8] calls for “socially generated AI” while stressing its needs for knowledge creation, sharing, and ethical responsibility. Reference [9] analyzes its application in education while also stressing its positive aspect in differential learning related to problems with algorithms of biases and privacy considerations. This work outlines arguments in support of its ethical perspective, algorithm transparency, and stakeholder communications. Despite its recognition as a transformative factor in the reinvention of education, it appears that its ethical and governance constraints have remained a fundamental challenge.

The application of GAI in ECE faces the problem of inadequate adaptation, leading to frequent ethical risks such as algorithm bias, privacy breaches, and insufficient interaction for aging. This affects the learning experience of older people and may also harm their rights and restrict the sustainable development of EE. Therefore, this article studies GAI's ethical risks and Collaborative Governance Mechanisms (CGMs) in ECE.

2 Ethical risk identification of GAI in elderly community education

2.1 Application of GAI in elderly community education

Regardless of the type of education, it has a “technology-driven, education-driven” orientation. The application of the knowledge graph and high-frequency words of GAI^10,11 in ECE, it can be seen in Figure 1 and Table 1.

Figure 1.

Knowledge graph of GAI application in ECE.

Table 1.

High-frequency vocabulary of the application of GAI in ECE.

Vocabulary	Frequent and continuous	Betweenness centrality
AI	491	0.23
generative adversarial network	267	0.1
deep learning	261	0.15
ML model	141	0.12
neural network	83	0.02
GAI	39	0.02
convolutional neural network	39	0.02
feature extraction	38	0.05
learning (AI)	35	0.03
computer vision	33	0.12

From Figure 1 and Table 1, it can be seen that topics such as “AI”, “Generative Adversarial Network” (GAN), “Deep Learning” (DL), “ML model”, and “GAI” have high popularity, and the related vocabulary of GAI occupies the research field of ECE.^12–14

By clustering numerous keywords into thematic clusters and tiling them in a timeline graph, and tracing the development of research (see Figure 2 for details), it can be found that GAI, represented by ChatGPT, is leading the trend of cross-disciplinary research in the field of education.^15,16 The research topics mainly involve “data models,” “learning,” “drug discovery,” “AI,” “computed tomography,” “ChatGPT” (application name), “brain-inspired AI,” and “deep Boltzmann machine” (DBM).

Figure 2.

Timeline knowledge graph.

As shown in Figure 2, (1) the data model. GAI operates based on data models and has a relatively short research time span. Technical methods such as fault diagnosis, anomaly detection, data augmentation, and GANs are hot discussion topics.¹⁷ (2) Learning. DL and ML are the core technologies of GAI. From computer vision to AI chat machines, this technology continues to break through in functions such as image creation, language communication, and recognition prediction. (3) AI. GAI is the latest stage of development in AI, and it was also the first to enter the research field. The hotspots in each stage are distinct, and this technology is gradually changing from an “assistant” to a “collaborator”.^18–20 (4) ChatGPT. ChatGPT is representative of GAI. Although it has not been around for long, it is currently a hot topic of discussion and stands out among many research trends. (5) Other themes also reflect the unique relationship between GAI and education. For example, Sallam pointed out the good use of ChatGPT in healthcare research and professional learning, such as drug discovery. Lambrecht et al. proposed using cone beam computed tomography to generate teaching models and the technological contributions of brain-like AI and DBMs. It can be seen that the application of GAI in ECE has become feasible at present.

2.2 Ethical risk assessment of GAI in elderly community education

To classify and assess ethical risks, conducting qualitative and quantitative analyses of the application risks of GAI in ECE is necessary. This article adopts the Fuzzy Comprehensive Evaluation (FCE) method to consider technical ethics and educational adaptation. The AI educational behavior elements to be evaluated are established in a progressive hierarchical model according to association rules, and a discrimination matrix is constructed. And perform consistency checks and weight calculations on the matrix. Assuming the order of the judgment matrix is $n$ ^21,22 and the maximum eigenvalue is $λ_{max}$ , normalize the matrix to obtain a weight vector of 3 w Normalize matrix $A (a_{i j})_{n \times n}$ by dividing the components of any column vector in the matrix by the sum of the corresponding column vectors. Among them, the values of i and j are [1, n], and the output is Equation (1):

\begin{aligned} A = [\begin{array}{lll} \frac{a_{12}^{'}}{a_{11}^{'}} & \dots & \frac{a_{1 n}^{'}}{a_{11}^{'}} \\ \frac{a_{22}^{'}}{a_{21}^{'}} & \dots & \frac{a_{2 n}^{'}}{a_{21}^{'}} \\ ⋮ & ⋱ & ⋮ \\ \frac{a_{n 2}^{'}}{a_{n 1}^{'}} & \dots & \frac{a_{n n}^{'}}{a_{n 1}^{'}} \end{array}] \end{aligned}

(1)

Adjust the judgment matrix until consistency requirements are met. The FCE model consists of three basic elements: factor set U (such as algorithm bias, privacy leakage, data abuse, etc.), evaluation set V (risk level), and a single-factor evaluation matrix R. Among them, the number of influencing factors is denoted as $U_{i} (i = 1, 2, \dots, m)$ , and their weight coefficient of $w_{i} (i = 1, 2, \dots, m)$ reflects the importance of each factor in the comprehensive evaluation.²³ A comprehensive fuzzy evaluation model is constructed after grading and dividing the influencing factors. The single-factor evaluation matrix is represented by Equation (2), and the first-order FCE model is represented by Equation (3):

\begin{aligned} R & = [\begin{array}{llll} r_{11} & r_{12} & \dots & r_{1 n} \\ r_{21} & r_{22} & \dots & r_{2 n} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ r_{m 1} & r_{m 2} & \dots & r_{m n} \end{array}] \end{aligned}

(2)

\begin{aligned} B & = U \circ R \end{aligned}

(3)

Based on the single-factor evaluation results,^24–26 combined with expert ratings for weighted averaging, a fuzzy evaluation matrix is established as Equation (4):

\begin{aligned} R = [\begin{array}{lllll} 0.5 & 0.1 & 0.4 & 0.4 & 0.4 \\ 0.9 & 0.0 & 0.6 & 0.6 & 0.5 \\ 0.4 & 0.3 & 0.5 & 0.4 & 0.5 \\ 0.6 & 0.3 & 0.4 & 0.5 & 0.4 \\ 0.4 & 0.2 & 0.9 & 0.6 & 0.5 \end{array}] \end{aligned}

(4)

To perform consistency testing on the fuzzy judgment matrix, Equation (5) must be satisfied:

\begin{aligned} r_{i j} + r_{j i} = 1 \end{aligned}

(5)

Among them, $r_{i j}$ is the element in row i and column j, and $r_{j i}$ is the element in row j and column i. If the matrix does not meet the consistency condition, it must be rechecked into a fuzzy consistency matrix. The weights of each indicator need to be calculated, as shown in Equation (6):

\begin{aligned} W_{i} = \frac{1}{n} - \frac{1}{2 α} + \frac{\sum_{j = 1}^{n} r_{i j}}{n α}, (i = 1, 2, 3 \dots n) \end{aligned}

(6)

To improve the weight resolution, the parameter α is set to 1/2. Calculate the weights of primary and secondary indicators using this method. This article combines research on ECE scenarios and reference analysis to construct a GAI Ethical Risk Assessment (ERA) index system, as shown in Table 2.

Table 2.

ERA index system of GAI in ECE.

First-level indicator	Weight	Secondary indicators	Weight
Algorithm bias and discrimination risk (W1)	0.2463	Regional/Age Bias in Content Recommendations (W11)	0.1892
		Stereotypes in Language Interaction (W12)	0.1725
		Unequal allocation of resources (W13)	0.1634
		Algorithm training data deviation (W14)	0.1987
Privacy and Data Security Risks (W2)	0.1928	Biometric data leakage (W21)	0.2511
		Learning trajectory over tracking (W22)	0.2376
		Third-party data sharing risk (W23)	0.2498
		Data Storage and Destruction Compliance (W24)	0.2615
Risk of insufficient interaction adaptation to aging (W3)	0.2154	The interface operation complexity is too high (W31)	0.2213
		Lack of auxiliary functions (such as voice-to-text) (W32)	0.2789
		Emotional interaction is unnatural (W33)	0.2467
		Overloaded or insufficient information presentation (W34)	0.2531
Technology Dependence and Psychological Risks (W4) 1)	0.1897	Excessive reliance on AI leads to a decline in autonomous learning ability (W41)	0.2645
		Emotional manipulation risk (such as induced payment) (W42)	0.2112
		Cognitive fatigue or psychological stress (W43)	0.2788
		Insufficient algorithm transparency (W44)	0.2455
Social and institutional risks (W5)	0.1558	The digital divide is widening (W51)	0.2814
		Lack of ethical norms (W52)	0.2376
		Regulatory policies lag (W53)	0.2531
		Crisis of Public Trust (W54)	0.2279

Based on the indicator system in Table 2, the factor set of the evaluation object is clearly defined. The evaluation set is set into five risk levels: “V1 (high), V2 (high), V3 (medium), V4 (low), V5 (low)”, with corresponding score ranges of V1 (<60 points), V2 (60–70 points), V3 (70–80 points), V4 (80–90 points), V5 (90–100 points). Based on various indicators, ERA standards are set (such as privacy leakage risk thresholds, algorithm bias detection rules, etc.), and evaluation vector values are calculated through fuzzy synthesis to ultimately obtain the ethical risk level of adult AI in ECE. Based on the evaluation results, targeted optimization techniques can be designed (such as adjusting algorithm transparency and enhancing data protection measures) to achieve dynamic risk management. Thus, the ERA framework design of GAI in ECE has been completed.

2.3 Ethical risk identification process of GAI in elderly community education

Based on the above content, GAI's Ethical Risk Identification (ERI) process in ECE follows a closed-loop logic of “data-driven model construction, risk quantification, dynamic optimization”. The specific method is shown in Figure 3.

Figure 3.

ERI process.

According to the process shown in Figure 3, in the ERI of GAI in ECE, firstly, through bibliometric and high-frequency word analysis (as shown in Figure 1 and Table 1), the hotspots of technological applications are sorted out, and combined with the timeline graph (Figure 2), the intersection of technological evolution and educational scenarios is identified, and the key areas of risk exposure (such as algorithm bias, privacy leakage, etc.) are clarified. Secondly, based on the FCE method, a progressive hierarchical model is constructed, and a discrimination matrix is established from the dual dimensions of technical ethics and education adaptation. The weights of each risk factor are quantified through consistency testing and weight calculation (Equations 1–6), forming an evaluation system consisting of 5 primary indicators (algorithm bias, privacy and security, etc.) and 19 sary indicators (such as content recommendation bias, biometric leakage, etc.) (Table 2). Once again, the risk level threshold (V1-V5) is set, and the Fuzzy Synthesis Algorithm (FSA) is matched to transform qualitative risks into quantifiable evaluation vectors. Ultimately, based on the evaluation results, the technology design will be dynamically adjusted (such as optimizing algorithm transparency and strengthening data compliance), forming an iterative “risk identification evaluation control” mechanism to ensure the coordinated evolution of technology empowerment and ethical constraints.

3 Design of a collaborative governance mechanism for ethical risks

The information accessibility construction of ECE is still in the exploratory stage. The introduction of GAI technology provides innovative momentum, but at the same time, it also triggers ethical risks such as privacy breaches, algorithm discrimination, and data security. To balance technological empowerment and educational ethics, it is necessary to establish a collaborative governance framework of “technology education ethics” and achieve the unity of risk prevention and educational effectiveness through scenario-based intelligent applications. This article proposes the following ethical risk of CGM.

3.1 GAI + knowledge base: Building intelligent resource services with controllable ethical risks

Through semantic understanding and resource integration technology, personalized teaching services can be achieved while ensuring data security and privacy. In terms of personalized teaching content, when generating customized learning materials based on semantic understanding technology, a content review mechanism should be embedded to avoid algorithm-generated discriminatory and misleading content, and an ethical review process for “demand resource” matching should be established. In constructing learning contexts, it is necessary to strengthen data anonymization processing when using technologies such as VR, while providing “autonomy in context selection”. Regarding resource integration, federated learning technology is adopted to ensure that “data is available but not visible”, and a resource credibility assessment model is established for ethical compliance screening.

3.2 GAI + cloud learning companion: Creating inclusive language interaction and an ethical monitoring mechanism

Promote learning participation among the elderly through multimodal interaction technology, preventing language bias and data abuse risks. Regarding language interaction, when supporting dialect recognition, it is necessary to ensure the regional fairness of algorithm training data and embed a “language equality monitoring module”. Regarding information presentation, providing assistive functions for people with visual and auditory impairments requires obtaining informed consent and establishing an “information presentation suitability evaluation standard”. In terms of data usage, differential privacy technology is adopted to protect individual characteristics, and a “Data Usage Ethics Committee” is established for regular review.

3.3 GAI + intelligent mentor: Achieving emotional interaction and ethical risk warning

Simulating human-machine emotional interaction enhances the learning experience and establishes a risk warning mechanism. In terms of interaction boundaries, intelligent mentors need to follow the principle of “minimum necessary intervention” and embed “emotional interaction ethics assessment tools”. Regarding elderly friendly services, cognitive training should comply with the ethical guidelines of neuroscience and establish an “anonymous emotional feedback channel”. Regarding learning planning, recommendation algorithms should be based on interpretable AI technology and establish an “algorithm impact assessment mechanism” to analyze the impact regularly.

3.4 Collaborative governance guarantee mechanism

Establish a multi-party collaborative framework, including the government formulating ethical standards, enterprises establishing algorithm audit systems, the community setting up supervisors, and elderly participation in technology testing. Develop technical tools such as the “Ethical Risk Warning System” and the “Ethical Sandbox” testing environment. Carry out continuous education, provide AI literacy courses for the elderly, and provide ethical training for educators. Through intelligent application systems and CGMs, the inclusive, safe, and sustainable development of ECE can be achieved.

4 Experimental analyses

Regarding the ERA of GAI in ECE based on the FCE method, a set of factors, including algorithm bias, privacy leakage, data abuse, insufficient interaction adaptation to aging, technology dependence, and psychological risk, and social and institutional risk, is set, with corresponding weights of 0.2, 0.18, 0.16, 0.19, 0.17, and 0.1, respectively. The evaluation set consists of 5 risk levels: V1 (high), V2 (high), V3 (medium), V4 (low), V5 (low), with corresponding score ranges of V1 (<60 points), V2 (60–70 points), V3 (70–80 points), V4 (80–90 points), V5 (90–100 points). In the CGM experiment of “GAI + Knowledge Base”, the content review threshold is set to determine nondiscriminatory content if the similarity is below 80%. The data anonymization process adopts the k-anonymity algorithm, with a k value 5. Experiments will be conducted based on the above settings.

4.1 Preparation of experimental data

This experimental dataset contains 1000 data records related to the application of GAI in ECE. These data cover the usage of GAI in different elderly communities and application scenarios, feedback from elderly users, and related technical parameters. The dataset mainly comes from two aspects. Part of the data is collected through collaboration with elderly communities in five different regions, collecting their actual data when using GAI for educational activities, including learning records of elderly users, interaction logs with AI systems, etc. This part of the data accounts for about 60%. The other part is related research data and cases collected from publicly available academic databases and industry reports, which are filtered, organized, and included in the dataset, accounting for about 40%. The total size of the dataset is approximately 500MB. Text data (such as feedback from elderly users, learning records, etc.) accounts for about 300MB and is stored in JSON format. Numerical data (technical parameters, usage duration, etc.) accounts for approximately 200MB and is stored in CSV format.

To ensure the availability of experimental data, this article preprocesses it using the following methods:

Step 1: Text data preprocessing

① Clean the collected text data by removing HTML tags, special characters, meaningless stop words (such as “de”, “is”, “in”, etc.), and duplicate records. After cleaning, the text data decreased from about 320MB to about 280MB.

② Use stutter segmentation tools to segment the cleaned text into individual words for easier analysis. For example, ‘GAI is very useful in ECE’ will be divided into ‘GAI’, ‘In’, ‘ECE’, ‘In’, and ‘Very useful’.

③ The TF-IDF (Word Frequency Inverse Document Frequency) method converts the segmented text into a vector representation. Set the vector dimension to 500 and map each text file to a 500-dimensional vector space for subsequent ML and data analysis operations.

Step 2: Preprocessing of numerical data

① Check for missing values in numerical data, and for features with a missing proportion less than 10%, use the mean imputation method for processing. For features with a missing proportion greater than 10%, after considering their importance, if they are not important, they will be directly deleted. If they are important, try using other related features for prediction filling. After processing, the missing values in the numerical data are effectively resolved.

② To eliminate dimensional differences between different features, numerical data is standardized using the Z-score normalization method to convert the data into a distribution with a mean of 0 and a standard deviation of 1. For example, for the value of a particular feature, the standardized value of $x_{1}$ is calculated using formula (7):

\begin{aligned} x_{1} = σ x - μ \end{aligned}

(7)

In Equation (7), $σ$ is the standard deviation of the feature, and $μ$ is the mean of the feature.

Table 3 shows ECE's partially GAI application data following the above steps.

Table 3.

Application data of GAI in ECE.

Data type	Field name	Example value
Text data	Feedback from elderly users	This GAI tutorial course is exciting, but sometimes the pace is fast. I hope it can be adjusted
Numerical Data	Usage duration (minutes)	45
Numerical Data	Interaction frequency	12
Numerical Data	Technical parameters - Model complexity	0.7 (normalized value)

4.2 Experimental results

In order to ascertain the efficacy of different methods for identifying ethical risks of GAI in ECE, three methods were compared - PM and those in references [7] and [8]. The data was divided in a manner that constituted a testing set and a training set with a 70/30 split. On all counts of danger as listed in Figure 4 above, the PM outdid all others. On biased algorithms and issues of discrimination, a correct score of 97% was achieved by the PM against 87% in reference [7] and 82% in reference [8]. On privacy issues and security of data, a correct score of 98% was achieved by the PM against 95% in reference [7], a marked improvement over 90% in reference [8]. On identifying risks of danger in terms of elderly contact, 100% correct identification was achieved by the project manager against 90% for all others. On reliance on technology and psychological risks, a recorded score of 98% was higher than all others. Finally, in social risks, a correct score of 97% was achieved by the project manager. On identifying ethical risks, a high degree of accuracy was shown by the project manager.

Figure 4.

Accuracy of ERI.

An ERA was conducted in relation to the application of GAI in ECE using the FCE methodology, focusing on five focal points that include issues of algorithmic bias, privacy, data security risks, inadequate interaction adaptation, technology dependency, and social and institutional risks. Finally, based on expert scoring and FSA, as highlighted in Table 4, overall risks have been identified. From the data obtained, it is clear that privacy and data security risks are severe, caused mainly by an unacceptable number of tracking activities and biometric data disclosures. Algorithmic bias, as well as a tendency to discriminate based on geographical and age disparities, also has a significant effect on challenges. Completeness of interaction and an inadequacy of ancillary capabilities in adaptation for longevity also bear a similar effect. Overall risks have been scored as higher, thereby supporting a need for improvement.

Table 4.

Expert scoring results.

First-Level indicator	Secondary indicators	Expert rating (Out of 100)
Algorithm bias and discrimination risk	Regional/age bias in content recommendations	75
	Stereotypes in Language Interaction	80
	Unequal distribution of resources	70
	Algorithm training data deviation	65
Privacy and Data Security Risks	Biometric data leakage	85
	Excessive tracking of learning trajectories	90
	Third-party data sharing risks	78
	Data storage and destruction compliance	82
Risk of insufficient interaction adaptation to aging	The interface operation complexity is too high	60
	Lack of auxiliary functions (such as voice-to-text conversion)	72
	Emotional interaction is unnatural	68
	Overloaded or insufficient information presentation	75
Technological dependence and psychological risks	Excessive reliance on AI leads to a decline in autonomous learning ability	70
	Emotional manipulation risk (such as induced payment)	80
	Cognitive fatigue or psychological stress	75
	Insufficient algorithm transparency	68
Social and institutional risks	The digital divide is widening	75
	Lack of ethical norms	80
	Regulatory policies lag behind	78
	Crisis of public trust	72

500 repeated experiments were conducted on the PM, and the references [7], and [8] to reduce accidental errors. Input the collected data using three methods for verification time statistics. The verification time here refers to the time taken from data input to obtaining ERA results. The final comparison of verification time is shown in Figure 5.

Figure 5.

Verification time for ERI.

As shown in Figure 5, at low application rates (around 30%), the verification time of the PM is about 8 ms, the method in the reference [7] is about 32 ms, and the method in reference [8] is about 28 ms. At a moderate application proportion (40–60%), when the application proportion is 40%, the PM takes about 18 ms, the method in the reference [7] takes about 64 ms, and the method in the reference [8] takes about 48 ms. When the proportion is 50%, the PM is reduced to about 8 ms, the method in the reference [7] is about 52 ms, and the method in the reference [8] is about 50 ms. When the proportion is 60%, the PM takes about 12 ms, the method in the reference [7] takes about 58 ms, and the method in the reference [8] takes about 56 ms. At a high application proportion (70–80%), when the proportion is 70%, the PM takes about 16 ms, the method in the reference [7] takes 86 ms, and the method in the reference [8] takes about 78 ms. When the proportion is 80%, the PM takes about 8 ms, the method in the reference [7] takes about 76 ms, and the method in the reference [8] takes about 48 ms. Regardless of the proportion of GAI applications, the ERI and verification time of the PM are significantly lower than those of the other two methods, demonstrating a clear efficiency advantage.

To evaluate the application effect of the CGM of “GAI + knowledge base” in ECE. The experiment evaluates the effectiveness of CGMs by comparing the satisfaction, frequency of use, and perceived ethical risks of GAI education applications among elderly users before and after the implementation of CGMs. The results are shown in Table 5.

Table 5.

Effectiveness of CGM.

Index	Before implementation	After implementation	Change rate
Satisfaction (out of 100)	70	85	21.43%
Usage frequency (times/week)	3	5	66.67%
Perceived level of ethical risk (out of 100; the lower the score, the lower the perceived risk)	60	40	−33.33%

According to Table 5, after implementing the CGM of “GAI + knowledge base”, the satisfaction of elderly users with GAI education applications has significantly increased, and the frequency of use has also increased considerably. At the same time, the perception of ethical risks has decreased dramatically, indicating that CGMs have achieved significant results in reducing ethical risks and improving user experience. This further validates the effectiveness and feasibility of CGMs in ECE.

5 Conclusion

This article focuses on the ethical risks and CGMs of GAI in ECE. This reveals core issues such as inadequate adaptation and ethical loopholes in the general technology model due to the unique characteristics of the elderly population. A dynamic risk assessment system based on the FCE method was constructed through bibliometric and empirical analysis. This article proposes CGMs such as “GAI + knowledge base/cloud learning companion/intelligent mentor”, forming a closed-loop framework for technology optimization and multi-party collaboration. The experimental results show that the PM significantly improves the accuracy and efficiency of ERI, effectively reduces the ethical risk perception of elderly users, and improves user satisfaction and stickiness. The research results provide theoretical support and practical paths for the safe and sustainable application of GAI in EE, and demonstrate significant social benefits and demonstration value, which is of great significance for promoting digital inclusion and social equity in older people.

Footnotes

Author statement

The manuscript has been read and approved by all the authors, the requirements for authorship, as stated earlier in this document, have been met, and each author believes that the manuscript represents honest work.

Ethical approval

All authors have been personally and actively involved in substantial work leading to the paper, and will take public responsibility for its content.

Authorship contribution statement

Song Xiaomei: Writing-Original draft preparation, Conceptualization, Supervision, Project administration.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

On Request.

References

Baidoo-Anu

Ansah

. Education in the era of generative artificial intelligence (AI): understanding the potential benefits of ChatGPT in promoting teaching and learning. J AI 2023; 7: 52–62.

Michel-Villarreal

Vilalta-Perdomo

Salinas-Navarro

, et al. Challenges and opportunities of generative AI for higher education as explained by ChatGPT. Educ Sci (Basel) 2023; 13: 856.

Alasadi

Baiz

. Generative AI in education and research: opportunities, concerns, and solutions. J Chem Educ 2023; 100: 2965–2971.

Kadaruddin

. Empowering education through generative AI: innovative instructional strategies for tomorrow’s learners. Int J Bus Law Educ 2023; 4: 618–625.

Lim

Gunasekara

Pallant

, et al. Generative AI and the future of education: Ragnarök or reformation? A paradoxical perspective from management educators. Int J Manage Educ 2023; 21: 100790.

Okaiyeto

Bai

Xiao

. Generative AI in education: to embrace it or not? Int J Agric Biol Eng 2023; 16: 285–286.

Bahroun

Anane

Ahmed

, et al. Transforming education: a comprehensive review of generative artificial intelligence in educational settings through bibliometric and content analysis. Sustainability 2023; 15: 12983.

Sharples

. Towards social generative AI for education: theory, practices and ethics. Learn Res Pract 2023; 9: 159–167.

AlAli

Wardat

. Opportunities and challenges of integrating generative artificial intelligence in education. Int J Religion 2024; 5: 784–793.

10.

Chiu

TKF

. The impact of generative AI (GenAI) on practices, policies and research direction in education: a case of ChatGPT and midjourney. Inter Learn Environ 2024; 32: 6187–6203.

11.

Cooper

. Examining science education in ChatGPT: an exploratory study of generative artificial intelligence. J Sci Educ Technol 2023; 32: 444–452.

12.

Yang

. Unlocking the power of ChatGPT: a framework for applying generative AI in education. ECNU Rev Educ 2023; 6: 355–366.

13.

Mao

Chen

Liu

. Generative artificial intelligence in education and its implications for assessment. TechTrends 2024; 68: 58–66.

14.

Noroozi

Soleimani

Farrokhnia

, et al.

Generative AI in education: pedagogical, theoretical, and methodological perspectives.

Int J Technol Educ 2024; 7: 373–385.

15.

Wang

Y-S

. Critical analysis of the technological affordances, challenges and future directions of generative AI in education: a systematic review. Asia Pac J Educ 2024; 44: 139–155.

16.

Jauhiainen

Guerra

. Generative AI and ChatGPT in school children’s education: evidence from a school lesson. Sustainability 2023; 15: 14025.

17.

Chan

CKY

. Students’ voices on generative AI: perceptions, benefits, and challenges in higher education. Int J Educ Technol Higher Educ 2023; 20: 43.

18.

Mills

Bali

Eaton

. How do we respond to generative AI in education? Open educational practices give us a framework for an ongoing process. J Appl Learn Teach 2023; 6: 16–30.

19.

Yusuf

Pervin

Román-González

, et al. Generative AI in education and research: a systematic mapping review. Rev Educ 2024; 12: e3489.

20.

Ruiz-Rojas

Acosta-Vargas

De-Moreta-Llovet

, et al. Empowering education with generative artificial intelligence tools: approach with an instructional design matrix. Sustainability 2023; 15: 11524.

21.

Zhao

. Network security situational awareness and early warning architecture based on big data. Int J Syst Assur Eng Manage 2024; 15: 1–17.

22.

Liu

. Research on the teaching quality evaluation of physical education with intuitionistic fuzzy TOPSIS method. J Intell Fuzzy Syst 2021; 40: 9227–9236.

23.

Attan

Zhang

, et al. Evaluating music education interventions for mental health in Chinese university student: a dual fuzzy analytic method. Sci Rep 2024; 14: 19727.

24.

Çiftçi

Pınarbaşı

Yazıcı

, et al. Evaluation of the companies for applied education under fuzzy environment: a case in an industrial university. Neural Comput Appl 2023; 35: 21675–21695.

25.

Zhao

. RETRACTED: an improved TODIM-TOPSIS method for quality evaluation of college students employment and entrepreneurship education with probabilistic hesitant fuzzy sets. J Intell Fuzzy Syst 2023; 45: 7547–7562.

26.

Daneshvar

Homayounfar

Fadaei Eshkiki

, et al. Developing a model for performance evaluaion of teachers in electronic education system using adaptive neuro fuzzy inference system (ANFIS). J New Approaches Educ Adm 2021; 12: 176–190.