To guide the systematic literature review and delimit the analysis of the study, 6 research questions were formulated focused on the ethical, pedagogical, and evaluative implications of generative artificial intelligence in education. These questions allow structuring the search, selection, and analysis of the collected information.

— RQ1: How should assessments in educational contexts be redesigned to mitigate the misuse of generative AI and promote authentic learning?

— RQ2: What institutional strategies can be implemented in educational settings to guarantee the safe and reliable use of generative AI tools?

— RQ3: How can educational institutions integrate digital sustainability principles (Green AI) into the adoption of generative AI tools?

— RQ4: How can algorithmic biases be mitigated in the implementation of generative AI to ensure equity and inclusion in educational processes?

— RQ5: How can generative AI literacy be promoted to strengthen critical thinking, autonomy, and digital ethics in students and teachers?

— RQ6: How can generative AI, used as an artificial evaluator, support pedagogical and academic decision-making in educational institutions?

Table 1. Definition of general concepts with PIPOH.

Source: Own elaboration.

To precisely structure the bibliographic search process, the main terms of the study were defined in Table 2, along with their synonyms and conceptual filters associated with each dimension of analysis. On this basis, the databases selected for the review were consolidated in Table 3, while Table 4 presents the search string used to identify the relevant documents. This procedure guided the literature exploration toward studies related to generative artificial intelligence in education and its ethical, pedagogical, and evaluative implications.

Table 2. Terms, synonyms, and filters for composing the search string.

Source: own elaboration.

Table 3. Main term and databases.

Source: own elaboration.

Table 4. Search string used to obtain the documents.

Source: own elaboration.

The studies identified during the search process were subjected to a selection process to ensure their relevance to the research objectives. For this purpose, a progressive review of the documents was carried out, considering elements such as the title, keywords, abstract, and general content, to identify those that addressed artificial intelligence in the educational context, especially from a generative AI perspective.

The selection process involved a gradual refinement of the results obtained in the initial search. In the first phase, duplicate documents or those not available in open access were eliminated. Subsequently, studies that were not directly related to the use of artificial intelligence in education or that did not provide relevant information to answer the research questions were excluded. Finally, documents that met the defined criteria and supported the analysis of the ethical, pedagogical, and evaluative implications of generative AI were selected.

Table 5. Dates of consultation and file download.

Source: own elaboration.

To guarantee the quality of the selected studies, an evaluation process was carried out based on the review of the results obtained in each of the databases consulted. This process included the identification of initial documents, the elimination of duplicate records, and the exclusion of studies that did not meet the defined criteria, such as thematic relevance, the relationship with generative artificial intelligence in education, availability in open access, and the recency of the publications.

Table 6. Quality assessment in search and selection processes.

Source: own elaboration.

Table 7. Studies included in the systematic review.

Source: own elaboration.

The redesign of assessments in the context of generative artificial intelligence use must focus on the learning process and the development of critical thinking, rather than evaluating only final products. This implies transforming assessment methods, integrating AI transparently into academic activities, and promoting continuous and formative approaches that validate learning more authentically[1], [2], [3], [4].

Priority should be given to assessing skills that AI does not yet replicate accurately, such as reflective judgment, creativity, and the application of knowledge to local contexts or real and current problems that fall outside the training base of models [2], [4], [5], [6]. In this regard, it is recommended to replace short-answer theoretical questions or generic essays with case-development tasks and complex problem-solving [2], [3], [7].

Similarly, it is proposed to reaffirm the value of supervised in-person assessments and the use of oral exams, recorded presentations, or project defenses to verify the student's actual internalization of the concepts [1], [2], [3], [4], [5], [8].

Likewise, instead of prohibiting technology, it should be integrated through didactic protocols where the student declares the use of IAG, attaches the prompts used, and performs a critical analysis of the generated results (identifying biases or "hallucinations"), providing personal contributions that go beyond the automatic result [3], [5], [8].

Finally, it is essential to move toward a model of continuous assessment with frequent lower-weight tasks (low-stakes) that allow monitoring of student progress and self-regulation, using IAG as a tutor or personalized feedback assistant rather than only as a final control tool [1], [3], [6].

The institutional response must focus on proactive adaptation that updates regulatory frameworks concerning AI, incorporating rules on authorship simulation and avoiding legal gaps that affect academic integrity. In addition, it is necessary to implement digital literacy and training programs for faculty and students, oriented toward both the technical and ethical use of these tools, in line with institutional guidelines [2], [1], [9], [8].

To ensure the reliability of results generated by AI, institutions must require transparency in its use, requesting that students declare the tools used, include the prompts, and justify their personal contribution. Likewise, it is necessary to strengthen critical thinking, redesign assessments with fraud-resistant questions, and establish internal verification systems to mitigate risks such as disinformation and model hallucinations [5], [2], [5], [10].

From an ethical and security perspective, institutional strategies must include the creation of clear guidelines that protect personal data privacy and ensure compliance with intellectual property and copyright laws, which implies establishing robust security policies and informed consent mechanisms for data handling within AI platforms [11], [12].

In addition, the formation of communities of practice and specialized subcommittees is recommended to continuously review assessment design and share experiences on the educational use of technology [1], [8]. The literature also highlights the importance of implementing pilot programs to evaluate the impact of AI on critical skills before its massive adoption, ensuring that technological integration is scientific and professional [10], [11].

Finally, the success of these strategies depends on an approach of distributed responsibility, in which institutional governance, teachers, and students share a commitment to academic honesty and transparency in the use of generative tools [1].

The integration of digital sustainability begins with the recognition that these large language models require intensive processing power both for training and for inference, which entails considerable energy consumption and a carbon footprint that institutions must not ignore when incorporating these applications into their teaching practices [1], [13], [14], [15].

Another key pathway for incorporating these principles consists of fostering digital literacy and critical thinking within the curriculum, training teachers and students to understand the limitations and ecological cost of the tools. Likewise, institutions must adopt international recommendations that promote ethical and transparent use of technology, protecting the environment. In this context, teaching work must guide responsible use of AI, prioritizing critical analysis and avoiding excessive dependence that ignores its environmental impact [5], [1], [17].

From a technical perspective, institutions can adopt Green AI models, characterized by a low carbon footprint, lower computational complexity, and greater efficiency, which enables their use on local devices and reduces dependence on large data centers [15], [18], [19]. Furthermore, it is necessary to manage the life cycle of educational hardware, addressing electronic waste through circular economy principles and strengthening technological resilience [13].

Finally, digital sustainability implies that AI should not be seen as absolute truth, but rather as a tool that requires human validation to avoid superficial learning without pedagogical value. Therefore, institutions must implement policies and governance frameworks that align the use of AI with the Sustainable Development Goals, promoting responsibility and academic integrity [1], [6], [16].

Technology is not neutral, since generative systems can inherit biases present in their training data, affecting marginalized populations. Their mitigation requires identifying and correcting these biases through representative data, rebalancing, and proper curation, considering that information taken from the internet does not always guarantee quality. Therefore, developers must treat bias as a critical factor before implementing these systems in educational contexts [1], [2], [6], [20], [21].

From a technical perspective, strategies such as fine-tuning with human feedback (RLHF) are proposed, which allow aligning the model's responses with ethical criteria. Equity-oriented approaches, such as FairGAN, and bias-removal techniques also stand out, together with the use of explainable AI (XAI) to ensure transparency, traceability, and detection of possible discrimination in the results [6], [14], [19], [22], [23], [24], [25]

At the institutional level, bias mitigation requires an educational approach that prioritizes AI literacy and the development of critical thinking. Rather than prohibiting these tools, teachers and students should be trained to evaluate their outputs in an ethical and contextual manner, teacher training in bias detection should be strengthened, and good practices should be promoted that turn students into critical verifiers of generated content [4], [8], [11], [14], [25], [26].

Governance and regulatory frameworks are key to preventing generative AI from reproducing racial, gender, or socioeconomic biases. Educational policies should include periodic audits; compliance with data protection regulations and fundamental rights; and the creation of ethical codes, validation committees, and university frameworks that regulate the use of these tools and prevent academic dishonesty [6], [11], [14]. Transparency and accountability are essential for ensuring that institutions oversee these tools and promote fairness and impartiality [20], [25], [27].

Finally, to guarantee equity and inclusion, it is necessary to address the digital divide and biases directed toward non-English-speaking populations or people with disabilities. Since many models are trained primarily in English, solutions must be developed for diverse languages, and accessibility must be improved through tools such as text-to-speech and simplified interfaces. Under principles of universal design, AI can support inclusion without replacing human judgment or empathy in education [8], [20], [23], [28], [29].

Digital literacy should be understood as part of students' comprehensive development, not only as a technical skill. It includes effective information search, security and privacy, ethical and empathetic use, and awareness of data collection and terms of service [30].

Strengthening critical thinking entails understanding that language models are not sources of truth but probabilistic systems that can generate coherent yet incorrect information [1], [9]. To prevent overconfidence in their outputs, the EIPP-CT model is proposed, which promotes interpretive judgment and critical reflection. This approach promotes a pedagogy that questions algorithmic assumptions and encourages source verification to preserve human agency [31], [32].

Student autonomy is strengthened by using AI as an assistant that facilitates adaptive learning and self-regulation. Strategies such as interactive example machines allow the analysis and correction of AI responses, promoting learning from errors. In this process, teacher training is essential for integrating these tools into pedagogical design and avoiding restrictive approaches [5], [1], [32].

Regarding digital ethics, the proposal is to move toward a post-plagiarism era that recognizes hybrid human-AI writing and redefines authorship. To this end, transparency-based protocols are proposed in which the use of AI is declared and results are verified with reliable sources. In addition, institutions must implement flexible regulatory frameworks and promote ethical training and originality in the educational process [5], [8], [1], [11], [30].

The integration of these systems makes it possible to automate the assessment of essays, exams, and written assignments, optimizing teaching time and enabling faculty to focus on higher-order cognitive activities. In addition, generative models trained on human-evaluated data can achieve high accuracy in identifying characteristics of well-written texts, yielding results comparable to those of human experts [2], [29], [21], [34].

The use of generative artificial intelligence as a dynamic evaluator makes it possible to offer formative and personalized feedback tailored to students' needs. This supports continuous monitoring and self-regulation of learning, identifying deficiencies and suggesting improvements. Furthermore, it acts as a writing assistant, facilitating the comparison and correction of texts to iteratively strengthen learning [17], [21], [36], [31], [32].

Learning analytics based on these technologies enables the detection of trends and the prediction of academic performance, facilitating early identification of areas for improvement. Furthermore, it acts as an artificial evaluator by analyzing specific skills and proposing adaptive lessons centered on students' needs, promoting more inclusive and effective learning [37], [38], [1], [39].

Despite their advantages, instructors must maintain an expert role to validate the results of these systems, since they are based on probabilities and may generate errors or inconsistencies. Although they can serve as dynamic evaluators, their use requires adjusting academic integrity policies and strengthening digital literacy to critically interpret the information generated by AI [31], [1], [8], [16].