Inteligencia artificial en la ingenieria de software

INTRODUCTION

The integration of large language models (LLMs) and generative artificial intelligence (GenAI) has driven a significant transformation in software engineering, making it possible to reduce the gap between human intent expressed in natural language and technical specifications, as they function as a universal translator, facilitating the transformation of unstructured inputs into actionable requirements, improving communication, and optimizing the development life cycle.

This study describes the state of the art in artificial intelligence applied to software engineering. It analyzes how the use of AI assistants allows for a transition from passive instruction models toward interactive collaboration environments that enhance critical competencies.

To delve deeper into this analysis, this document poses 6 questions that seek to explore the impact of these tools on productivity, mental workload, and the quality of the generated code.

Regarding its organization, the document is structured into the following sections: Methods and Research Questions, where the research approach is detailed; Results, which present the findings on the use of AI in technical education; and Conclusions, where best practices for training future software developers are synthesized.

METHODS

To conduct the systematic literature review presented in this article, the Software Engineering research process approach was used. This methodological approach was defined by Barbara Kitchenham and allows for the structured identification, evaluation, and synthesis of the available evidence on a research topic. The main objective was to analyze the impact of artificial intelligence on different cycles of software engineering, considering aspects such as requirements engineering, software maintainability, cybersecurity, competency development, and prompt engineering.

Results

The results obtained from the systematic literature review process are presented below, organized into 6 sections addressing the use of artificial intelligence in different phases of software development.

Impact of language models on the quality and management of software requirements.

The integration of language models (LLM) and generative artificial intelligence has driven a significant transformation in requirements engineering, acting as a universal translator that narrows the gap between human intent expressed in natural language and operational technical specifications. In the initial stages of capture and elicitation, the use of natural language processing (PLN) tools and conversational assistants enables the transformation of unstructured inputs from stakeholders into actionable requirements, facilitating the identification of key entities and the resolution of communication conflicts in real time [12], [13]. This conversational capability not only democratizes the development process by enabling non-technical users to participate through the generation of domain-specific languages (DSL), but also optimizes the software lifecycle by significantly reducing requirements production times in agile environments [14], [26].

With regard to quality assurance, methodological advances have enabled a shift from simple lexical analyses to deep semantic evaluations that address dimensions such as clarity, consistency, and completeness [15].

The application of vision and language models (VLM) represents a crucial innovation in this field, as it enables the automatic detection of visual defects in flowcharts and traffic scenarios, identifying local inconsistencies and spatial incompleteness that previously required error-prone manual reviews [12], [16]. Furthermore, the systematic alignment of user stories with standardized quality frameworks such as the INVEST model has been enhanced by the generative capability of LLM, allowing not only the detection of problems but also the suggestion of precise reformulations and the generation of acceptance criteria by simulating collaboration among roles [15].

The management and organization of large volumes of technical information also benefit from automation through scalable architectures; models based on architectures such as BERT have demonstrated high accuracy in the automatic classification of functional and non-functional requirements, surpassing traditional techniques by capturing contextual nuances and ambiguous phrases [12], [17]. Complementarily, the use of fuzzy logic and machine learning algorithms in prioritization tasks helps to manage the uncertainty and volatility inherent to dynamic projects, enabling constant reprioritizations without the prohibitive computational costs of manual methods [6].

Finally, language models contribute to traceability and systemic consistency by automating the mapping between various development artifacts, such as requirements, source code, and test plans [18]; the use of generative AI allows comparison of different views of the system—for example, the developer's code perspective versus the functionality perceived by the user—to identify semantic mismatches and ensure that all actors share a coherent understanding of the product [19]. However, the adoption of these technologies faces critical challenges related to model hallucinations, the lack of transparency in their decision-making processes, and the need to integrate specific technical knowledge of industry standards; these challenges underscore the importance of maintaining rigorous human oversight and formal validation processes [6], [12], [16], [17].

Effects of generative AI on technical debt and maintainability of developed software.

The implementation of generative artificial intelligence (AGI) in software development has an ambivalent influence on technical debt and maintainability, establishing a tension between the streamlining of tasks and the increase in structural complexity. On the one hand, analyses of the use of language models in software engineering projects indicate that the final code integrated in repositories tends to have a higher cyclomatic complexity, length and Halstead effort compared to the initial suggestions of AI, suggesting that human adaptation of these products tends to increase the logical density of the system [20].

This structural complexity is aggravated by the persistence of algorithmic errors that do not disappear with large language models (LLM), but instead are transformed into new forms of opacity dependent on data and prompt sensitivity, which demands a robust theoretical foundation to guarantee the reproducibility and validity of the [7] software.

Likewise, the presence of "hallucinations" or logically inconsistent results creates distrust in professional environments, compelling developers to perform exhaustive manual validation to prevent the introduction of vulnerabilities or failures that would compromise the future stability of the code [3]. Moreover, there is a latent risk of overreliance that could erode programmers' technical skills, hindering independent corrective maintenance in the long term as in-depth knowledge of system logic is diluted [8].

Nevertheless, GenAI and advanced techniques such as fuzzy logic offer advantages for managing volatile requirements and facilitating dynamic reprioritization, which is critical for avoiding software degradation caused by constant changes in scope [6]. In global development contexts, maintainability continues to depend on the standardization of documentation to reduce inconsistencies, a practice that can be reinforced by AI tools to improve communication in distributed teams [3]. Finally, the sustainability of AI-based platforms can be strengthened through the use of permissive open-source licenses that foster a community capable of contributing updates and new functionalities collaboratively [21].

Contribution of explainable AI to the resilience and automation of software testing.

First, explainability is essential to mitigate "algorithmic error," which arises when there is a mismatch between the software's capabilities and the system requirements; methodological transparency allows developers to understand why a text analysis or code-generation tool makes certain decisions, which facilitates the replicability of tests and the validity of results [7].

In the field of automatic code generation, the use of transparency tools, such as sequence salience maps (salience maps), helps users trust the results generated by language models (LLM) by making the tool's output process visible; open and transparent AI is essential in critical environments to ensure accountability and mitigate biases [8], [22].

Resilience in software testing is enhanced by AI-driven automation tools that include "self-healing" capabilities; tools such as Testim allow test scripts to automatically adjust to changes in source code, maintaining the reliability of continuous testing without constant human intervention [22]. Furthermore, the use of fuzzy logic techniques contributes to resilience by enabling systems to manage uncertainty and ambiguity in dynamic environments, adapting to frequent changes in project requirements [6]. On the other hand, the application of reference architectures and software forensics techniques enables the systematic evaluation and comparison of the integrity of complex systems, such as agent-based systems, reinforcing their structural stability [9].

Generative AI has revolutionized the creation of test cases and the identification of faults, significantly reducing time to market (time-to-market). The use of machine learning algorithms for predictive analytics also makes it possible to identify bottlenecks and prioritize error detection in critical areas of high complexity [22]; these algorithms have demonstrated superior capabilities in generating unit tests and in automated program repair (APR), and can correct errors in both human-written and machine-written code [20].

Strategies for guaranteeing cybersecurity and privacy in software development using AI.

To guarantee security, it is necessary to implement a multidimensional approach that spans system architecture, data management, and algorithmic validation; a fundamental strategy is to adopt the "local-first software" paradigm, which allows researchers and developers to maintain ownership and full control over their information by processing it directly in the browser or on local machines, minimizing risks associated with storage in third-party infrastructures or cloud servers [21]. Additionally, in critical sectors such as open banking, advanced data encryption techniques, cloud servers with high security standards, and access control systems based on user consent must be employed [22].

To mitigate privacy and copyright risks, it is crucial to configure AI tools to block the generation of code based on protected public sources; users must rigorously validate the generated code to verify the existence of omitted licenses and avoid ethical or legal infringements [8]; likewise, the use of Explainable Artificial Intelligence (XAI) is recommended to understand the logic behind model decisions, facilitating the audit of potential security failures [22].

The implementation of advanced prompt engineering makes it possible to refine the instructions given to AI, ensuring that the resulting code is more accurate and aligned with the technical requirements [20]. To combat "hallucinations" (false information generated by AI), external verification tools, input validation, and sequence saliency maps that increase the transparency of the results should be used. In addition, the use of strongly typed languages (such as TypeScript) in the development of AI interfaces helps to reduce the occurrence of logical errors and vulnerabilities [8], [21].

Development must align with models such as the Agent Systems Reference Architecture (ASRA), which defines specific components for the safety and survivability of the system [9]. It is essential to integrate AI-driven automated software testing that spans the entire software development life cycle (SDLC), including the use of "safety guardrails" based on natural language processing models to protect and supervise the outputs of AI models [22].

Finally, the selection of any AI analysis tool must be grounded in theories that ensure the validity of the constructs, avoiding algorithmic errors that compromise the integrity of the system [7].

Conclusions

Conducting a systematic literature review in the field of AI-assisted software engineering is essential due to the ambiguous nature of these technologies. Although language models (LLM) act as a universal translator that optimizes requirements capture and code generation, their implementation faces critical challenges such as "hallucinations" and a lack of transparency in their decision-making processes. Therefore, the importance of the review lies in the need to maintain rigorous human oversight that mitigates algorithmic errors and validates semantic consistency, ensuring that the final product meets standards and does not introduce security vulnerabilities.

Within the framework of this study, the review of the state of the art has made it possible to identify how the use of AI assistants transforms technical education into a competency-based learning model. The importance of this analysis lies in showing that tools such as "Vibe Coding" enable teachers to simulate expert reasoning, helping students close knowledge gaps in a self-directed manner; moreover, this examination makes it possible to understand that success in implementing these tools depends not only on technology, but also on the user's ability to counterbalance the assistants' synthesis with their own mental model, thus avoiding the weakening of technical skills due to overdependence.

Finally, the six research questions posed allow us to conclude that generative AI significantly optimizes productivity and software resilience through "self-repair" capabilities in testing; however, a persistent tension is confirmed, as generated code can increase cyclomatic complexity and technical debt if there is no manual intervention. Therefore, prompt engineering and techniques such as "Chain-of-Thought" are key competencies for guaranteeing accurate and secure results, allowing future programmers to evolve from imprecise queries toward structured technical instructions that protect system integrity.

References

1. 1 B. Kitchenham, «Procedures for Performing Systematic Reviews», Keele, UK, Keele Univ., vol. 33, ago. 2004.
2. 2 J. D. Martínez Díaz, V. Ortega Chacón, y F. J. Muñoz Ronda, «El diseño de preguntas clínicas en la práctica basada en la evidencia: modelos de formulación», Enfermería Global, vol. 15, n.o 43, pp. 431-438, jul. 2016.
3. 3 H. U. Khan, M. Niazi, M. El-Attar, N. Ikram, S. U. Khan, y A. Q. Gill, «Empirical Investigation of Critical Requirements Engineering Practices for Global Software Development», IEEE Access, vol. 9, pp. 93593-93613, 2021, doi: 10.1109/ACCESS.2021.3092679..
4. 4 D. Herremans, «aiSTROM–A Roadmap for Developing a Successful AI Strategy», IEEE Access, vol. 9, pp. 155826-155838, 2021, doi: 10.1109/ACCESS.2021.3127548..
5. 5 L. M. Amugongo, A. Kriebitz, A. Boch, y C. Lütge, «Operationalising AI ethics through the agile software development lifecycle: a case study of AI-enabled mobile health applications», AI Ethics, vol. 5, n.o 1, pp. 227-244, feb. 2025, doi: 10.1007/s43681-023-00331-3..
6. 6 N. Tasneem, H. B. Zulzalil, y S. Hassan, «Enhancing Agile Software Development: A Systematic Literature Review of Requirement Prioritization and Reprioritization Techniques», IEEE Access, vol. 13, pp. 32993-33034, 2025, doi: 10.1109/ACCESS.2025.3539357..
7. 7 J. Wobst, «Avoiding algorithm errors in textual analysis: A guide to selecting software, and a research agenda toward generative artificial intelligence», 2025.
8. 8 M. Basha, «Trust, transparency, and adoption in generative AI for software engineering: Insights from Twitter discourse», Information and Software Technology, 2025.
9. 9 W. C. Regli et al., «Development and Specification of a Reference Architecture for Agent-Based Systems», IEEE Trans. Syst. Man Cybern, Syst., vol. 44, n.o 2, pp. 146-161, feb. 2014, doi: 10.1109/TSMCC.2013.2263132..
10. 10 S. B. Nettur, S. Karpurapu, U. Nettur, y L. S. Gajja, «Cypress Copilot: Development of an AI Assistant for Boosting Productivity and Transforming Web Application Testing», IEEE Access, vol. 13, pp. 3215-3229, 2025, doi: 10.1109/ACCESS.2024.3521407..
11. 11 J. D. Weisz, J. He, M. Muller, G. Hoefer, R. Miles, y W. Geyer, «Design Principles for Generative AI Applications», en Proceedings of the CHI Conference on Human Factors in Computing Systems, Honolulu HI USA: ACM, may 2024, pp. 1-22. doi: 10.1145/3613904.3642466..
12. 12 M. Esposito et al., «Generative AI for software architecture. Applications, challenges, and future directions», Journal of Systems and Software, vol. 231, p. 112607, ene. 2026, doi: 10.1016/j.jss.2025.112607..
13. 13 Md. A. Haque, «LLMs: A game-changer for software engineers?», BenchCouncil Transactions on Benchmarks, Standards and Evaluations, vol. 5, n.o 1, p. 100204, mar. 2025, doi: 10.1016/j.tbench.2025.100204..
14. 14 N. Malamas, E. Tsardoulias, K. Panayiotou, y A. L. Symeonidis, «Toward efficient vibe coding: An LLM-based agent for low-code software development», Journal of Computer Languages, vol. 85, p. 101367, nov. 2025, doi: 10.1016/j.cola.2025.101367..
15. 15 E. Wolf, A. Trendowicz, y J. Siebert, «Quality assessment of software requirements using artificial intelligence methods: A systematic literature review», Information and Software Technology, vol. 191, p. 107979, mar. 2026, doi: 10.1016/j.infsof.2025.107979..
16. 16 N. Bukhary et al., «Few-Shot Evaluation of Vision Language Models for Detecting Visual Defects in Autonomous Vehicle Software Requirement Specifications», IEEE Access, vol. 13, pp. 117914-117942, 2025, doi: 10.1109/ACCESS.2025.3586554..
17. 17 H. Ali, «Cloud-based machine learning for scalable classification of software requirements: Insights from the PROMISE dataset», 2025.
18. 18 U. K. Durrani, M. Akpinar, H. Bektas, y M. Saleh, «Impact of Artificial Intelligence on Software Engineering Phases and Activities (2013–2024): A Quantitative Analysis Using Zero- Truncated Poisson Model».
19. 19 J. R. Laracy, Z. Meng, V. D. Kirova, C. S. Ku, y T. J. Marlowe, «Software Quality Assurance and AI: A Systems-Theoretic Approach to Reliability, Safety, and Security», Software, vol. 4, n.o 4, p. 30, nov. 2025, doi: 10.3390/software4040030..
20. 20 A. Naman, R. Shariffdeen, G. Wang, S. Rasnayaka, y G. N. Iyer, «Analysis of Student-LLM Interaction in a Software Engineering Project», 3 de febrero de 2025, arXiv: arXiv:2502.01273. doi: 10.48550/arXiv.2502.01273..
21. 21 A. P. Placitelli, S. Hochscheid, F. Maffucci, Al. F. Rees, G. Treglia, y A. Staiano, «Towards decentralized ecological monitoring: A local-first web-based AI platform for sea turtle detection», Ecological Informatics, vol. 93, p. 103569, feb. 2026, doi: 10.1016/j.ecoinf.2025.103569..
22. 22 L. S. Ismail, A. Salim Jamil, A. Amantur Dastanbekovich, I. H. Mohammed AlDosari, y K. Salman, «AI-Driven Automation for Transforming the Future of Software Development», پژوهشنامه پردازش و مدیریت اطلاعات, vol. 40, n.o ویژه نامه انگلیسی 4, jul. 2025, doi: 10.22034/jipm.2025.728106..
23. 23 W.-C. Chen y S.-Y. Chen, «TIPS: A Prompt Engineering Framework for Code Classification and Generation on Resource-Constrained Systems», IEEE Access, vol. 14, pp. 15723-15735, 2026, doi: 10.1109/ACCESS.2026.3658502..
24. 24 F. Pesce y W. Cheungpasitporn, «Vibe Coding in nephrology education: clinician-led, AI-assisted development of open-source interactive learning tools», Renal Failure, vol. 47, n.o 1, p. 2581933, dic. 2025, doi: 10.1080/0886022X.2025.2581933..
25. 25 N. Arshad, T. A. Butt, y M. Iqbal, «A Comprehensive Framework for Intelligent, Scalable, and Performance-Optimized Software Development», IEEE Access, vol. 13, pp. 74062-74077, 2025, doi: 10.1109/ACCESS.2025.3564139..
26. 26 S. Sarferaz, «Implementing Agentic AI Into ERP Software», IEEE Access, vol. 13, pp. 178945-178960, 2025, doi: 10.1109/ACCESS.2025.3621887..
27. 27 N. A. Mohd Salim, A. Azman, P. S. Sulaiman, N. S. Roslan, R. Hod, y S. K. Adam, «Development and validation of virtual patients for assessing history taking and communication skills among medical students», BMC Med Educ, vol. 26, n.o 1, p. 264, ene. 2026, doi: 10.1186/s12909-025-08503-x..
28. 28 M. Glas, C. Nirschl, B. Lanyado, y J. Van Niekerk, «Insecure by design? A human-centric security perspective on AI-assisted software development», Computers & Security, vol. 164, p. 104842, may 2026, doi: 10.1016/j.cose.2026.104842..
29. 29 F. Lonetti, A. Bertolino, P. Antonino, y D.-H. Bae, «Recent developments in software engineering for systems-of-systems and software ecosystems», Journal of Systems and Software, vol. 236, p. 112819, jun. 2026, doi: 10.1016/j.jss.2026.112819..