Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology

ISSN NO-2584-2706

A Study: Code Review in Software Development

using AI
Dhanashri Thakur; Gaurav Talse; Yogesh Sonvane
Dept. Master in Computer Application, GHRCEM, Nagpur, India

Abstract Code Security, Deep Learning, Static

A critical stage in software development, Analysis, Developer Productivity.
code review guarantees the quality,
security, and maintainability of the code. 1. Introduction
Traditional bottlenecks in the development Code review is a crucial component of the
cycle and are laborious and prone to programdevelopmentprocess as they identi
human error. Code reviews are now more fy errors, security vulnerabilities, and
accurate and efficient thanks to automated inconsistencies in the code ensuring
solutions brought about by the its reliability, accessibility.
development of artificial intelligence (AI). and confidentiality. Code has historically
AI-powered tools evaluate source code, been assessed manually by software
identify defects, pinpoint security flaws, developers and peers. Despite its
and suggest fixes using Machine Learning effectiveness, this approach
(ML) and Natural Language Processing is often laborious, prone to mistakes,
(NLP) techniques. This study investigates and influenced by human behaviour. The
the role of AI in automated code review increasingly complex nature of computer
systems by examining several AI-driven software can turn manual code audits into
tools and frameworks utilized in obstacles in the design process, delaying
contemporary software development. We releases and increasing costs. With the
assess the efficacy of deep learning-based advent of automated solutions enabled by
and conventional static analysis methods artificial intelligence (AI), source
in detecting anomalies in the code. evaluation techniques are becoming
According to experimental findings, AI- increasingly reliable and precise. Machine
enhanced review systems outperform learning (ML) and natural language
traditional techniques in terms of accuracy, processing (NLP) methods can
efficiency, and fewer false positives. In be employed by AI systems to analyse
this paper author demonstrates how AI- source code in real-time, detect faults, and
powered code review tools may improve suggest enhancements. Software
software quality and developer Engineering (SE) is one of the numerous
productivity. Lastly, discussion of the processes where Artificial Intelligence
shortcomings of existing AI models and (AI) is increasingly being integrated. SE
propose ways to enhance automated code has been a human activity for many years,
review procedures in the future, such as although many processes have been
integrating AI with human-in-the-loop mechanized. The adoption of AI into a
techniques to create more effective socio-technical procedure like SE, which
feedback systems. has historically relied on human
interaction, control, and decision-making,
Keywords could signal a paradigm shift in software
Automated Code Review, Artificial engineering [1]. Code review is a quality
Intelligence, Machine Learning, Natural assurance procedure that involves
Language Processing, Software Quality, developers reviewing each other's code

IJMSRT25MAY97 www.ijmsrt.com 557

DOI: https://doi.org/10.5281/zenodo.15567600
Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

modifications. It has several variations manual evaluations rely on developers’

depending on the business. Code review, knowledge. Machine learning and natural
sometimes known as Modern Code language processing are used by AI-driven
Review (MCR), began as formal code systems, such as DeepCode and Amazon
inspections and has since developed into a CodeGuru, to lower false positives and
more relaxed procedure. MCR is offer real-time feedback. Modern code
distinguished by its regularity, tool-based reviews combined with CI/CD workflows
nature, and informality. Knowledge allow for continuous quality assurance. A
exchange, learning, flaw detection, and thorough explanation of Google's well-
code improvement are the process's established, change-based, and tool-
common advantages [2]. assisted code review procedure can be
found here. At least one additional
2. Background developer must review any modification
For an extended period, automated code made to the codebase. Tens of thousands
auditing has been an essential element of of developers participate in the review
software development. Basic rule-based process each day as both code authors and
advice was provided by early tools such as reviewers, and tens of thousands of
SonarQube, Checkstyle, and Find Bugs, modifications are made to the codebase
which identified syntax errors, security [3].
vulnerabilities, and technical infractions.
Although these tools depended on
established norms to enforce acceptable
practices, they frequently had significant
false positive rates and lacked contextual
comprehension of the code. As Artificial
Intelligence (AI) and Machine Learning
(ML) advanced, better automated code
review techniques appeared. DeepCode,
Codacy, and CodeScene are examples of
contemporary AI-driven applications that Fig1. AI-Powered Automated Code
use deep learning, large code datasets, and Review Workflow
Natural Language Processing (NLP) to
produce context-aware recommendations. 2.2 Presentation of Review Results
Unlike conventional static analysis tools, AICodeReview displays the review results
AI-powered solutions are able to recognize in an easy-to-use interface after the
logical errors, examine source code analysis is finished, offering detailed
patterns, and suggest major enhancements explanations and precise recommendations
based on past learning. for improving the code. To help
developers make well-informed changes,
2.1 Code Review these suggestions are produced using
In the creation of software, code scrutiny industry standards, coding best practices,
is an essential phase that ensures code and AI-driven insights. By analysing
quality, safety, maintainability. It entails syntax, logic, security flaws, and
carefully examining source code before performance problems, the system
delivery to find mistakes, security holes, provides project-specific, context-aware
and performance problems. The approach recommendation. In contrast to
includes human peer evaluations, conventional static analysis tools that
automatic reviews driven by AI, and static depend on preset rules, AICodeReview
analysis tools. While SonarQube and other uses Natural Language Processing (NLP)
static analysis tools enforce preset criteria, and machine learning to deliver intelligent,

IJMSRT25MAY97 www.ijmsrt.com 558

DOI: https://doi.org/10.5281/zenodo.15567600
Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

adaptive feedback [5]. Each suggestion programming languages, and black-box

includes a detailed reasoning process, decision-making, developers may be
helping developers understand why a hesitant to depend on AI recommendations
change is needed and how it improves in automated code review. AI models
code quality. This enhances developer should include human-AI cooperation,
learning, trust, and efficiency in the review Explainable AI (XAI), and adaptive
process [6]. By integrating AI-driven learning techniques to increase confidence.
feedback within CI/CD pipelines and Building trust also requires establishing
IDEs, AICodeReview ensures real-time moral standards and guaranteeing
code analysis, reduced manual effort, and objective AI conduct.
higher code reliability. This makes
software development more efficient, 2.4 Challenges in AI-Powered Code
scalable, and secure, reinforcing AI’s role Review
in modernizing automated code review. Although AI-powered code review
solutions improve automation, correctness,
2.3 Trust on AI and efficiency, a number of obstacles
Widespread use of Artificial Intelligence prevent their widespread use. Reliability is
(AI) in vital fields including code review, decreased by false positives and false
health care, banking, and robotics relies on negatives since AI may miss real problems
public trust in the innovation. To win over or mistakenly flag correct code.
consumers, AI models must be objective, Developers are reluctant to accept AI
clear, and trustworthy. However, recommendations due to deep learning
confidence is affected by issues including models lack of explainability. When AI
verdict reliability, bias in training data, and models trained on one language have
explanations. Because trust in AI is trouble with another, generalization
defined as the user's belief that "an agent problems occur. Aligning AI tools with
will help achieve an individual's goals in a IDEs, CI/CD pipelines, and version control
situation characterized by uncertainty and systems presents integration issues. AI
vulnerability," it is particularly crucial models should allow multi-language
when users are involved in high-stakes adaptation, integrate Explainable AI
situations where errors could have serious (XAI), and facilitate human-AI
consequences. collaboration for improved decision-
making in order to address these problems.
By addressing these issues, AI-driven code
No Trust Low Trust
review will become more effective, secure,
Moderate trust High Trust and trustworthy [11,13].
10%
2.4.1 Contextual Understanding and Best
40% 15%
Practices:
35% AI models often struggle to fully grasp
project-specific contexts, coding styles, or
domain-specific best practices. While they
Fig2. Developers Trust in AI-Powered can identify syntax errors and common
Code Review security flaws, they may not recognize
In contrast to reliance or obedience, which more nuanced design flaws,
are often examined as behaviours, maintainability issues, or architectural
confidence in AI is subjective and should inconsistencies.
be viewed as an attitude, according to a
review paper [4]. Because of false
positives, generalization problems across

IJMSRT25MAY97 www.ijmsrt.com 559

DOI: https://doi.org/10.5281/zenodo.15567600
Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

2.4.2 Security and Ethical Concerns: AI-

AI-based code review tools may Traditional
Powered
Feature Code
inadvertently introduce security risks, such Code
Review
as leaking sensitive code snippets when Review
processing code in cloud-based AI models. Speed Slow Fast
Additionally, biased training data can lead Human
High Reduced
to AI models reinforcing existing coding Effort
biases or incorrect recommendations. Context-
Error
Limited aware and
Detection
automated
2.4.3.Human-AICollaborationandTrust:
False Optimized
Developers may be reluctant to rely on AI- Moderate
Positives using ML
generated code reviews, especially when
Self-
they contradict human intuition. Hardcoded
Adaptability learning &
Encouraging human-AI collaboration, rules
adaptive
where AI suggests improvements while Manual and
developers retain control over final Integration Automated
separate
decisions, is critical for adoption. with CI/CD & seamless
steps

3. Literature Survey Deployment automation powered by AI

Automated code review has gained guarantees optimal resource allocation and
significant attention in software reduces human error in production
engineering, particularly with the settings. Although these developments,
integration of Artificial Intelligence (AI). there are still issues with making sure AI-
This section reviews existing studies on generated suggestions adhere to industry
traditional code review methods, AI-driven norms, security regulations, and best
techniques, and their impact on software practices. Careful thought must be given to
quality. In software engineering, the ethical ramifications of AI-driven
generative AI has drawn a lot of interest, software engineering decision-making,
especially for automated code generation, especially with regard to bias,
refactoring, and debugging. Research has explainability, and responsibility [7].
indicated that AI-powered models that
have been trained on extensive code 4. Approaches
Repositories are capable of correctly The AI-powered code review system
anticipating such weaknesses and follows a structured approach that
recommending the best fixes. integrates machine learning, Natural
Machine learning models increase Language Processing (NLP), and static
productivity by lessening the cognitive code analysis to enhance software quality,
strain on developers, according to research security, and maintainability. The approach
in AI-assisted programming environments. focuses on accurate code evaluation,
Deep learning algorithms and Natural effective presentation of review results,
Language Processing (NLP) are used by and building trust in AI-generated
AI-powered code review tools to recommendations.
comprehend programming structures and
identify trends that point to bad coding 4.1 Gathering and Preparing Data
practices. AI capabilities have also been We gathered an extensive set of open-
added to automated testing frameworks, source software files from services such as
allowing for predicted failure analysis and GitHub, GitLab, and Bitbucket for the
intelligent test case development. purpose of developing and evaluating the
Table 1 AI model. To ensure diversity, the dataset
Traditional vs. AI-based Code Review includes code from multiple programming

IJMSRT25MAY97 www.ijmsrt.com 560

DOI: https://doi.org/10.5281/zenodo.15567600
Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

languages, including Python, Java, quality are quantitatively evaluated. The

JavaScript, and C++ [9,10]. capabilities and limits of AI technologies
like DeepCode, OpenAI Codex, and AI-
4.2 Model Training and Development powered static analysis frameworks are
The AI-powered code review system is investigated. Case studies of businesses
trained using deep learning models, using AI-driven software engineering
specifically transformer-based techniques are used to gather empirical
architectures like CodeBERT, to analyze data. Analysis is done on performance
source code, detect issues, and provide parameters such deployment success rates,
intelligent recommendations. The training code quality enhancements, and defect
process involves data preprocessing, detectionratesTogaugeefficiencyimprovem
supervised learning, transfer learning, and ents and error reduction, a comparison
reinforcement learning to ensure high between workflows enhanced by AI and
accuracy and adaptability across different conventional manual procedures is carried
programming languages [8,11,14]. out. The study also examines developer
viewpoints on AI adoption and evaluates
4.3 Metrics for Evaluation their level of confidence in
We employed the following evaluation recommendations and proposals for AI-
measures to gauge the efficacy of the AI- generated code. Interviews and surveys
powered code review system. with cybersecurity specialists, DevOps
Precision & Recall: Assesses how well professionals, and software engineers shed
false positives and false negatives are light on the practical applications of
balanced. generative AI in software development [7].
A number of performance indicators, such
4.4.IntegrationwithDevelopment as accuracy, execution time, false positive
Workflow rate, and precision-recall balance, are used
The AI-powered code review system is to assess how well AI-driven code review
designed to seamlessly integrate into works. AI-powered reviews and
modern software development workflows, conventional static analysis tools are
ensuringcontinuous feedback,automation, compared to gauge advances in bug
and efficiency. By embedding AI into identification rates and efficiency.
CI/CD pipelines, IDEs, and version control Additionally, case studies of companies
systems, developers receive real-time implementing AI-driven code review are
insights on code quality, security used to evaluate the system's impactThese
vulnerabilities, and best practices [9]. studies examine important business
indicators like CI/CD pipeline
5. Methodology acceleration, defect detection rates, and
The proposed AI-powered automated code improvements in code quality [16]. To
review system follows a structured assess developer confidence in AI
methodology comprising data collection, recommendations, software engineers,
preprocessing, model training, evaluation, DevOps specialists, and cybersecurity
and integration with development specialists are surveyed and interviewed.
workflows. This section outlines the step- Modern IDEs, version control systems,
by-step process used to build and CI/CD pipelines are connected with
and evaluate the system. In order to assess the AI-powered code review system for
the efficacy of generative AI in automated smooth adoption. By automating code
code review, testing, and deployment, this analysis preliminary to pull request
study takes a multifaceted approach. merging, real-time security checking, and
Existing AI models are analysed compliance implementation, the AI
qualitatively, and their effects on software provides CI/CD pipelines supported by

IJMSRT25MAY97 www.ijmsrt.com 561

DOI: https://doi.org/10.5281/zenodo.15567600
Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

platforms such as Jenkins, GitLab CI/CD, development. Ultimately, AI-driven code

and GitHub Actions [17]. It offers review will become an integral part of
refactoring advice, auto-correction, and CI/CD pipelines, ensuring high-quality,
real-time inline suggestions in IDEs like secure,andmaintainablecode, thereby trans
VS Code, IntelliJ IDEA, and PyCharm. forming the future of software
Because AI-powered feedback is development [12,16].
immediately integrated into development
workflows, the technology promises 7. Conclusion
ongoing software quality tracking and This study concludes that the creation of
improvement. This process assures that AI- software is being transformed by AI-
powered autonomous code review is not powered automated code reviews that
only reliable and efficient but also scalable promote code reliability, safety and
and flexible, which will revolutionize accuracy. Although successful, traditional
software development by improving manual code audits are often difficult and
accuracy, security, and maintainability prone to human error. AI-driven solutions
while lowering human error and manual analyze source code, identify
labour. vulnerabilities, and offer valuable
suggestions with the help of Machine
6. Future Scope Learning (ML), Natural Language
By increasing precision, effectiveness, and Processing (NLP), and Deep Learning.
security, AI-powered automated code review This study indicates the superior accuracy,
is set to transform software development in speed and scalability of AI-based
the future. More intelligent, context-aware code technologies over classic static analysis
analysis will be enabled by advancements in methodologies. To increase adoption
Machine Learning, Explainable AI (XAI), and and dependability, however issues such as
Natural Language Processing (NLP). While explainability, security risks, false
human-AI collaboration will enhance positives, and biases in AI models must be
decision-making, future AI models will be addressed. Enhancing the effectiveness of
self-learning and adapt to project-specific AI-driven code review will require a
coding styles through continuous combination of explainable AI (XAI), self-
feedback. AI will evolve to not only detect learning systems, and human-AI
errors but also suggest and apply fixes collaboration. Future developments will
automatically, leading to intelligent auto- concentrate on automatic bug fixes, cross-
correction and code refactoring. Real-time language code reviews, security
AI-powered assistants will provide instant compliance, real-time AI support, and
feedback within IDEs, boosting developer ethical AI management.
productivity. Cross-language support and
domain-specific AI models will ensure 8. References
greater applicability across industries such [1] N. Ernst, A. Alami, "Human and
as finance, healthcare, and cybersecurity Machine: How Software Engineers
[13,14,15]. security and compliance will Perceive and Engage with AI-Assisted
be enhanced through AI-driven Code Reviews Compared to Their Peers,"
vulnerability detection, automated Department of Computer Science,
compliance enforcement, and blockchain- University of Victoria; Mærsk Mc-Kinney
based audit trails. Cloud and edge Møller Institute, University of Southern
computing will make AI-powered code Denmark [2025].
review scalable and accessible. Ethical AI [2] U. Cihan, V. Haratian, A. İcöz, M. K.
considerations, including bias reduction, Gül, Ö. Devran, E. F. Bayendur, B. M.
accountability, and intellectual property Uçar, and E. Tüzün, "Automated Code
protection, will shape responsible AI Review in Practice." [2024].

IJMSRT25MAY97 www.ijmsrt.com 562

DOI: https://doi.org/10.5281/zenodo.15567600
Volume-3, Issue-5, May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

[3] M. Vijayvergiya, "AI-Assisted [12] K. Sharma and A. Gupta, "Blockchain

Assessment of Coding Practices in Modern and AI for Secure Software Development:
Code Review," Google [2024]. A Future Perspective," Future Internet,
[4] R. Wang and D. Ford, "Investigating vol. 14, no. 3, pp. 1-20, [2022].
and Designing for Trust in AI-powered [13] S. Lin, J. Luo, and T. Wang,
Code Generation Tools," University of "Explainable AI for Code Review:
Washington, Seattle, WA, USA; Microsoft Improving Trust in Automated Systems,"
Research, Redmond, WA, USA [2024]. Proceedings of the IEEE International
[5] Y. Almeida, D. Albuquerque, E. D. Conference on Artificial Intelligence
Filho, and F. Muniz, "AICodeReview: (ICAI) [2021].
Advancing Code Quality with AI- [14] Z. Guo, Z. Yang, Y. Zhang, H. Sun,
enhanced Reviews." [2024]. and H. Peng, "CodeBERT: A Pre-Trained
[6] Resolving Code Review Comments Model for Programming and Natural
with Machine Learning. In International Languages," Findings of the Association
Conference on Software Engineering: for Computational Linguistics: EMNLP,
Software Engineering in Practice (ICSE- pp. 1536-1547, [2020].
SEIP) [2024]. [15] D. Vassallo, R. Bavota, G. Canfora,
Alexander Frömmgen, Jacob Austin, Peter and M. Di Penta, "Context-Aware Code
Choy, Nimesh Ghelani, Lera Kharatyan, Review Using Machine Learning,"
Gabriela Surita, Elena Khrapko, Pascal Proceedings of the 42nd International
Lamblin, Pierre-Antoine Manzagol, Conference on Software Engineering
Marcus Revaj, Maxim Tabachnyk, Daniel (ICSE) [2020].
Tarlow, Kevin Villela, Daniel Zheng, [16] J. Devlin, M. Chang, K. Lee, and K.
Satish Chandra, and Petros Maniatis. Toutanova, "BERT: Pre-training of Deep
[7] Generative AI for Automated Code Bidirectional Transformers for Language
Review, Testing, and Deployment, Bright Understanding," Proceedings of NAACL-
Matthew [2024]. HLT [2019].
[8] A. Roberts, H. W. Chung, G. Mishra, [17] M. Allamanis, M. Brockschmidt, and
A. Levskaya, J. Bradbury, D. Andor, S. E. T. Barr, "Learning to Represent
Narang, B. Lester, C. Ganey, A. Programs with Graphs," Proceedings of
Mohiuddin, et al., "Scaling Up Models and the International Conference on Learning
Data with T5X and SeqIO," Journal of Representations (ICLR) [2018].
Machine Learning Research, vol. 24, no.
377, [2023].
[9] J. Smith, K. Doe, and L. Brown,
"Integrating AI-Powered Code Review in
CI/CD Pipelines," Journal of Software
Engineering and Applications, vol. 15, no.
4, pp. 120-134, [2023].
[10] M. Pradel, T. Bernard, and B. Baudry,
"Deep Learning-Based Automated Code
Review," Proceedings of the 44th
ACM/IEEE International Conference on
Software Engineering (ICSE) [2022].
[11] M. Pradels, T. Bernard, and B.
Baudry, "Deep Learning-Based Automated
Code Review," Proceedings of the 44th
ACM/IEEE International Conference on
Software Engineering (ICSE) [2022].

IJMSRT25MAY97 www.ijmsrt.com 563

DOI: https://doi.org/10.5281/zenodo.15567600