The Role of AI in Application Development: A Comparative Study with Manual Coding

Abstract

Application development has traditionally depended on manual coding, in which developers write every line of code themselves. This approach offers precision and control, but it is time-consuming, labour-intensive and prone to human error. The emergence of Artificial Intelligence (AI) has introduced tools that generate, test, debug and optimise code, raising the question of how AI-assisted development actually compares with manual practice. This paper presents a comparative study of the two approaches using secondary data from the Stack Overflow Annual Developer Survey 2024, comprising 65,437 responses from developers across 185 countries. Seven hypotheses were formulated covering productivity, job satisfaction, accuracy, compensation, challenges, sentiment and instrument reliability, and were tested using non-parametric methods (Mann-Whitney U, chi-square) at a 5% significance level. The job-satisfaction scale demonstrated excellent internal consistency (Cronbach\'s alpha = 0.931, 9 items, n = 29,095). The analysis found that 57.6% of respondents currently use AI tools, that 81.0% of adopters identify increased productivity as a benefit, and that 72.0% hold a favourable or very favourable view of AI. However, two widely assumed advantages did not survive testing. The difference in job satisfaction between AI users and manual coders was statistically significant but negligible in magnitude (means 6.97 vs 6.89; Cohen\'s d = 0.039). The apparent compensation advantage reversed direction once national context was controlled: pooled data showed manual coders earning more, yet within the United States alone the difference disappeared entirely (p = 0.203), indicating that the pooled gap is a confound arising from higher AI adoption in lower-income economies rather than an effect of AI itself. Trust remains the principal barrier, with 65.1% of respondents distrusting AI output and 61.9% reporting that AI tools lack context of their codebase. The study concludes that AI meaningfully augments developer productivity but does not yet demonstrably improve satisfaction or earnings, and that a hybrid human-AI model, supported by governance and training, remains the most defensible direction for application development.

Introduction

This text examines the emergence of AI-driven cybercrime and autonomous malware and argues that rapid advances in artificial intelligence are creating cybersecurity threats that existing legal and regulatory frameworks do not fully address. The paper combines cybersecurity research with doctrinal, comparative, and policy-oriented legal analysis to examine technological threats, criminal responsibility, jurisdiction, evidence, and possible regulatory reforms.

1. AI-Driven Cybercrime and Autonomous Malware

AI-driven cybercrime refers broadly to cyber offences enhanced by AI, including automated phishing, deepfake fraud, adaptive ransomware, vulnerability discovery, and AI-assisted attacks. Autonomous malware is a more specific category in which malicious software contains AI capabilities that allow it to make decisions, adapt its behavior, and pursue objectives with limited or no direct human control.

The text identifies three defining characteristics of autonomous malware:

• Context-aware decision-making
• Real-time behavioral adaptation
• Decentralized or autonomous operation

Unlike traditional malware that follows predetermined instructions, autonomous malware can evaluate its environment and change tactics, potentially making it more difficult to detect and attribute.

2. Technological Threats

AI can potentially be incorporated throughout the cyberattack lifecycle, including:

• Reconnaissance and vulnerability discovery
• Social engineering and phishing
• Privilege escalation and lateral movement
• Persistence and evasion
• Data exfiltration
• Ransomware operations
• Botnet coordination
• Deepfake-enabled fraud
• Cyberterrorism and attacks against critical infrastructure

AI-enhanced ransomware and botnets could potentially automate target selection, adapt to defensive measures, and operate at much greater scale. The major concern is that increasing autonomy may reduce the predictability of malicious behavior.

3. Legal Responsibility

A central legal issue is who should be held responsible when autonomous software commits a crime.

Current criminal-law frameworks generally do not treat AI systems as independent legal persons. Instead, AI and autonomous malware are treated as instruments used by human actors. Responsibility can therefore potentially be attributed to developers, operators, deployers, or other individuals depending on factors such as:

• Intent
• Knowledge
• Foreseeability
• Control
• Benefit obtained from the attack

Cases involving the Morris worm, Ancheta botnet, and Brovko illustrate the traditional principle that humans remain responsible even when malicious software performs actions autonomously.

However, increasingly autonomous AI systems create more difficult questions. If an AI system changes its behavior or discovers an attack path that its creator did not specifically anticipate, determining intent, foreseeability, and control becomes considerably more complicated.

4. Attribution and Evidence

AI-driven attacks also create significant forensic and evidentiary challenges. Investigators may need to analyze enormous quantities of logs, malware artifacts, network activity, and AI-generated actions.

Attribution becomes harder because attackers can use:

• Autonomous systems
• Synthetic identities
• False flags
• AI-generated misinformation
• Distributed infrastructure
• Rapidly changing attack behavior

The use of AI in forensic investigations creates another challenge: courts may need to determine whether evidence produced by potentially opaque AI systems is sufficiently reliable, transparent, and scientifically valid.

5. International Legal Gaps

The paper examines major international cybersecurity instruments, particularly the Budapest Convention on Cybercrime and the United Nations Convention against Cybercrime (Hanoi Convention).

These frameworks provide important foundations for criminalizing cyber offences and facilitating international cooperation. However, the text argues that they were developed before today's widespread AI-enabled threats and therefore do not specifically address issues such as:

• Autonomous malware
• AI-driven ransomware
• Deepfake-enabled cybercrime
• AI-enhanced cyberterrorism
• AI-specific attribution
• AI-related evidentiary problems

Their generally technology-neutral approach allows them to cover many AI-enabled offences, but the paper argues that technology neutrality alone may not adequately address the unique risks created by autonomous systems.

6. Need for an AI-Aware Cybersecurity Framework

The proposed solution is not to replace existing cybercrime law entirely but to supplement technology-neutral legal frameworks with AI-specific requirements.

The paper proposes principles including:

• Clear accountability chains connecting developers, deployers, and operators
• AI risk assessment and governance
• Transparency and explainability requirements
• Secure AI development practices
• Behavioral monitoring and anomaly detection
• Stronger forensic standards
• Improved international evidence sharing
• Greater cooperation between governments and industry

The framework should remain flexible enough to accommodate rapidly changing AI technology.

7. Policy Recommendations

The paper recommends reforms at several levels.

Internationally:

• Develop an AI-and-cybersecurity treaty or dedicated protocols/annexes.
• Establish common definitions and minimum standards for AI-driven cybercrime.
• Improve international cooperation and electronic-evidence sharing.
• Consider victim compensation mechanisms.
• Include technical experts and civil society in regulatory development.

Domestically:

• Clarify criminal responsibility for AI-enabled offences.
• Consider aggravating circumstances when AI significantly increases the scale or sophistication of an offence.
• Train judges, prosecutors, and investigators in AI and autonomous malware.
• Strengthen cyber-forensics capabilities.

Technically:

• Establish secure AI development and testing standards.
• Require appropriate safeguards and fail-safes for high-risk AI systems.
• Strengthen vulnerability disclosure practices.
• Monitor dual-use AI tools.
• Deploy behavioral AI-based detection in critical infrastructure.
• Improve public-private threat intelligence sharing.

Conclusion

This study set out to compare AI-assisted and manual application development using large-scale secondary data, and to test rather than assume the benefits attributed to AI. Analysis of 65,437 responses from the Stack Overflow Annual Developer Survey 2024, with a measurement instrument of demonstrated reliability (Cronbach\'s alpha = 0.931), supports three conclusions. AI adoption is now majority practice and is perceived as productive. A majority of developers use AI tools, 81.0% of adopters report increased productivity, and 72.0% view AI favourably. The direction of travel is not in doubt. The benefits are narrower than commonly claimed. Neither of the two outcome advantages tested survived examination. The job-satisfaction difference between AI users and manual coders, while statistically significant, is negligible in magnitude (Cohen\'s d = 0.039) with identical medians. The apparent compensation advantage reverses in pooled data and disappears entirely when national context is controlled (p = 0.203), a clear instance of Simpson\'s paradox. The evidence supports AI as a productivity aid, not as a driver of improved satisfaction or earnings. Trust, not capability, is the binding constraint. With 65.1% of respondents distrusting AI output and 61.9% citing absent codebase context, the obstacle to deeper integration is verification rather than generation quality. This finding aligns with the correctness and security literature and indicates where effort is best directed. The broader implication is methodological as much as substantive. Where evidence about AI is frequently asserted rather than tested, this study demonstrates that two widely repeated claims do not hold under controlled analysis. AI is not replacing human developers, and on the available evidence it is not yet making them measurably more satisfied or better paid. It is making them faster at routine work, which is a real and valuable outcome, and it is doing so within a hybrid model in which human judgement retains responsibility for architecture, security and verification. That model, supported by governance, training and better verification tooling, represents the most defensible direction for application development.

References

[1] L. Banh, F. Holldack, and G. Strobel, “Copiloting the future: How generative AI transforms software engineering,” Information and Software Technology, vol. 183, art. 107751, 2025. [2] A. Odeh, N. Odeh, and A. S. Mohammed, “A comparative review of AI techniques for automated code generation in software development: Advancements, challenges, and future directions,” TEM Journal, vol. 13, no. 1, pp. 726–739, 2024. [3] D. Cotroneo, A. Foggia, C. Improta, P. Liguori, and R. Natella, “Automating the correctness assessment of AI-generated code for security contexts,” Journal of Systems and Software, vol. 216, art. 112113, 2024. [4] M. F. Kharma, S. Choi, M. Alkhanafseh, and D. Mohaisen, “Security and quality in LLM-generated code: A multi-language, multi-model analysis,” arXiv preprint arXiv:2502.01853, 2025. [5] V. Terragni, A. Vella, P. Roop, and K. Blincoe, “The future of AI-driven software engineering,” ACM Transactions on Software Engineering and Methodology, 2025. [6] V. Jackson, B. Vasilescu, D. Russo, P. Ralph, M. Izadi, and R. Prikladnicki, “Creativity, generative AI, and software development: A research agenda,” arXiv preprint arXiv:2406.01966, 2024. [7] C. Negri-Ribalta, R. Geraud-Stewart, A. Sergeeva, and G. Lenzini, “A systematic literature review on the impact of AI models on the security of code generation,” Frontiers in Big Data, vol. 7, art. 1386720, 2024. [8] Z. Xu, C. Li, Q. Li, and S. H. Tan, “Defining ethically sourced code generation,” arXiv preprint arXiv:2507.19743, 2025. [9] M. Alenezi and M. Akour, “AI-driven innovations in software engineering: A review of current practices and future directions,” Applied Sciences, vol. 15, no. 3, art. 1344, 2025. [10] P. S. Khade and R. U. Sambhe, “Artificial intelligence in software development: A review of code generation, testing, maintenance and security,” International Journal of Current Science Research and Review, vol. 8, no. 4, 2025. [11] N. Upadhyaya, “The role of artificial intelligence in software development: A literature review,” Preprint, 2022. [12] S. Kumar, “Artificial intelligence in software engineering: A systematic exploration of AI-driven development,” International Journal of Innovative Research in Science, Engineering and Technology, vol. 13, no. 6, 2024. [13] F. Nyaga, “AI-driven software engineering: A systematic review of machine learning’s impact and future directions,” Preprints.org, 2025. [14] K. Qiu, N. Puccinelli, M. Ciniselli, and L. Di Grazia, “From today’s code to tomorrow’s symphony: The AI transformation of developer’s routine by 2030,” arXiv preprint arXiv:2405.12731, 2024. [15] I. Zakharov, E. Koshchenko, and A. Sergeyuk, “AI in software engineering: Perceived roles and their impact on adoption,” arXiv preprint arXiv:2504.20329, 2025. [16] A. A. Abbassi, L. Da Silva, A. Nikanjam, and F. Khomh, “A taxonomy of inefficiencies in LLM-generated Python code,” arXiv preprint arXiv:2503.06327, 2025. [17] N. Ashrafi, S. Bouktif, and M. Mediani, “Enhancing LLM code generation: A systematic evaluation of multi-agent collaboration and runtime debugging for improved accuracy, reliability, and latency,” arXiv preprint arXiv:2505.02133, 2025. [18] J. Oertel, J. Klünder, and R. Hebig, “Don’t settle for the first! How many GitHub Copilot solutions should you check?” Information and Software Technology, vol. 183, art. 107737, 2025. [19] M. F. Wong and C. W. Tan, “Aligning crowd-sourced human feedback for reinforcement learning on code generation by large language models,” arXiv preprint arXiv:2503.15129, 2025. [20] Y. Dong, X. Jiang, J. Qian, T. Wang, K. Zhang, Z. Jin, and G. Li, “A survey on code generation with LLM-based agents,” arXiv preprint arXiv:2508.00083, 2025. [21] J. Li, H. Le, Y. Zhou, C. Xiong, S. Savarese, and D. Sahoo, “CodeTree: Agent-guided tree search for code generation with large language models,” Univ. of Texas at Austin and Salesforce Research, 2025. [22] A. Ramadan, H. Yasin, and B. Pekta?, “The role of artificial intelligence and machine learning in software testing,” Üsküdar University and Dhofar University, 2024. [23] J. Maarleveld, Y. Yang, S. Wang, and A. Bacchelli, “A systematic mapping study on graph machine learning for static source code analysis,” Information and Software Technology, vol. 183, art. 107722, 2025. [24] M. A. Job, “Automating and optimizing software testing using artificial intelligence techniques,” International Journal of Advanced Computer Science and Applications, vol. 12, no. 5, 2021. [25] P. Narvekar, R. Maurya, S. Parab, and V. Prasad, “CODE-GENIE: An AI-powered code generation model,” International Research Journal of Modernization in Engineering Technology and Science, vol. 7, no. 3, 2025. [26] C. Hymel, “The AI-native software development lifecycle: A theoretical and practical new methodology,” Crowdbotics, 2024. [27] Stack Overflow, “Stack Overflow annual developer survey 2024,” [Data set], 2024. [Online]. Available: https://survey.stackoverflow.co/2024/

Copyright

Copyright © 2026 Perseus Bhavnagri. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.