Artificial intelligence applied to the development of web and mobile applications encompasses a set of techniques, from natural language processing to generative models, that intervene at every stage of the software production cycle. Their adoption by development teams is now massive: the Stack Overflow Developer Survey 2025, conducted with over 49,000 respondents, indicates that 84% of developers are using or planning to use AI tools.
The striking fact lies elsewhere: only about 3% report having high confidence in them.
Gap between adoption and trust: what the data reveals about AI in development
This dissociation between widespread use and very low trust deserves to be highlighted first. Developers are integrating code assistants into their daily routines but systematically verify the outputs produced.
The DORA 2025 State of AI-assisted Software Development report, published by Google Cloud among nearly 5,000 professionals, associates the use of AI with an increase in delivery throughput. At the same time, it notes more instability in deployments. Teams are delivering faster, but post-deployment incidents are on the rise.
A randomized study conducted on experienced open-source developers found an increase in task execution time with certain AI tools, despite a subjective perception of time savings. The professionals involved possess advanced skills in application development and are familiar with the relevant codebase, which puts into perspective the idea of a universal accelerator.
Specialized platforms in web and mobile application design document this reality. The agency presented on https://www.codyx.org/ supports the creation of digital solutions while considering these constraints related to the reliability of AI-assisted code.
Security of AI-generated code: a documented structural risk
Competitors often present AI as a tool that enhances application security. The available data nuances this claim.
The Veracode 2025 GenAI Code Security Report states that 45% of tested programming tasks introduced a known vulnerability. This figure pertains to categories covered by standard security frameworks, not exotic flaws.
The issue is not isolated. Code generated by language models reproduces vulnerable patterns present in the training data. A developer who accepts a suggestion without manual analysis potentially integrates a flaw into their codebase. For publicly exposed web applications or mobile applications handling sensitive user data, this mechanism poses a structural risk.
- SQL injections and XSS vulnerabilities are among the most frequently introduced vulnerabilities by generated code, as models replicate examples from insecure tutorials
- Human code review remains the primary safety net, but it is often shortened when the team perceives the code as “validated” by the AI tool
- Static analysis tools (SAST) detect some of these flaws, provided they are systematically executed in the continuous integration pipeline
The security of produced solutions therefore depends less on the chosen AI tool than on the verification process maintained by the team.
European regulation and application development: the AI Act in practice
The European AI Act introduces obligations that directly impact the design of applications incorporating AI. The articles related to high-risk systems impose technical documentation, traceability of training data, and human oversight.
For companies developing web or mobile applications intended for the European market, compliance with the AI Act conditions production deployment. A personalized recommendation tool, a user scoring module, or an automated moderation system may fall into the “high risk” category depending on the context of use.
Obligations include maintaining a register of automated decisions and the ability to explain the model’s functioning to the relevant users. These requirements modify the software architecture from the design phase: it is necessary to plan for logging layers, audit interfaces, and manual deactivation mechanisms.
Impact on the development process
Integrating these regulatory constraints from the outset of the project is cheaper than adding them later. Teams that use AI to generate code must also document the automatically generated portion to trace the origin of unexpected behavior during an audit.
AI tools for development: what works and what remains fragile
Code assistants like GitHub Copilot or Cursor currently cover three main uses: real-time autocompletion, function generation from natural language descriptions, and analysis of existing codebases.
Autocompletion works well on repetitive code (API routes, forms, CRUD queries). For these tasks, the time savings are real and measurable. Generating unit tests from existing functions also reduces manual work, provided the relevance of the produced test cases is verified.
On the other hand, generating complex business logic remains fragile. A language model does not understand domain-specific rules (finance, health, logistics). The produced code may be syntactically correct but functionally wrong, which is more dangerous than a visible compilation error.
- For rapid prototyping and proofs of concept, AI significantly accelerates the initial design phase
- For maintaining large codebases, AI-assisted analysis helps understand undocumented legacy code
- For the architecture of distributed systems or fine-grained concurrency management, human oversight remains the determining factor
Artificial intelligence in web and mobile application development is neither a universal accelerator nor a marketing gimmick. Productivity gains exist in repetitive tasks and prototyping. The documented risks regarding the security of generated code and deployment instability impose a rigorous verification framework. Teams that make the most of these tools are those that treat them as assistants, not replacements.
