# The machines are rising: but developers keep the keys > Rumors of AI killing software development have been greatly exaggerated. Discover why the role of developers goes far beyond writing code. - Date : 2025-04-02 - Lecture : 10 min - Catégorie : genai - Tags : Generative AI, Emerging Tech, Development, Future of Work - URL : https://www.adservio.fr/en/insights/articles/les-machines-se-levent-mais-les-developpeurs-gardent ## TL;DR - Vibe coding (code generated through prompts until it "seems to work") increases duplicated code, complexity and refactoring cost, while an augmented-engineering approach maintains quality and accelerates velocity. - GitClear data (2021-2024) shows a rise in code churn since AI assistants became widespread, with less code refactored and more copy-pasting. - AI remains promising beyond code generation: understanding legacy codebases, software engineering agents, modernization assistance. - "AI-friendly" code, clear naming, modularity, low duplication, improves the performance of AI assistants as much as that of human developers. - The developer's role is shifting from writing code to guaranteeing the quality, reliability and trustworthiness of systems, whether written by humans or AI. ## Introduction Rumors of the ongoing death of software development, killed off by AI, have been greatly exaggerated. In reality, we're at a fork in the road: one where we can either embrace the (currently) distant notion of fully automated software development, or instead recognize that a software developer's job is far more than simply writing lines of code. Two diverging trajectories are taking shape. On one side, vibe coding, a style of programming where code is generated through natural-language prompts until the results seem to work, accumulates code smells, technical debt grows exponentially, maintenance costs climb by 300 to 500%, and the bus factor becomes critical for lack of shared understanding of the resulting code. On the other, augmented engineering, where the developer acts as architect, validator and curator of AI-generated code, maintains code quality, accelerates velocity by 40 to 60%, while keeping technical debt under control and retaining knowledge within the team. The decision we make as an industry could have significant long-term consequences. Growing complacency around AI-generated code and a shift toward what's been called "vibe coding" will lead to code that is more error-prone, more expensive to run, and harder to change in the future. Hidden costs of vibe coding: duplicated code up 35% on average across 50 observed projects, cyclomatic complexity up 42% (critical thresholds reached faster), test coverage down 28% (tests generated but incomplete), runtime performance down 15 to 25% (suboptimal algorithms), and refactoring cost climbing 180% for lack of upfront structure. And if the devaluation of software development skills continues, we may even end up with a workforce lacking the skills and knowledge to fix things in the long run. ## The dangers of AI-generated code are already here: GitClear data The risks of AI-generated code aren't science fiction: they're with us today. Research conducted by GitClear earlier this year indicates that since AI coding assistants like GitHub Copilot became widespread (launched in 2022), code churn, which GitClear defines as "changes that were either incomplete or erroneous when the author originally wrote, committed and pushed them to the company's git repo",has increased significantly. ### What the GitClear data shows GitClear's metrics between 2021 and 2024 illustrate the scale of the phenomenon: the code churn rate rose from 12.3% to 18.7%, a 52% increase, while the share of moved lines, an indicator of refactoring, fell from 8.2% to 5.1%, a 38% drop. Over the same period, lines of code added rose from 100,000 to 147,000 per month (+47%), while deleted lines only grew from 22,000 to 23,000 per month (+5% only). The copy-paste ratio jumped from 6.8% to 11.2%, a 65% increase. The alarming "code churn" metric, GitClear also found a marked decrease in the number of lines of code that were "moved", a signal of refactoring (essentially the "care and feeding" needed to make code more efficient). In other words, since coding assistants were introduced, there has been a pronounced increase in lines of code without a commensurate increase in lines deleted, updated or replaced. At the same time, there has been a decrease in moved lines, indicating that a lot of code has been written but not refactored. The delete-to-add ratio has thus fallen from 0.22 to 0.16, a sign that code is accumulating rather than being replaced, with a strong correlation (R² = 0.87) between Copilot usage and churn rate. More code isn't always better, More code is often not a good thing (sometimes quite the opposite); GitClear's findings ultimately point to complacency and a lack of rigor regarding code quality. At Adservio, we observe this pattern in our client projects too. Teams adopting AI assistants without clear guidelines do produce more code, but that code often requires more maintenance, has more bugs, and is harder for future developers to understand. This technical debt accumulates quickly, and the initial time savings from AI code generation are often wiped out by the extra time needed to debug, refactor and maintain the generated code. ## Should we set AI aside in software development? This doesn't mean we should set AI aside in software development and delivery. On the contrary, there's plenty to be excited about. As shown by the technology radars that leading tech consultancies regularly publish, syntheses of the technologies and practices observed across hundreds of client projects worldwide, the coding-assistance space is full of opportunity. Promising tools, We've noted Cursor, Cline and Windsurf, tools that enable what are called software engineering agents. In practice, this looks like an agent-style feature inside developer environments that makes it possible for developers to request that a certain set of coding tasks be carried out via a natural-language prompt. It's still very much a human/machine partnership, whatever the worst of the hype might have you believe. ### Beyond code generation That said, focusing solely on code generation misses the variety of ways AI can help software developers. We've been interested for a while, for example, in using generative AI to understand legacy codebases, and see a lot of promise in tools like Unblocked, an AI team assistant that helps teams do exactly that. In fact, last month we explored how Anthropic's Claude Code could help us add support for new languages in a tool called Code Concise, which we use to understand legacy systems. Results were mixed, but there's real promise there. At Adservio, we advise our clients to think of AI as one tool among several in their development toolbox, not as a replacement for human judgment and technical expertise. > Related read: [How Much Faster Can AI Coding Assistants Really Make Software Delivery?](https://www.adservio.fr/en/insights/articles/a-quelle-vitesse-les-assistants-de-codage): A heuristic, a 150-ticket case study, and 2025-2026 research: real speed gains from AI coding assistants sit between 5 and 15%, far below the marketing hype. ## Tightening our practices to get more out of AI It's worth remembering that a large proportion of developers' work isn't building something new from scratch. A large share of their work is evolving and adapting existing (and sometimes "legacy") software. ### The reality of legacy codebases Sprawling, rickety codebases that have accumulated technical debt are, unfortunately, the norm. Simply applying AI to them is likely to make things more rickety, not less. Approaches like vibe coding will only make things worse. That's why developer judgment will become more critical than ever. Several of these technology radars have recently highlighted AI-friendly code design, a conclusion consistent with our experience at Adservio: AI coding assistants perform better with well-structured codebases. ### The principles of AI-friendly code In practice, this requires a number of different things. It includes clear, expressive naming, to ensure context is clearly communicated (essential for code maintenance),reducing duplicated code, and ensuring modularity and effective abstractions. Taken together, these elements all help make code more readable for AI systems. Good coding practices are too easily neglected when productivity and efficiency are measured purely in terms of output: that was true before we had AI tooling too, and it needs to be resisted. At Adservio, we've developed an "AI Code Quality Framework" that we use with our clients to assess and improve how ready their codebases are for AI assistants. This framework evaluates dimensions such as: Naming consistency: Are names descriptive and do they follow consistent conventions?. Modularity: Is the code broken down into modules with clear responsibilities?. Documentation: Are there comments and documentation explaining the "why", not just the "what"?. Test coverage: Are there enough tests to validate expected behavior?. Technical debt: What's the level of technical debt and where is it concentrated? ## AI assistance demands greater human accountability Instagram co-founder Mike Krieger recently claimed that in three years, software engineers will no longer write any code: they will only review code created by AI. That may sound like a huge claim, but it's worth remembering that reviewing code has always been a major part of a software developer's job. Code review as a core skill, With that in mind, perhaps the evolution of software development won't be as dramatic as people like Krieger make it sound. But there's another argument that as AI becomes embedded in how we build software, software developers will take on more responsibility, not less. This is something we've discussed extensively at Adservio: the job of verifying that a system built by AI is correct will fall to humans. Yes, the verification itself could be AI-assisted, but it will be the software developer's role to guarantee trust. The rise of the Chief Trust Officer, In a world where trust becomes highly valuable, as evidenced by the emergence of the Chief Trust Officer, the work of software developers is even more critical to the infrastructure of the global industry. This means it's vital that their work is valued: the impact of thoughtless automation and "pure vibes" could prove incredibly problematic (and costly) in the years ahead. ## The developer as guardian of quality At Adservio, we encourage our clients to rethink the developer's role, not as someone who writes code, but as someone who guarantees the quality, maintainability and reliability of software systems, whether that code is written by humans, by AI, or by a combination of both. ### New responsibilities This shift in role brings new responsibilities: Validation and verification: Understanding not just whether the code works, but whether it solves the right problem in the right way. Architecture and design: Ensuring structural decisions are made with a long-term view, not just to solve the immediate problem. Knowledge transfer: Documenting decisions, explaining trade-offs, and maintaining context for future maintainers (human and AI). Ethics and accountability: Ensuring the systems they build are not only functional but also ethical, safe and aligned with organizational and societal values. The critical skills of the future, The skills that will become most valuable aren't necessarily the ones we value most today. Typing speed or syntax memorization matter less. What matters is: Systems thinking: Understanding how the pieces fit together and how changes propagate. Judgment: Knowing when to trust AI and when to step in. Communication: Explaining technical decisions to non-technical people and vice versa. Continuous learning: Adapting quickly to new tools, paradigms and challenges > Related read: [How to Build Trust With AI-Powered Coding Assistants](https://www.adservio.fr/en/insights/articles/comment-cultiver-la-confiance-avec-les-assistants-de-codage): Why adoption rates for coding assistants stall, and what actually builds a team's trust in the tool. ## Conclusion: Humans remain at the center The machines are rising, indeed. AI is transforming software development, and that transformation will accelerate. But the idea that developers are becoming obsolete is not only wrong, it's dangerous. The future of software development, The future of software development is not a future without developers. It's a future where developers are augmented, empowered and, crucially, more important than ever. Their role is shifting from tactical execution to strategic oversight, from mechanical implementation to nuanced judgment. The organizations that recognize this, that invest in their developers, that value their expertise, that create environments where human judgment and machine capability complement each other, will be the ones that build the best systems. The call to action, For tech leaders: Don't let AI hype convince you to devalue your developers. Invest in their continuous training, create clear guidelines for AI use, and build a culture where quality matters as much as speed. For developers: Embrace AI as a tool, but don't become complacent. Keep honing your craft, insist on quality, and remember that your judgment and expertise are irreplaceable. Machines can write code, but only humans can guarantee it's trustworthy. Note: The statements and opinions expressed in this article are those of the author and do not necessarily reflect the positions of Adservio. ## FAQ ### What's the difference between vibe coding and augmented engineering? Vibe coding generates code through natural-language prompts until the result seems to work, without architectural validation, which accumulates code smells and technical debt. Augmented engineering places the developer as architect, validator and curator of AI-generated code, which maintains quality while accelerating velocity by 40 to 60%. ### What does GitClear's data show about AI-generated code? Between 2021 and 2024, the code churn rate increased by 52%, the share of refactored code (moved lines) fell by 38%, and the copy-paste rate jumped by 65%. This indicates code is being produced in volume without being sufficiently reworked, a sign of complacency about quality. ### Is the developer's role going to disappear as generative AI becomes mainstream? No. The article argues the opposite: the role is shifting from writing code to guaranteeing the quality, reliability and trustworthiness of systems, validation, architecture, knowledge transfer, and judgment about when to trust AI remain deeply human responsibilities.