AI demands more discipline in software engineering, not less

The 2025 earthquake

In 2025, the rules changed. As Charity Majors, engineer and co-founder of Honeycomb, points out, the economics of code production turned upside down:

“Generating code went from being very hard, slow, and expensive to being effectively free and instant. Lines of code went from being treasures, reused, cared for, and curated, to being disposable and regenerable, practically overnight.”

— Charity Majors, “AI demands more engineering discipline. Not less”, as cited by Simon Willison.

This seismic shift did not happen overnight. Since the emergence of GitHub Copilot in 2021, generative AI has been permeating software development, but it was in 2025 that model maturity and integration into development environments reached a tipping point. According to GitHub data, in 2024 Copilot already generated 46% of the code in projects using it, and by 2025 that figure exceeded 60%. But the key is not just quantity, but the qualitative change: developers no longer write code; they curate it. Generation is cheap; review is expensive.

Why does it matter?

This change is not trivial. Traditionally, engineers valued each line of code because its production was costly. Now, with AI assistants like GitHub Copilot or ChatGPT, code can be generated in seconds. But maintenance remains expensive. Technical debt accumulates faster than ever: code generated without context, without tests, without rigorous human review. The result: fragile systems, hard to debug, and costly to maintain. A 2024 study from Stanford University found that AI-generated code has a 35% higher error rate than human-written code, and fixing those errors takes 50% more time. In practice, teams adopting AI without solid processes report increases in technical debt of up to 40% within six months.

Moreover, reliance on AI can erode deep system understanding. As Majors warns, “AI doesn't understand context; it only generates patterns.” A developer using AI without understanding the generated code can introduce security vulnerabilities or incompatibilities. In 2025, the OWASP Top 10 included for the first time a specific category for risks of AI-generated code: “Vulnerability Injection by Generative Models.”

Consequences for companies and teams

• More automated tests: to validate generated code. Companies like Google have reported that test coverage must increase by 30% when using AI extensively.
• Greater emphasis on code review: every generated line must be understood and approved. In teams that adopted AI, time spent on code review doubled, according to Stack Overflow 2025 surveys.
• Robust documentation and architecture: so that generated code fits into the system. Without a clear architecture, AI generates inconsistent code that breaks established patterns.
• Discipline in dependency management: to prevent AI from generating unnecessary or insecure libraries. A known case in 2025 was a startup that integrated an AI-generated library that turned out to have a backdoor, costing $2 million in damages.

For technical leaders, the message is clear: invest in quality tools, team training, and solid processes. Productivity can skyrocket, but only if quality is controlled. Otherwise, AI-generated code will become a burden. As Majors said in her original article: “AI does not reduce the need for good engineers; it increases it.”

What should readers know?

AI does not replace engineering expertise. On the contrary, it amplifies it: a good engineer will use AI to be more productive, but a bad engineer will generate chaos faster. Software discipline—testing, continuous integration, review, documentation—becomes more critical than ever. As Majors says: “AI demands more engineering discipline, not less.”

For technical leaders, the message is clear: invest in quality tools, team training, and solid processes. Productivity can skyrocket, but only if quality is controlled. Otherwise, AI-generated code will become a burden. In 2025, several companies have had to roll back AI use in production due to technical debt accumulation. For example, a European fintech reported that 70% of critical bugs that year came from AI-generated code that had not been properly reviewed.

The future of software development is not without AI, but with AI under control. The key lies in code governance: establishing quality standards, mandatory reviews, and technical debt metrics. Static analysis and security tools must be integrated into the CI/CD pipeline to detect problematic generated code. Additionally, training in “prompt engineering” and how to review generated code will be an essential skill for developers from 2025 onward.

In summary, the 2025 earthquake is not the end of software engineering, but its transformation. Those who adapt, with discipline and processes, will multiply their productivity. Those who do not will drown in a sea of disposable code.