(versione italiana qua)
Let me say upfront — to pre-empt the usual hasty comments from those who won't read to the end — that I do believe tools based on generative Artificial Intelligence (GenAI, hereafter) are genuinely useful in software development, as they are in many other fields. But we need to be able to tell dreams from reality.
In 2023, Big Tech CEOs were more or less unanimously declaring that GenAI would replace most software developers by 2025. As recently as March 2025, Dario Amodei, CEO of Anthropic, was claiming that within 3 to 6 months GenAI would be writing 90% of all code.
The story went that the future would be "agentic" — that developers would have digital colleagues who never slept and never complained and would therefore take their jobs.
2024 closed with no fewer than 152,000 employees laid off across the tech sector worldwide, while by December 2025 the number of layoffs had reached 123,000 — all of it justified, according to the executives themselves, as "realigning towards an AI-centric future."
2026 has now begun, and there is no trace of the miraculous transformations that until recently were being touted as certainties — as I had anticipated several months ago. There I cited the Generative AI Gap report from MIT's NANDA project, published in July 2025, which had found that despite $40 billion in global investment, a staggering 95% of GenAI pilot projects had failed to produce any measurable economic return. Most organisations were seeing a net impact of zero on their bottom line.
Here are some updates that help explain why.
The GitClear report of February 2025 examined 211 million lines of code modified in the open source repositories used by Big Tech, finding that between 2020 and 2024 there was a near-doubling (from 3.05% to 5.67%) in the proportion of code that is added and then deleted within two weeks — so-called code churn. This is a clear indicator of a progressive decline in the quality of software being produced. The only technological or organisational change of any significance that occurred over the same period is that the proportion of developers using GenAI tools rose from zero to 63%. Correlation is not causation, of course, but this looks very much like a smoking gun. Furthermore, the report also recorded an increase in duplicated code blocks and a decrease in code reuse. These too are signs of declining quality. A December 2025 study by CodeRabbit, analysing 470 pull requests on the GitHub platform (used by more than 100 million developers worldwide to share and collaboratively develop code), found that GenAI-generated code produces on average 1.7 times more issues than code written by humans.
Stack Overflow, following a survey involving nearly 50,000 developers across 177 countries, reported that the proportion of developers favourably disposed towards GenAI tools fell from over 70% in 2024 to around 60% in 2025. The main source of frustration, cited by 66% of respondents, was that GenAI-produced solutions are "almost right, but not quite" — which leads to wasted time fixing errors in the code generated this way. That said, 69% of them acknowledged that GenAI does increase productivity in software writing. Several studies carried out in 2024 found that this increase ranges from 10–20% for senior developers to 35–40% for junior ones, who are better placed to benefit from a system that has knowledge of virtually all the software ever written in the world.
These findings are corroborated by studies conducted by the Stanford software engineering productivity research group involving more than 100 developers at major technology companies, which showed that the productivity gains in software development achieved through the use of GenAI tools are on average between 10% and 20%. Using them uncritically for software development, however, risks being a very costly mistake in the long run. While the volume of code produced increases by almost 40%, its quality decreases, requiring additional time for correction. And these are just the short-term consequences. Over the medium or long term — that is, in the context of the adaptive maintenance of software systems that have been in operation for some time — the situation risks becoming explosive.
There are no results yet pointing in this direction, given how little time has passed, but two significant findings emerge from the Stanford group's studies. The first is that without strong discipline in keeping a company's codebase structured and organised in a clear and clean manner, productivity gains evaporate or even reverse. The second is that productivity gains are greatest for new, low-complexity projects, while they diminish considerably for mature, high-complexity ones.
Throughout 2025, the wonders of vibe coding were extolled — the approach to software development in which a developer interacts with GenAI in natural language to "materialise" a complex software system by describing its overall vision. The problem is that this approach is fine for a demo but not for production-grade systems, because — like a sandcastle — it does not hold up solidly over time. This is the so-called GenAI technical debt: the future costs that accumulate when, in the rush of software development, one takes advantage of GenAI's code generation speed without bothering with thorough checks and verification.
During 2025, an approach emerged — conceptually known for some time, though now gaining traction — called specification-driven development, in which the developer uses GenAI tools starting from a specification — that is, a high-level description of the desired behaviour of the entire system — defined by the developer and progressively refined in ever greater detail, also with the support of GenAI. In this way, one partially overcomes the main limitation of these tools, which are based on statistics rather than symbolic representation: their inability to represent concepts that abstract away from the literal context under examination. With this approach, the software developer transforms into a specification developer. It is too early to say where this will lead, however, not least because the problems afflicting GenAI systems will not be overcome until the statistical approach is coupled with the symbolic one. The need to review AI-generated code in order to be confident it can be trusted therefore remains, at least until that integration occurs — and no one can reliably say when that will be. Which means that graduates in informatics and informatics engineering will always be in high demand — but that is a separate conversation.
And then there is the ever-present elephant in the room when it comes to informatics systems: security.
The Veracode October 2025 report on the security of GenAI-generated code reveals that 45% of it contains at least one of the ten most critical vulnerabilities identified by OWASP (Open Worldwide Application Security Project). For Java programs the situation is even worse. This sheds light on another important point: since GenAI tools have been trained on all existing code, and Java systems contain — for partly historical reasons — far more vulnerabilities than those written in other languages, the code that GenAI generates for this language is correspondingly more flawed.
Finally, the most damaging effect of all: the decline in hiring for entry-level software development roles. Because companies believed GenAI could handle junior-level tasks, hiring for these profiles collapsed. In 2024, the 15 largest Silicon Valley companies hired 25% fewer people for positions requiring less than one year of experience, while the reduction at start-ups was 11%. In its wake, we will likely also see a decline in enrolments in university degree programmes and upper secondary school diploma courses centred on informatics.
A suicidal choice, at every level. If companies do not hire inexperienced young people, they will never be able to develop experienced leaders. If young people do not learn to develop informatics systems with their own minds, they will never be able to manage those developed by GenAI.
Companies that have understood how to use GenAI are deploying it to help people be more effective in their work, freeing them from repetitive, low-level tasks. Those that are blindly relying on GenAI to get rid of new hires or pay them less will be forced to think again — and it will hurt.
The pendulum always swings back. What do you think?
--The original version (in italian) has been published by "StartMAG" on 16 February 2026.