AI solved syntax. It did not solve judgement.
Three independent measurements, taken between 2025 and 2026, all point the same way: AI-generated code passes over 95% on syntax and 55% on security, and that second number has not moved in two years. The part you write got cheap. The part you decide did not.
For thirty years, knowing how to program was an advantage. Writing code was expensive, slow and scarce, and the people who could do it set the pace of any project. That changed quickly, and it changed measurably. Generating code is no longer the hard step.
What follows is not a prediction about the future. It is three independent measurements, taken between 2025 and 2026, that all point the same way: the machine has solved the part of software you write, and it has not solved the part you decide.
What got cheap
Google's 2025 DORA report found that 90% of development professionals had adopted AI, fourteen points more than the year before. And unlike the year before, that adoption now comes with higher throughput: teams write and ship faster.
In March 2026, Veracode tested more than 150 language models across 80 coding tasks in four languages. The syntactic correctness of the generated code passed 95%. Producing code that compiles and does what it was asked to do is no longer the problem.
What did not
In the same study, on the same tasks, only 55% of the generated code passed security testing. Close to half of it carries known vulnerabilities when nobody gives explicit security instructions. And the number does not budge: it has sat at that same 55% for two years, through GPT-5.1 and 5.2, Gemini 3, and Claude 4.5 and 4.6.
| Language | Generated code passing security testing |
|---|---|
| Python | 62% |
| C# | 58% |
| JavaScript | 57% |
| Java | 29% |
Put the two numbers side by side: over 95% correct syntax, 55% correct security. The gap between them is exactly the gap between writing and deciding. The models learned to produce code that works. They did not learn to prefer the safe approach when an unsafe one is equally functional.
DORA sees the same effect from the other side. AI adoption still brings more instability with it: more failed changes, more rework, more time to recover. Speed went up. Soundness did not follow.
There is a third finding, and it is the most uncomfortable. In July 2025, METR ran a randomised trial with experienced developers on mature repositories they already knew well. With access to AI tools, they were 19% slower. Afterwards, they estimated that they had been 20% faster. It is a specific setting, and the authors themselves warn against generalising, but the lesson holds: the feeling of speed and actual speed came apart.
What stayed scarce
If code has stopped being the scarce resource, the value has moved to what automatic generation does not cover:
- knowledge of the business, which sits in no repository;
- architecture, whose main job is deciding what not to build;
- security, which the numbers above show is still unsolved;
- integration, which is where real systems break;
- accountability, which cannot be delegated to a model.
None of these got cheaper over the last two years. All of them became more valuable, because far more code now passes through them, and passes faster.
AI is an amplifier, not a cure
DORA's conclusion sums this up best: AI does not fix a team, it amplifies the team that is already there. Teams with automated tests, real review and fast feedback use AI to go further. Teams without them produce the same defects, only in greater volume and sooner.
That is why the useful question is no longer how much code we can produce. It has become a different one: what, in this particular organisation, is AI going to amplify?
What this changes in practice
The work moves upstream. Before connecting an AI tool to company data, you have to know who can read what, and that is a permissions and governance problem rather than a model problem: we covered it in what to put in order before switching Copilot on and in the AI assistant that only reads what the user can read. You still have to decide where the systems themselves sit, which was the subject of the state of SharePoint in 2026.
Two companies, the same tool
A few years from now, the difference between two companies in the same sector will not be the amount of software they produced. It will be what they chose to build, and above all what they had the discipline not to build. The tool is the same for both; the judgement is not.
AI lowered the cost of writing. It did not lower the cost of being wrong. That one is still paid in full, and it now arrives sooner.
