When AI has favourite programming languages

A customer recently shared an observation that got us thinking.

One of their development teams had been using AI coding assistants like Cursor and ChatGPT extensively. Working in Python, the experience was everything we've come to expect from modern AI tools. They generated useful code, suggested sensible refactorings, and generally acted like an experienced pair programmer sitting alongside the team.

Then they switched to one of their Scala teams.

Nothing else had changed, but the experience was noticeably different. The suggestions became less reliable, the models seemed less confident, and language-specific idioms were often missed.

This isn't a criticism of Scala. It's a mature language with a rich ecosystem and years of production use behind it. Nor is it particularly surprising that AI models perform better in some languages than others.

The interesting question is what that observation means for the future of programming languages.

For decades, languages have competed through network effects. A language attracts developers, those developers build libraries, write documentation, publish open source projects, answer questions online, and create successful products. That attracts more developers, and the cycle reinforces itself.

AI introduces another feedback loop.

The more public code and documentation that exists for a language, the better AI models become at understanding and generating it. Better AI support makes developers more productive. Greater productivity encourages more teams to adopt that language, creating even more code for future models to learn from.

Programming languages are no longer competing solely for the attention of developers. Increasingly, they're also competing for the attention of the machines that help developers write software.

Recent research suggests this isn't simply anecdotal. Studies have found significant differences in LLM performance across programming languages, with models generally performing best in languages that have extensive public code and documentation. The precise reasons are still being explored, but the effect is becoming increasingly familiar to anyone using AI coding assistants every day.

The practical consequence is that AI has quietly become another factor in technology selection.

Historically, teams compared programming languages on performance, ecosystem maturity, maintainability, hiring, and developer experience. Those questions still matter, but another one is beginning to appear alongside them:

How well do our AI tools support this language?

That might sound like a small consideration, but it changes the economics of adoption. If one language consistently enables developers to move faster because AI tools understand it better, that becomes a genuine productivity advantage, regardless of whether the language itself is technically superior.

Language ergonomics now extends beyond the developer. It increasingly includes the machines that help developers write software.

This doesn't mean AI will determine which programming languages succeed. Strong ideas, good engineering, and vibrant communities will always matter. But it does raise the bar for emerging languages.

Historically, a new language needed to persuade developers that it offered something fundamentally better. Today, it may also need to persuade AI.

Fortunately, this isn't a one-way street. Better documentation, richer examples, active open source communities, and publicly available code all improve the quality of future AI models. We may even see languages designed with AI-assisted development as an explicit goal rather than an afterthought.

For decades, programming languages have existed primarily as a conversation between developers and computers. Developers express intent, compilers and runtimes execute it, and the language bridges the gap between the two.

Today there is a third participant in that conversation.

The next successful programming language won't just need to be expressive enough for developers or efficient enough for computers. Increasingly, it may also need to be understandable by the machines that help us build software.

That wasn't something language designers needed to think about a few years ago.

It probably is now.