A new analysis published by CEPR takes a sober look at one of the tech industry's loudest promises: that AI coding tools dramatically boost programmer productivity. According to CEPR, the picture is more complicated than the marketing suggests.
The study, titled "Writing code versus shipping code: Productivity effects across generations of AI coding tools," draws a distinction that sits at the heart of the debate. Writing code — the raw act of generating lines with an AI assistant's help — is not the same as shipping code, meaning getting working software reviewed, tested, and out the door to users.
That gap matters. An AI tool can churn out code quickly, but if that output still needs heavy review, debugging, or rework before it can be deployed, the headline productivity gains can shrink. CEPR frames its findings around productivity effects measured across different generations of these tools, suggesting that newer versions do not automatically deliver bigger real-world wins.
The framing of the story is that the gains are mixed rather than uniformly large. Faster code generation does not always translate cleanly into faster delivery of finished, reliable software.
Why it matters: companies are investing heavily in AI coding assistants on the assumption they make developers far more productive, and this research is a reminder that the gap between generating code and actually shipping it may determine whether those investments pay off.