The METR study wasn't a verdict on every situation
The study described in the first module specifically examined experienced developers doing complex work in codebases of around one million lines of code. That's a specific, demanding setup – not the average coding task. Commentary on the study points out that results from this setup don't automatically transfer to other situations.
The adoption-productivity contradiction
A related pattern confirms this: even though most developers use AI coding tools, industry-wide productivity metrics don't show a correspondingly large, uniform jump. The obvious explanation isn't that the tools don't help at all – it's that their benefit depends heavily on context and therefore gets diluted on average.
Three context factors that make the difference
Three factors largely determine whether AI coding tools tend to help or tend to slow you down: the task type (boilerplate and scaffolding benefit more than complex architecture decisions), the codebase (greenfield projects and smaller, clearly scoped repositories benefit more than large, grown legacy systems), and experience level (beginners often benefit more from suggestions because they need to hold less of their own context in mind – experienced developers are more likely to lose flow through constant re-prompting).
Practice section: assess your own situation instead of guessing
Instead of blindly trusting industry promises or a single study, it's worth doing a simple self-assessment against the three context factors – and building on that with your own small time measurement on a typical task, with and without the tool. That also explains why developers use an average of 2.3 different tools in parallel: no single tool fits all three context factors equally well.